The Executive Guide To Artificial Intelligence How Identify And Implement Applications For AI%2

User Manual:

Open the PDF directly: View PDF .
Page Count: 187

Download
Open PDF In Browser	View PDF

THE

EXECUTIVE GUIDE

TO

ARTIFICIAL INTELLIGENCE

How to identify and implement applications
for AI in your organization

ANDREW BURGESS

The Executive Guide to Artificial Intelligence

Andrew Burgess

The Executive Guide
to Artificial
Intelligence
How to identify and implement
applications for AI in your organization

Andrew Burgess
AJBurgess Ltd
London, United Kingdom

ISBN 978-3-319-63819-5    ISBN 978-3-319-63820-1
https://doi.org/10.1007/978-3-319-63820-1

(eBook)

Library of Congress Control Number: 2017955043
© The Editor(s) (if applicable) and The Author(s) 2018
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether the
whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does
not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective
laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book are
believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors
give a warranty, express or implied, with respect to the material contained herein or for any errors or omissions
that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps
and institutional affiliations.
Cover illustration: Ukususha/iStock/Getty Images Plus
Printed on acid-free paper
This Palgrave Macmillan imprint is published by Springer Nature
The registered company is Springer International Publishing AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland

This book is dedicated to my wonderful wife, Meg, and our two amazing
children, James and Charlie.

Foreword

I remember well the AI work I did whilst at college studying computer
science, how different and fascinating it was and still is. We were set a very
open challenge to write an AI programme on any subject. I decided to
write mine so that it could tell you if the building in a photo was a house,
a flat or a bungalow. Somewhat impractical, but a great learning experience for me, particularly in understanding how AI is different from traditional software.
Although my college days were a number of years ago, since that time the
concept of computers learning has always intrigued me and I have since wondered how long it will take for AI to have a truly widespread impact. In recent
years, we’ve seen massive improvements in processing power, big data collection via sensors and the Internet of Things, cloud services, storage, ubiquitous
connectivity and much more. These technological leaps mean that this is the
right time for AI to become part of the ‘here and now’ and I strongly believe
we will see a dramatic increase in the use of AI over the next few years.
The AI in use today is known as narrow AI because it can excel at thousands
of relatively narrow tasks (e.g. doing internet searches, playing Go or looking
for fraudulent transactions). Things will certainly get even more exciting when
‘general AI’ can outperform humans at nearly every task we do, but we simply
don’t know when this might be, or what the world will then look like. Until
then, what excites me most is how we can apply AI now to solve our day-today problems at home and work.
So why is AI important and how can we use it? Firstly, if you are impatient
(like I am), doing small manual, repetitive tasks on computers simply takes
too much time. I want the computer to do a better job of anticipating my
needs and to just get on with it. If I could, I would prefer to talk to Alexa or
vii

viii

Foreword

Google Assistant and just tell the computer what to do. For example, I would
love to be able to ask Alexa to buy the most convenient train ticket and take
the money out of my account. Compare this to buying a train ticket on any
website, where after something like 50 key strokes you might have bought a
ticket. I don’t think future generations, who are becoming increasingly impatient, will put up with doing these simple and time-consuming tasks. I see my
children and future generations having more ‘thinking time’ and focusing on
things that are outside the normal tasks. AI may in fact free up so much of my
children’s time that they can finally clean up their bedrooms.
In the workplace, how many of the emails, phone calls and letters in a call
centre could be handled by AI? At Virgin Trains, we used AI to reduce the time
spent dealing with customer emails by 85% and this enabled our people to focus
on the personable customer service we’re famous for. Further improvements will
no doubt be possible in the future as we get better at developing conversational
interfaces, deep learning and process automation. One can imagine similar
developments revolutionising every part of the business, from how we hire people to how we measure the effectiveness of marketing campaigns.
So, what about the challenges of AI? One that springs to mind at Virgin is
how to get the ‘tone of voice’ right. Our people are bold, funny and empathetic, and our customers expect this from us in every channel. Conversational
interfaces driven by AI should be no different.
Today it may be a nuisance if your laptop crashes, but it becomes all the
more important that an AI system does what you want it to do if it controls
your car, your airplane or your pacemaker. With software systems that can
learn and adapt, we need to understand where the responsibility lies when
they go wrong. This is both a technical and an ethical challenge. Beyond this,
there are questions about data privacy, autonomous weapons, the ‘echo chamber’ problem of personalised news, the impact on society as increasing numbers of jobs can be automated and so on.
Despite these challenges, I am incredibly excited about the future of technology, and AI is right at the heart of the ‘revolution’. I think over the next
five to ten years AI will make us more productive at work, make us more
healthy and happy at home, and generally change the world for the better.
To exploit these opportunities to the full, businesses need people who
understand these emerging technologies and can navigate around the challenges. This book is essential reading if you want to understand this transformational technology and how it will impact your business.
John Sullivan, CIO and Innovation at Virgin Trains

Acknowledgements

I would like to thank the following people for providing valuable input, content and inspiration for this book:
Andrew Anderson, Celaton
Richard Benjamins, Axa
Matt Buskell, Rainbird
Ed Challis, Re:infer
Karl Chapman, Riverview Law
Tara Chittenden, The Law Society
Sarah Clayton, Kisaco Research
Dana Cuffe, Aldermore
Rob Divall, Aldermore
Gerard Frith, Matter
Chris Gayner, Genfour
Katie Gibbs, Aigen
Daniel Hulme, Satalia
Prof. Mary Lacity, University of Missouri-St Louis
Prof. Ilan Oshri, Loughborough University
Stephen Partridge, Palgrave
Mike Peters, Samara
Chris Popple, Lloyds Bank
John Sullivan, Virgin Trains
Cathy Tornbaum, Gartner

ix

x

Acknowledgements

Vasilis Tsolis, Congnitiv+
Will Venters, LSE
Kim Vigilia, Conde Naste
Prof. Leslie Willcocks, LSE
Everyone at Symphony Ventures

Contents

1	Don’t Believe the Hype   1
2	Why Now?  11
3	AI Capabilities Framework  29
4	Associated Technologies   55
5	AI in Action   73
6	Starting an AI Journey   91
7	AI Prototyping   117
8	What Could Possibly Go Wrong?   129
9	Industrialising AI  147
10	Where Next for AI?  165
Index  177

xi

List of Figures

Fig. 2.1
Fig. 2.2
Fig. 2.3
Fig. 3.1
Fig. 3.2
Fig. 3.3
Fig. 4.1
Fig. 4.2
Fig. 4.3
Fig. 6.1
Fig. 6.2
Fig. 6.3
Fig. 6.4
Fig. 9.1

Basic neural network
Training a neural network
A trained neural network
AI objectives
Knowledge map
The AI framework
Human in the loop
Training through a human in the loop
Crowd-sourced data training
Aligning with the business strategy
AI maturity matrix
AI heat map first pass
AI heat map
AI Eco-System

21
21
22
30
43
51
68
69
69
93
100
102
104
148

xiii

1
Don’t Believe the Hype

Introduction
Read any current affairs newspaper, magazine or journal, and you are likely to
find an article on artificial intelligence (AI), usually decrying the way the
‘robots are taking over’ and how this mysterious technology is the biggest risk
to humanity since the nuclear bomb was invented. Meanwhile the companies
actually creating AI applications make grand claims for their technology,
explaining how it will change peoples’ lives whilst obfuscating any real value
in a mist of marketing hyperbole. And then there is the actual technology
itself—a chimera of mathematics, data and computers—that appears to be a
black art to anyone outside of the developer world. No wonder that business
executives are confused about what AI can do for their business. What exactly
is AI? What does it do? How will it benefit my business? Where do I start? All
of these are valid questions that have been, to date, unanswered, and which
this book seeks to directly address.
Artificial Intelligence, in its broadest sense, will have a fundamental impact
on the way that we do business. Of that there is no doubt. It will change the
way that we make decisions, it will enable completely new business models to
be created and it will allow us to do things that we never before thought possible. But it will also replace the work currently being done by many knowledge workers, and will disproportionally reward those who adopt AI early and
effectively. It is both a huge opportunity and an ominous threat wrapped up
in a bewildering bundle of algorithms and jargon.
But this technological revolution is not something that is going to happen in
the future; this is not some theoretical exercise that will concern a few businesses.
© The Author(s) 2018
A. Burgess, The Executive Guide to Artificial Intelligence,
https://doi.org/10.1007/978-3-319-63820-1_1

1

2

1 Don’t Believe the Hype

Artificial Intelligence is being used today in businesses to augment, improve and
change the way that they work. Enlightened executives are already working out
how AI can add value to their businesses, seeking to understand all the different
types of AI and working out how to mitigate the risks that it inevitably brings.
Many of those efforts are hidden or kept secret by their instigators, either because
they don’t want the use of AI in their products or services to be widely known,
or because they don’t want to give away the competitive advantage that it
bestows. A persistent challenge for executives that want to get to grips with AI is
where to find all the relevant information without resorting to fanciful articles,
listening to vendor hyperbole or trying to understand algorithms. AI is firmly in
the arena of ‘conscious unknowns’—we know that we don’t know enough.
People generally experience AI first as consumers. All our smartphones
have access to sophisticated AI, whether that is Siri, Cortana or Google’s
Assistant. Our homes are now AI enabled through Amazon’s Alexa and Google
Home. All of these supposedly make our lives easier to organise, and generally
they do a pretty good job of it. But their use of AI is actually pretty limited.
Most of them rely on the ability to turn your speech into words, and then
those words into meaning. Once the intent has been established, the rest of
the task is pretty standard automation; find out the weather forecast, get train
times, play a song. And, although the speech recognition and natural language understanding (NLU) capabilities are very clever in what they achieve,
AI is so much more than that, especially in the world of business.
Artificial Intelligence can read thousands of legal contracts in minutes and
extract all the useful information out of them; it can identify cancerous
tumours with greater accuracy than human radiologists; it can identify fraudulent credit card behaviour before it happens; it can drive cars without drivers;
it can run data centres more efficiently than humans; it can predict when
customers (and employees) are going to desert you and, most importantly, it
can learn and evolve based on its own experiences.
But, until business executives understand what AI is, in simple-enough
terms, and how it can help their business, it will never reach its full potential.
Those with the foresight to use and exploit AI technologies are the ones that
need to know what it can do, and understand what they need to do to get
things going. That is the mission of this book. I will, over the course of the ten
chapters, set out a framework to help the reader get to grips with the eight
core capabilities of AI, and relate real business examples to each of these. I will
provide approaches, methodologies and tools so that you can start your AI
journey in the most efficient and effective way. I will also draw upon interviews and case studies from business leaders who are already implementing
AI, from established AI vendors, and from academics whose work focuses on
the practical application of AI.

Introducing the AI Framework

3

Introducing the AI Framework
My AI Framework was developed over the past few years through a need to be
able to make sense of the plethora of information, misinformation and
marketing-­speak that is written and talked about in AI. I am not a computer
coder or an AI developer, so I needed to put the world of AI into a language
that business people like myself could understand. I was continually frustrated by the laziness in the use of quite specific terminology in articles that
were actually meant to help explain AI, and which only made people more
confused than they were before. Terms like Artificial Intelligence, Cognitive
Automation and Machine Learning were being used interchangeably, despite
them being quite different things.
Through my work as a management consultant creating automation strategies for businesses, through reading many papers on the subject and speaking
to other practitioners and experts, I managed to boil all the available information down into eight core capabilities for AI: Image Recognition, Speech
Recognition, Search, Clustering, NLU, Optimisation, Prediction and
Understanding. In theory, any AI application can be associated with one or
more of these capabilities.
The first four of these are all to do with capturing information—getting
structured data out of unstructured, or big, data. These Capture categories are
the most mature today. There are many examples of each of these in use today:
we encounter Speech Recognition when we call up automated response lines;
we have Image Recognition automatically categorising our photographs; we
have a Search capability read and categorise the emails we send complaining
about our train being late and we are categorised into like-minded groups
every time we buy something from an online retailer. AI efficiently captures
all this unstructured and big data that we give it and turns it into something
useful (or intrusive, depending on your point of view, but that’s a topic to be
discussed in more detail later in the book).
The second group of NLU, Optimisation and Prediction are all trying to
work out, usually using that useful information that has just been captured,
what is happening. They are slightly less mature but all still have applications
in our daily lives. NLU turns that speech recognition data into something
useful—that is, what do all those individual words actually mean when they
are put together in a sentence? The Optimisation capability (which includes
problem solving and planning as core elements) covers a wide range of uses,
including working out what the best route is between your home and work.
And then the Prediction capability tries to work out what will happen next—
if we bought that book on early Japanese cinema then we are likely to want to
buy this other book on Akira Kurosawa.

4

1 Don’t Believe the Hype

Once we get to Understanding, it’s a different picture all together.
Understanding why something is happening really requires cognition; it
requires many inputs, the ability to draw on many experiences, and to conceptualise these into models that can be applied to different scenarios and
uses, which is something that the human brain is extremely good at, but AI,
to date, simply can’t do. All of the previous examples of AI capabilities have
been very specific (these are usually termed Narrow AI) but Understanding
requires general AI, and this simply doesn’t exist yet outside of our brains.
Artificial General Intelligence, as it is known, is the holy grail of AI researchers
but it is still very theoretical at this stage. I will discuss the future of AI in the
concluding chapter, but this book, as a practical guide to AI in business today,
will inherently focus on those Narrow AI capabilities that can be implemented
now.
You will already be starting to realise from some of the examples I have
given already that when AI is used in business it is usually implemented as a
combination of these individual capabilities strung together. Once the individual capabilities are understood, they can be combined to create meaningful
solutions to business problems and challenges. For example, I could ring up a
bank to ask for a loan: I could end up speaking to a machine rather than a
human, in which case AI will first be turning my voice into individual words
(Speech Recognition), working out what it is I want (NLU), deciding whether
I can get the loan (Optimisation) and then asking me whether I wanted to
know more about car insurance because people like me tend to need loans to
buy cars (Clustering and Prediction). That’s a fairly involved process that
draws on key AI capabilities, and one that doesn’t have to involve a human
being at all. The customer gets great service (the service is available day and
night, the phone is answered straight away and they get an immediate response
to their query), the process is efficient and effective for the business (operating
costs are low, the decision making is consistent) and revenue is potentially
increased (cross-selling additional products). So, the combining of the individual capabilities will be key to extracting the maximum value from AI.
The AI Framework therefore gives us a foundation to help understand what
AI can do (and to cut through that marketing hype), but also to help us apply
it to real business challenges. With this knowledge, we will be able to answer
questions such as; How will AI help me enhance customer service? How will
it make my business processes more efficient? And, how will it help me make
better decisions? All of these are valid questions that AI can help answer, and
ones that I will explore in detail in the course of this book.

The Impact of AI on Jobs

5

Defining AI
It’s interesting that in most of the examples I have given so far people often
don’t even realise they are actually dealing with AI. Some of the uses today,
such as planning a route in our satnav or getting a phrase translated in our
browser, are so ubiquitous that we forget that there is actually some really
clever stuff happening in the background. This has given rise to some tongue-­
in-­cheek definitions of what AI is: some say it is anything that will happen in
20 years’ time, others that it is only AI when it looks like it does in the movies.
But, for a book on AI, we do need a concise definition to work from.
The most useful definition of AI I have found is, unsurprisingly, from the
Oxford English Dictionary, which states that AI is “the theory and development of computer systems able to perform tasks normally requiring human
intelligence”. This definition is a little bit circular since it includes the word
‘intelligence’, and that just raises the question of what is intelligence, but we
won’t be going into that philosophical debate here.
Another definition of AI which can be quite useful is from Andrew Ng, who
was most recently the head of AI at the Chinese social media firm, Baidu, and
a bit of a rock star in the world of AI. He reckons that any cognitive process
that takes a human under one second to process is a potential candidate for
AI. Now, as the technologies get better and better this number may increase
over time, but for now it gives us a useful benchmark for the capabilities of AI.
Another way to look at AI goes back to the very beginnings of the technology and a very fundamental question: should these very clever technologies
seek to replace the work that human beings are doing or should they augment
it? There is a famous story of the two ‘founders’ of AI, both of whom were at
MIT: Marvin Minsky and Douglas Engelbart. Minsky declared “We’re going
to make machines intelligent. We are going to make them conscious!” To
which Engelbart reportedly replied: “You’re going to do all that for the
machines? What are you going to do for the people?” This debate is still raging
on today, and is responsible for some of those ‘robots will take over the world’
headlines that I discussed at the top of this chapter.

The Impact of AI on Jobs
It is clear that AI, as part of the wider automation movement, will have a
severe impact on jobs. There are AI applications, such as chatbots, which can
be seen as direct replacements for call centre workers. The ability to read thousands of documents in seconds and extract all the meaningful information

6

1 Don’t Believe the Hype

from them will hollow out a large part of the work done by accountants and
junior lawyers. But equally, AI can augment the work done by these groups as
well. In the call centre, cognitive reasoning systems can provide instant and
intuitive access to all of the knowledge that they require to do their jobs, even
if it is their first day on the job—the human agent can focus on dealing with
the customer on an emotional level whilst the required knowledge is provided
by the AI. The accountants and junior lawyers will now have the time to properly analyse the information that the AI has delivered to them rather than
spend hours and hours collating data and researching cases.
Whether the net impact on work will be positive or negative, that is, will
automation create more jobs than it destroys, is a matter of some debate.
When we look back at the ‘computer revolution’ of the late twentieth century
that was meant to herald massive increases in productivity and associated job
losses, we now know that the productivity benefits weren’t as great as people
predicted (PCs were harder to use than first imagined) and the computers
actually generated whole new industries themselves, from computer games to
movie streaming. And, just like the robots of today, computers still need to be
designed, manufactured, marketed, sold, maintained, regulated, fixed, fuelled,
upgraded and disposed of.
The big question, of course, is whether the gains from associated activities
plus any new activities created from automation will outweigh the loss of jobs
that have been replaced. I’m an optimist at heart, and my own view is that we
will be able to adapt to this new work eventually, but not before going through
a painful transition period (which is where a Universal Basic Income may
become a useful solution). The key factor therefore is the pace of change, with
all the indicators at the moment suggesting that the rate will only get faster in
the coming years. It’s clear that automation in general, and AI in particular,
are going to be huge disruptors to all aspects of our lives—most of it will be
good but there will be stuff that really challenges our morals and ethics. As
this book is a practical guide to implementing AI now, I’ll be exploring these
questions in a little more detail toward the end of the book, but the main
focus is very much on the benefits and challenges of implementing AI today.

A Technology Overview
The technology behind AI is fiendishly clever. At its heart, there are algorithms: an algorithm is just a sequence of instructions or a set of rules that are
followed to complete a task, so it could simply be a recipe or a railway timetable. The algorithms that power AI are essentially very complicated statistical

A Technology Overview

7

models—these models use probability to help find the best output from a
range of inputs, sometimes with a specific goal attached (‘if a customer has
watched these films, what other films would they also probably like to watch?’).
This book is certainly not about explaining the underlying AI technology; in
fact, it is deliberately void of technology jargon, but it is worth explaining
some of the principles that underpin the technology.
One of the ways that AI technologies are categorised is between ‘supervised’
and ‘unsupervised’ learning. Supervised learning is the most common
approach out of the two and refers to situations where the AI system is trained
using large amounts of data. For example, if you wanted to have an AI that
could identify pictures of dogs then you would show it thousands of pictures,
some of which had dogs and some of which didn’t. Crucially, all the pictures
would have been labelled as ‘a dog picture’ or ‘not a dog picture’. Using
machine learning (an AI technique which I’ll come on to later) and all the
training data the system learns the inherent characteristics of what a dog looks
like. It can then be tested on another set of similar data which has also been
tagged but this time the tags haven’t been revealed to the system. If it has been
trained well enough, the system will be able to identify the dogs in the pictures, and also correctly identify pictures where there is no dog. It can then be
let loose on real examples. And, if the people using your new ‘Is There a Dog
in My Picture?’ app are able to feed back when it gets it right or not, then the
system will continue to learn as it is being used. Supervised learning is generally used where the input data is unstructured or semi-structured, such as
images, sounds and the written word (Image Recognition, Speech Recognition
and Search capabilities in my AI Framework).
With Unsupervised Learning, the system starts with just a very large data
set that will mean nothing to it. What the AI is able to do though is to spot
clusters of similar points within the data. The AI is blissfully naive about what
the data means; all it does is look for patterns in vast quantities of numbers.
The great thing about this approach is that the user can also be naive—they
don’t need to know what they are looking for or what the connections are—all
that work is done by the AI. Once the clusters have been identified then predictions can be made for new inputs.
So, as an example, we may want to be able to work out the value of a house
in a particular neighbourhood. The price of a house is dependent on many
variables such as location, number of rooms, number of bathrooms, age, size
of garden and so on, all of which make it difficult to predict its value. But,
surely there must be some complicated connection between all of these variables, if only we could work it out? And that’s exactly what the AI does. If it
is fed enough base data, with each of those variables as well as the actual price,

8

1 Don’t Believe the Hype

then it uses statistical analysis to find all the connections—some variables may
be very strong influencers on price whilst others may be completely irrelevant.
You can then input the same variables for a house where the price is unknown
and it will be able to predict that value. The data that is input is structured
data this time, but the model that is created is really a black box. This apparent lack of transparency is one of AI’s Achilles’s heels, but one that can be
managed, and which I’ll discuss later in the book.
As well as the above two types of training, there are various other terms
associated with AI, and which I’ll cover briefly here, although for business
executives they only need to be understood at a superficial level. ‘Neural
Networks’ is the term used to describe the general approach where AI is mimicking the way that the brain processes information—many ‘neurons’ (100
billion in the case of the brain) are connected to each other in various degrees
of strength, and the strength of the connection can vary as the brain/machine
learns.
To give an over-simplified example, in the dog picture app above, the ‘black
nose’ neuron will have a strong influence on the ‘dog’ neuron, whereas a ‘horn’
neuron will not. All of these artificial neurons are connected together in layers, where each layer extracts an increasing level of complexity. This gives rise
to the term Deep Neural Networks. Machine Learning, where the machine
creates the model itself rather than a human creating the code (as in the examples I have given above), uses DNNs. So, think of these terms as concentric
circles: AI is the over-arching technology, of which Machine Learning is a core
approach that is enabled by DNNs.
There are obviously many more terms that are in common use in the AI
world, including decision tree learning, inductive logic programming, reinforcement learning and Bayesian networks, but I will cover these only when
absolutely necessary. The focus of this book, as you will now hopefully
appreciate, is on the business application of AI rather than its underlying
technologies.

About This Book
My working experience has been as a management consultant, helping organisations cope with the challenges of the time, from productivity improvement,
through change management and transformation to outsourcing and robotic
process automation, and now AI. I first came across AI properly in my work
in 2001. I was working as Chief Technology Officer in the Corporate Venturing
unit of a global insurance provider—my role was to identify new technologies

About This Book

9

that we could invest in and bring into the firm as a foundation client (it was
what we used to call the ‘incubator’ model). One of the technologies we
invested in was based around the idea of ‘smart agents’ that could be used as
an optimisation engine—each agent would have a specific goal and would
‘negotiate’ with the other agents to determine the ideal collective outcome. So,
for example, the system could determine the most effective way for trucks to
pass through a port, or the best way to generate the most revenue from the size
and arrangement of advertisements in newspapers. Although we didn’t call it
AI at the time, this is effectively what it was—using computer algorithms to
find optimal solutions to real problems.
Fast forward to 2017, and my work is focused almost exclusively on AI. I
work with enterprises to help create their AI strategy—identifying opportunities for AI, finding the right solution or vendor and creating the roadmap for
implementation. I don’t do this as a technologist, but as someone that understands the capabilities of AI and how those capabilities might address business
challenges and opportunities. There are plenty of people much cleverer than
me that can create the algorithms and design the actual solutions, but those
same people rarely understand the commerciality of business. I see myself as a
‘translator’ between the technologists and the business. And with AI, the challenge of translating the technology is orders of magnitude greater than with
traditional IT. Which is why I wanted to write this book—to bring that
understanding to where it can be used the best: on the front line of business.
So, this is not a book about the theoretical impact of AI and robots in 10
or 20 years’ time and it is certainly not a book about how to develop AI algorithms. This is a book for practitioners of AI, people who want to use AI to
make their businesses more competitive, more innovative and more future-­
proof. That will happen only if the business leaders and executives understand
what the capabilities of the technology are, and how it can be applied in a
practical way. That is the mission of this book: be as informed as possible
about AI, but without getting dragged down by the technology, so that you
can make the best decisions for your business. And it is also a heartfelt appeal:
whatever you read or hear about AI, don’t believe the hype.

2
Why Now?

A Brief History of AI
For anyone approaching AI now, the technology would seem like a relatively
new development, coming off the back of the internet and ‘big data’. But the
history of AI goes back over 50 years, and includes periods of stagnation
(often referred to as ‘AI Winters’) as well as acceleration. It is worth providing
a short history of AI so that the developments of today, which really is AI’s big
moment in the sun, can be put into context.
I’ve already mentioned in the previous chapter the two people that are considered some of the key founders of AI, Marvin Minsky and Douglas
Engelbert, who both originally worked at the Massachusetts Institute of
Technology in Boston, USA. But the person who coined the phrase ‘artificial
intelligence’ was John McCarthy, a professor at Stanford University in
California. McCarthy created the Stanford Artificial Intelligence Lab (SAIL),
which became a key focus area for AI on the west coast of America. The technology driving AI at that time would seem rudimentary when compared to
the neural networks of today, and certainly wouldn’t pass as AI to anyone with
a basic understanding of the technology, but it did satisfy our earlier definition of “the theory and development of computer systems able to perform
tasks normally requiring human intelligence”, at least to a very basic level.
Much of the work in the early development of AI was based around
‘expert systems’. Not wanting to disparage these approaches at all (they are
still being used today, many under the guise of AI), they were really no
more than ‘if this then that’ workflows. The programmer would lay out the
knowledge of the area that was being modelled in a series of branches and
© The Author(s) 2018
A. Burgess, The Executive Guide to Artificial Intelligence,
https://doi.org/10.1007/978-3-319-63820-1_2

11

12

2

Why Now?

loops, with each branch depending on an input from the user or a rule. So,
in a system designed to model a bank account recommendation process, the
user would be asked a series of questions (employment status, earnings, savings, etc.) with each answer sending the process down different branches
until it came to a conclusion. And, because that was essentially performing
a task ‘normally requiring human intelligence’ it was, at the time, considered AI. Today it wouldn’t really pass muster because it doesn’t have any
self-learning capability, which is a key facet of how AI is perceived to be
defined.
Interestingly, even now, this same approach is being used in many of the
chatbots that have proliferated across the internet. Most of these claim to use
AI, and some do, but most are passive decision trees. There are a number of
online chatbot platforms (most are free to use) that allow you to create your
own bots using this approach and they do a reasonable job for simple processes (I actually built my own a while back—it took about half a day and was
very basic—but it proved to me that it could be done by a non-technical
person and that there was very little, if any, AI actually involved).
There were a couple of ‘AI Winters’, when advances in AI stagnated for a
good number of years. Both of these were the result of over-inflated expectations and the withdrawal of funding. The first occurred between 1974 and
1980, triggered by three events: a 1973 report by Sir James Lighthill for the UK
Government which criticised the ‘grandiose objectives’ of the AI community
and its failure to get anywhere near reaching those objectives; and in the United
States where the Mansfield Amendment required the Advanced Research
Projects Agency (ARPA, now known as DARPA) to only fund projects with
clear missions and objectives, particularly relating to defence, which AI at the
time couldn’t satisfy; and the perceived failure of a key project for ARPA that
would have allowed fighter pilots to talk to their planes. These events meant
that much of the funding was withdrawn and AI became unfashionable.
The second AI Winter lasted from 1987 to 1993 and was chiefly due to the
failure of ‘expert systems’ to meet their over-inflated expectations in 1985,
when billions of dollars were being spent by corporations on the technology.
As with my own chatbot experience that I described above, expert systems
ultimately proved difficult to build and run. This made them overly expensive
and they quickly fell out of favour in the early 1990s, precipitated by the
simultaneous collapse of the associated hardware market (called Lisp
machines). In Japan, a 1981 program costing $850m to develop a ‘fifth generation computer’ that would be able to ‘carry on conversations, translate languages, interpret pictures, and reason like human beings’ failed to meet any of
its objectives by 1991 (some remain unmet in 2017). And, although DARPA
had started to fund AI projects in the United States in 1983 as a response to

The Role of Big Data

13

the Japanese efforts, this was withdrawn again in 1987 when new leadership
at the Information Processing Technology Office, which was directing the
efforts and funds of the AI, supercomputing and microelectronics projects,
concluded that AI was ‘not the next wave’. It dismissed expert systems as simply ‘clever programming’ (which, with hindsight, they were pretty close on).
I talk about these AI Winters in a little detail because there is the obvious
question of whether the current boom in AI is just another case of over-­
inflated expectations that will lead to a third spell of the technology being left
out in the cold. As we have seen in the previous chapter, the marketing
machines and industry analysts are in a complete froth about AI and what it
will be capable of. Expectations are extremely high and businesses are in danger of being catastrophically let down if they start to believe everything that is
being said and written. So, we need to understand what is driving this current
wave and why things might be different this time.
From a technology point of view, the only two words you really need to
know for now are ‘machine learning’. This is the twenty-first century’s version
of expert systems—the core approach that is driving all the developments and
applications (and funding). But, before I describe what machine learning is
(in non-technical language, of course), we need to understand what all the
other forces are that are contributing to this perfect storm, and why this time
things could be different for AI. There are, in my mind, four key drivers.

The Role of Big Data
The first driver for the explosion of interest and activity in AI is the sheer volume of data that is now available. The numbers vary, but it is generally agreed
that the amount of data generated across the globe is doubling in size every
two years—that means that by 2020 there will be 44 zettabytes (or 44 trillion
gigabytes) of data created or copied every year. This is important to us because
the majority of AI feeds off data—without data this AI would be worthless,
just like a power station without the fuel to run it.
To train an AI system (like a neural network) with any degree of accuracy
you would generally need millions of examples—the more complex the model,
the more examples are needed. This is why the big internet and social media
companies like Google and Facebook are so active in the AI space—they simply
have lots of data to work with. Every search that you make with Google—there
are around 3.5 billion searches made per day—and every time you post or like
something on Facebook—every day 421 billion statuses are updated, 350 million photos are uploaded and nearly 6 trillion Likes are made—more AI fuel is
generated. Facebook alone generates 4 million gigabytes of data every 24 hours.

14

2 Why Now?

This vast volume of data is then consumed by the AI to create value. To
pick up again on the simple example I used in the previous chapter, to train a
DNN (essentially a machine learning AI) to recognise pictures of dogs you
would need to have many sample images of dogs, all labelled as ‘Dog’, as well
as lots of other images that didn’t contain dogs, all labelled ‘No Dog’. Once
the system had been trained to recognise dogs using this set of data (it could
also go through a validation stage where the algorithm is tuned using a sub-set
of the training data), then it will need to be tested on a ‘clean’, that is, unlabelled, set of images. There are no strict guidelines for how much testing data
is required but, as a rule of thumb, this could represent around 30% of the
total data set.
These huge amounts of data that we create are being exploited every minute
of the day, most of the time without our knowledge (but implicit acceptance).
Take, for example, your Google searches. Occasionally, as you type in your
search term you may spell a word incorrectly. Google usually offers you results
based on a corrected or more common spelling of that word (so for example
if I search for ‘Andrew Durgess’ it shows me results for Andrew Burgess), or
you can choose to actually search for the uncommonly spelled version. What
this means is that Google is constantly collecting data on commonly misspelt
versions of words and, most importantly, which corrections that it suggests are
acceptable or not. All the data is then used to continually tune their AI-powered
spell checker. If, in my example, there was actually someone called Andrew
Durgess who suddenly became famous tomorrow so that many people were
searching for his name, then the correction that Google used on my name
would be quickly phased out as less and less people accepted it and instead
clicked on ‘Search instead for Andrew Durgess’.
But it’s not just social media and search engines that have seen exponential
increases in data. As more and more of our commercial activities are done
online, or processed through enterprise systems, more data on those activities
will be created. In the retail sector our purchases don’t have to be made online
to create data. Where every purchase, whilst not necessarily connecting it to
an identified buyer, is recorded, the retailers can use all that data to predict
trends and patterns that will help them optimise their supply chain. And
when those purchases can be connected to an individual customer, through a
loyalty card or an online account for example, then the data just gets richer
and more valuable. Now the retailer is able to predict what products or services you might also want to buy from them, and proactively market those to
you. If you are shopping online, it is not only the purchase data that is
recorded—every page that you visit, how long you spend on each one and
what products you view are all tracked, adding to the volume, and the value,
of data that can be exploited by AI.

The Role of Big Data

15

After the purchase has been made, businesses will continue to create, harvest
and extract value from your data. Every time you interact with them through
their website or contact centre, or provide feedback through a third-­party recommender site or social media, more data is created that is useful to them.
Even just using their product or service, if it is connected online, will create
data. As an example, telecoms companies will use data from your usage and
interactions to try and predict, using AI, whether you will leave them soon for
a competitor. Their ‘training data’ comes from the customers that have actually
cancelled their contracts—the AI has used it to identify all of the different
characteristics that make up a ‘churn customer’ which it can then apply to the
usage and behaviours of all of the other customers. In a similar way, banks can
identify fraudulent transactions in your credit card account simply because
they have so much data available of genuine and not-so-­genuine transactions
(there are roughly 300 million credit card transactions made every day).
Other sources of ‘big data’ are all the text-based documents that are being
created (newspapers, books, technical papers, blog posts, emails, etc.), genome
data, biomedical data (X-rays, PET, MRI, ultrasound, etc.) and climate data
(temperature, salinity, pressure, wind, oxygen content, etc.).
And where data doesn’t exist, it is being created deliberately. For the most
common, or hottest, areas of AI complete training sets of data have been
developed. For example, in order to be able to recognise handwritten numbers, a database of 60,000 samples of handwritten digits, as well as 10,000 test
samples, has been created by the National Institute of Standards (the database
is called MNIST). There are similar databases for face recognition, aerial
images, news articles, speech recognition, motion-tracking, biological data
and so on. These lists actually become a bellwether for where the most valuable applications for machine learning are right now.
The other interesting aspect about the explosion and exploitation of data is
that it is turning business models on their heads. Although Google and
Facebook didn’t originally set out to be data and AI companies, that is what
they have ended up as. But what that means now is that companies are being
created to harvest data whilst using a (usually free) different service as the
means to do it. An example of this, and one which uses data for a very good
cause, is Sea Hero Quest. At first sight this looks like a mobile phone game,
but what it actually does is use the data from how people play the game to
better understand dementia, and specifically how spatial navigation varies
between ages, genders and countries. At the time of writing 2.7 million people have played the game which means it has become the largest dementia
study in history. Commercial businesses will use the same approach of a ‘window’ product or service that only really exists to gather valuable data that can
be exploited elsewhere.

16

2

Why Now?

The Role of Cheap Storage
All the data that is being created needs to be stored somewhere. Which brings
me on to the second driver in favour of AI: the rapidly diminishing cost of
storage, coupled with the speed that data can be accessed and the size of the
machines that store it all.
In 1980 one gigabyte of storage would cost on average $437,500. Five years
later that had dropped to around a quarter, and by 1990 it was a fortieth of
the 1980 price, at $11,200. But that was nothing compared to the subsequent
reductions. At the turn of the century it was $11.00, in 2005 it was $1.24 and
in 2010 it was 9 cents. In 2016, the cost stood at just under 2 cents per gigabyte ($0.019).
Because of all that data that is generated through Facebook that I described
above, their data warehouse has 300 petabytes (300 million gigabytes) of data
(the amount of data actually stored is compressed down from that which is
originally generated). Accurate numbers are tricky to come by, but Amazon’s
Web Services (its commercial cloud offering) probably has more storage
capacity than Facebook. It is this sort of scale that results in a sub–2 cent
gigabyte price.
It’s not just costs that have shrunk; size has too. There is a photograph from
1956 that I use in some of the talks I give of an IBM hard drive being loaded
onto a plane with a fork-lift truck. The hard drive is the size of a large shed,
and only has a capacity of 5 Mb. Today that would just about be enough to
store one MP3 song. Amazon now has a fleet of trucks that are essentially
huge mobile hard drives, each capable of storing 100 petabytes (the whole of
the internet is around 18.5 petabytes). At the time of writing IBM have just
announced that they have been able to store information on a single atom. If
this approach can be industrialised it would mean that the entire iTunes
library of 35 million songs could be stored on a device the size of a credit card.

The Role of Faster Processors
All this cheap storage for our massive data sets is great news for AI, which, as
I have hopefully made very clear, lives off large amounts of data. But we also
have to be able to process that data too. So, the third driver in AI’s perfect
storm is faster processor speeds.
This is where the often-cited Moore’s Law comes in. The founder of Intel,
Gordon Moore, predicted in 1965 that the number of transistors that could

The Role of Ubiquitous Connectivity

17

fit onto a transistor would double every year. In 1975, he revised this to a
doubling every two years. It was actually David House, an Intel executive,
who came up with the most commonly used version which is that chip performance (as an outcome of more and faster transistors) would double every
18 months. Although there have been deviations from this trend, particularly
in the past few years, it means that the processors being used today are orders
of magnitude faster than during the last AI Winter.
One of the curious things to come out of technology for AI is that traditional computer chips (central processing units, or CPUs) are not ideally
suited to the crunching of large data sets, whereas GPUs (graphical processing
units), which had been developed to run demanding computer visualisations
(such as for computer games), were perfect for the job. Therefore NVidia, a
maker of GPUs, has taken most of the market for computer chips in the
world of AI.
So, faster AI-friendly processors mean that more complex problems, that
use more data, can be solved. This is important because managing and processing all that data does take time—the systems are fine at the second part of
the process, that of scoring and making a decision based on the learnt data,
but the training part can be a slog. Even relatively simple training sessions
need to run overnight, and more complex ones can take days. So any improvement in processor speed will help enormously in the usefulness of AI systems
both in the original development and design of the models but also on a day-­
to-­day basis so that the systems can use the most up-to-date models. Being
able to provide real-time training as well as real-time decision making is one
of the last frontiers of really useful AI.

The Role of Ubiquitous Connectivity
The final driver working in AI’s favour is connectivity. Clearly the internet has
had a huge enabling effect on the exploitation of data, but it is only in the past
few years that the networks (both broadband and 4G) have become fast
enough to allow large amounts of data to be distributed between servers and
devices. For AI this means that the bulk of the intensive real-time processing
of data can be carried out on servers in data centres, with the user devices
merely acting as a front end. Both Apple’s Siri (on the iPhone) and Amazon’s
Alexa (on the Echo) are prime examples of very sophisticated AI applications
that exploit processing power in data centres to carry out the bulk of the hard
work. This means there is less reliance on the device’s processor but it does put
a burden on having the network available and effective enough.

18

2 Why Now?

It’s not only the real-time processing where the internet can provide benefits. The sheer effort in training an AI model, with each training run taking
days or weeks on ‘standard’ hardware, can be significantly sped up using a
cloud-based solution that has specialised hardware available.
Better communication networks can also help AI systems in other ways.
The data sets that I mentioned in the previous section are huge but are made
available, usually publicly, to many users to help train their systems—that
wouldn’t have been nearly as easy previously.
Also, AI systems can be connected to each other using the internet so that
they can share learnings. A program run by a consortium of Stanford
University, University of California at Berkeley, Brown University and Cornell
University, called Robobrain, uses publicly available data (text, sounds,
images, film) to train AI systems which can be accessed by other AI systems.
And, of course, like any good AI system, the ‘recipient’ systems will feed back
everything that they learn back into Robobrain. The challenge for Robobrain,
which mirrors the overall challenge of AI, is that systems tend to be very narrow in their focus whereas Robobrain wants to be all things to all robots (or
‘multi-modal’ to use the lingo).

About Cloud AI
Nowhere has all four of these drivers come together more effectively than in
the concept of Cloud AI. The idea of AI-as-a-service, where the heavy-lifting
of the AI is done on the cloud and on-demand, is now a key catalyst for the
democratisation of AI. Many of the large technology firms, such as Google,
IBM and Amazon, all have cloud solutions for AI that offer easily accessible
APIs (Application Programming Interfaces, which are essentially standardised
access points to programming capabilities) for developers to create AI ‘front-­
ends’ out of them. IBM’s Watson—its heavily branded and lauded AI capability—is ‘just’ a series of APIs, with each one carrying out a specific function,
such as speech recognition or Q&A. Google’s TensorFlow is an open-source
AI platform that offers similar capabilities, as well as additional features such
as pre-trained models.
What this means for business, and specifically any entrepreneur wanting to
start an AI business, is that the value of the AI is not going to be where everyone assumes it be: in the algorithm. If every new customer-service focused AI
business used the open-source speech recognition algorithm from, say,
Amazon, then the competitive advantage would have to lie with the quality of
the training data, the way that the algorithm is trained or how easily the

What Is Machine Learning?

19

resulting system is to use. When you connect to an IBM Watson API, for
example, there is still a lot of work to do training it before you can derive any
real value.
Clearly some AI businesses will create competitive advantage from their
algorithm, but they must compete with the other firms who use an off-the-­
shelf one. It’s fairly straightforward, as I have done myself, to download a free,
open-source Named Entity Recognition algorithm (which is used to extract
names such as people, places, dates, etc., from a body of text) from a university website (in my case it was Stanford’s), feed it some sample text and get it
to return a reasonable answer. But that is not yet a viable AI solution, let alone
a business. To make it commercial I need to train and tune the algorithm
using as much data as I can get my hands on and create a user-friendly interface for it, and that is where the real skills—in data science, in optimisation,
in user experience—come into play. Put all that together and you may have
the seed of an AI business.
For the executive looking to use AI in their business (and the fact that you
are reading this book would suggest that is the case) the provenance of the
value in an AI company is important to understand. Why should you choose
one vendor over another, if they say they can do exactly the same thing?
Getting to the heart of the differences—is it down to the algorithm, the data,
the ease of implementation, the ease of training, the ease of use and so on?—
will allow you to make the right choice for your business. At the moment,
there are far too many smoke-and-mirrors AI businesses that are based purely
on open-source algorithms and built by very clever but inexperienced
20-somethings. They will be riding the hype of AI but will offer little in the
way of long-term value. My point being that you can create a successful AI
business based on great open-source algorithms, but you need all the other
aspects to be great too. As I said right at the start of this book: don’t believe
the hype.
I’ll be covering these selection considerations, as well as the buy-versus-­
build question, in more detail later in the book.

What Is Machine Learning?
These four developments then—big data, cheap storage, faster processors and
ubiquitous networks—are what is driving the acceleration and adoption of AI
in business today. If you were to take any one of those away, AI would struggle
to perform, and we’d probably still be in an AI Winter. Each one has also been
helping drive development in the others—if we didn’t have such cheap storage

20

2

Why Now?

then we couldn’t keep all that data which means there would be no need for
faster processors, for example. But, at the centre of all this activity, feeding off
each of the drivers, is machine learning, one of the most popular AI approaches.
It’s important to have an appreciation of what machine learning is and
what it does, even if you don’t need to understand how it actually works. As
the name suggests, it is the machine that does all of the hard computational
work in learning how to solve a problem—the machine, rather than a human
being, essentially writes the ‘code’. The human developer plants the seed by
defining the algorithm or algorithms that are to be used, and then the machine
uses the data to create a solution specific to it. We have already seen that the
problem could be undefined initially (as in unsupervised learning, where patterns or clusters are found in data) or defined (and therefore trained with large
data sets to answer a particular question, which we know as supervised learning). For machine learning it is usually easier to think about a supervised
learning example, such as my go-to case of being able to tell the difference
between pictures of dogs and pictures of cats.
In the previous chapter, I introduced the idea of DNNs, which is the ‘architecture’ that machine learning uses. A DNN will consist of a number of different layers; the more complex the problem being solved, the more layers are
needed. (More layers also mean that the model is much more complicated,
will require more computing power and will take much longer to resolve
itself.) The Input Layer takes the data in and starts to encode it. The Output
Layer is where the answer is presented—it will have as many nodes as there are
classes (types) of answer. So, in my dog versus cat example, there would be
two output nodes, one for dog and one for cat (although we could define
three if we had pictures that had neither a dog nor a cat in) (Fig. 2.1).
In between the Input Layer and the Output Layer are the Hidden Layers.
It is here that all the hard work happens. Each Hidden Layer will be looking
for different features within the data with increasing complexity, so with
image recognition different layers will look for outlines, shadows, shapes,
colours and so on. In these Hidden Layers, there will be more nodes (­ ‘neurons’)
than the Input or Output Layers, and these nodes will be connected to each
other across each layer. Each of these connections will have a certain ‘weight’
or strength which determines how much of the information in one node is
taken into the next layer: a strong link, which had, through training, led to a
‘right’ answer at the Output Layer, will mean that information is propagated
down through to the next layer. A weak link with a low weighting, which had
led to a ‘wrong’ answer in training, will not pass as much information down
(Fig. 2.2).

What Is Machine Learning?

Input
Layer

Hidden
Layer

Hidden
Layer

Output
Layer

Hidden
Layer 2

Output
Layer

21

Fig. 2.1 Basic neural network

Input
Layer

Hidden
Layer 1

Fig. 2.2 Training a neural network

As the model is being trained with more and more data, the weights are
continually adjusted (this is the machine learning) until it reaches an optimal
solution. The more data we have fed into the model, the more chances the
machine has to refine the weightings (but the harder it has to work), and
therefore the more accurate the solution will be. The ‘matching function’ (the

22

2

Why Now?

DOG

Input
Layer

Hidden
Layer

Hidden
Layer

Output
Layer

Fig. 2.3 A trained neural network

final version of the model) can then be used to solve for a new piece of data;
for example, I give it a picture of a dog that it hasn’t seen before and it should
then be able to correctly identify it as a dog (Fig. 2.3).
From this brief description of machine learning you should be able to see
that it is wholly dependent on the four drivers I have described—we need lots
and lots of data to train the hidden layers to come up with the right weightings, but that means we need to be able to store our data cheaply and process
our models as quickly as possible, as well as needing access to data sets from
as many sources as possible. Without any one of these, machine learning just
wouldn’t be viable, either because it wouldn’t be accurate enough or it wouldn’t
be easy enough to design and implement (and it’s already difficult enough in
the first place).
It’s not just machine learning that benefits from this ‘perfect storm’ of technology. Other types of AI, the main one being ‘symbolic AI’, have also accelerated and found new life through faster, cheaper and connected computers.

The Barriers to AI
So, with all this data, cheap storage, whizzy processors and inter-connectivity,
we are certain to never see another AI Winter, right? Well, there are a few
things that could spoil the AI party.

The Barriers to AI

23

The biggest barrier to AI achieving escape velocity in my opinion is the
over-inflation of expectations. Far too much is being written and said about
what AI might be capable of, especially when Artificial General Intelligence is
considered and not the Narrow AI that we have today. Both the previous AI
Winters have been precipitated by excessive and unrealistic expectations, and
therefore they need to be managed carefully. A key motivation for me to write
this book was to ‘bring AI down to earth’ so that it could be assessed for its
real value today and tomorrow, and not in 10 or 20 years’ time.
Another factor that will affect the uptake of AI, which has been fuelled by
the hype, is a general fear of what changes AI might bring, especially when it
could mean fundamentally changing the way that people work. There have
been many other seismic shifts in ways of working, including the introduction
of personal computers and outsourcing, but AI has the potential to do this on
a much greater scale: speculative reports by academics and institutions repeatedly report the decimation of ‘white collar’ jobs and the ‘hollowing out’ of the
middle classes. All of this may or may to be true, but the fear that it generates
leads to a natural defensiveness from those in businesses that may be affected
by the implementation of AI.
A third aspect, which is connected to the first two, is ignorance. How can
people expect to get real value from implementing AI if they don’t have a good
enough understanding of what it is? The hype doesn’t help, but AI is inherently a complex subject, and it’s not an easy task for a non-technical business
person to understand it. The maxim that a little knowledge is a dangerous
thing is certainly true for AI, but also having all the knowledge is not useful
unless you want to be a data scientist or AI developer. There is a ‘Goldilocks’
level of understanding AI that this book aims to provide—it should be just
enough that you can make the most out of AI but not too much that you
become overwhelmed with technical detail and jargon.
The final thing that could put the brakes on AI is regulation. As I’ve mentioned already, much of the computational work that is done by an AI system
is hidden from the developer or user—there is (generally) no audit trail that
details how it came to a certain decision. If I fed a complex credit decisioning
AI system with lots of loan applications as training data and told it which had
been approved and which hadn’t, it could then make a recommendation of
whether to approve or reject a new loan. But no one would really know why,
and that causes a problem for regulators who need to see that decision process
written down. There are some approaches which can mitigate this but the
opaqueness of AI can be a real challenge to its usefulness.
I’ll cover these aspects in more detail in Chap. 8, but I think it’s important
that you have an early appreciation of them at this point.

24

2 Why Now?

Some AI Case Studies
Although it may be fuelled by perhaps too much hype, there is certainly
momentum in the AI market right now. But who is benefiting from all these
advancements? Is it just research labs and some start-ups that have had success
playing computer games, or is there something more? Is AI providing value to
businesses today?
In my work, I see plenty of examples of AI being used in businesses to provide insights and efficiencies, adding value that would be beyond the means
of humans to create.
One of the sweet spots for AI at the moment is in customer services. At a
train operating company in the United Kingdom they use AI to categorise
emails that come in from their customers. The challenge, of course, is for the
AI to be able to ‘read’ the email, which has been written freeform, and determine what the customer actually wants. For example, if I couldn’t get a seat
on the 08:06 from Euston Station in London to Glasgow Central, for example, I would email the train company, saying something like “I can’t believe
it – I paid all that money for a ticket on the 08:06 from London to Glasgow
and I couldn’t even get a seat. You’ve really excelled yourselves this time!” All
the salient information is extracted from my unstructured text and validated
(Was there a train at 08:06? Have other people complained about the same
service? Is the customer a regular complainer? etc.) so that the AI can immediately route my query to the relevant person in the customer service organisation. Of course, my wording in the second sentence is sarcastic in tone, so
the AI needs to understand the difference between my complaint and an
actual compliment. What it all means is that when the human agent logs in,
they are the right person to deal with my query and they have everything at
hand to respond.
In a similar vein, a UK-based insurance firm that provides a claims processing service to their own insurance customers has automated the input of
unstructured and semi-structured data (incoming claims, correspondence,
complaints, underwriters’ reports, cheques and all other documents relating to
insurance claims) so that it is directed into the right systems and queues. Using
an AI solution, a team of four people process around 3,000 claims documents
per day, 25% of which are paper. The AI automation tool processes the scanned
and electronic documents automatically, identifies claim information and
other metadata, and deposits the results in databases and document stores
ready for processing by the claims handlers and systems. It also adds service
metadata so performance of the process can be measured end to end. Some

Conclusion

25

documents can be processed without any human intervention, and others
need a glance from the human team to validate the AI’s decisions or fill in
missing details.
A good example of AI in the legal sector is ROSS, a system built around
IBM’s Watson. ROSS actually uses a number of AI capabilities including
NLU, Search and Optimisation. So, when a lawyer has a matter that requires
some specific research (e.g. what is the precedent when an employee is fired
for gross misconduct within a few days of the end of their probationary
period?), she can turn to ROSS and type in that query in plain English. ROSS
will then interrogate the whole corpus of employment law and provide
answers ranked by relevancy back to the lawyer within seconds. The alternative would be hours of research by the lawyer, a junior lawyer and/or a paralegal, and likely without being able to examine every relevant document
available. Like all good AI systems, ROSS is self-learning—the lawyer can
evaluate the quality of the responses that ROSS returns and therefore allow it
to provide better answers in the future.
Some companies have used AI to create whole new lines of business. Charles
Schwab, the US bank and brokerage firm, has created an investment vehicle,
called Schwab Intelligent Portfolios, which uses AI to manage its clients’ portfolios (these are sometimes called ‘robo-advisors’). It focuses on low-cost
exchange-traded funds and has no advisory fees, account service fees or commission fees. Since it launched in 2015, a few other firms have appeared with
a similar model, including Betterment, Wealthfront and FutureAdvisor
(although these charge a very small admin fee). The attractiveness of the low
or zero fees and the simplicity of dealing with an AI rather than a person have
meant that these sorts of services have proved popular amongst beginner
investors, thus providing new clients for the banks that they normally wouldn’t
have been able to access.

Conclusion
So, AI is being used successfully by businesses around the world. The perfect
storm of big data, cheap storage, faster processors and ubiquitous connectivity
has allowed researchers to exploit the near-magic that is machine learning.
But AI has managed to escape the lab and is being commercialised by both
start-ups and internet giants. Real opportunities exist for executives to use AI’s
capabilities to deliver new sources of value into their business as well as to
challenge and disrupt their existing business models. In order to do that they
need to understand what those capabilities are and how they can be used.

26

2

Why Now?

The Academic’s View
This is an extract from an interview with Dr Will Venters, Assistant Professor
in Information Systems at the London School of Economics.
AB What’s the basis of all the fuss around AI right now?
WV	There are a few trajectories that are particularly important. The underlying algorithms have become much smarter and are used more effectively,
but that’s not actually that significant. What has really made the difference is the processing capacity of the computer chips as well as the ability
to handle much bigger volumes of data at scale.
Previously we saw IBM’s Big Blue beat Kasparov at chess, then Watson
won at Jeopardy!, but essentially these are just systems that can search
massive databases very quickly. But what AlphaGo is doing (beating the
world’s number one player in the Chinese game of Go) is amazing—it
watched, then played millions of games of Go against itself. The complexity and the sheer quantity of data that it was able to analyse is phenomenal—
I’m sure the future of AI will be around this sort of capability.
It’s interesting if you look at data complexity versus task complexity
for processes that computers want to take on: we are now able to manage
situations that are complex in both of these aspects. Managing elevators
is a complex task, but the data is relatively simple. Self-driving cars are
not doing complex things (just driving and turning and stopping) but the
data they need to do that is vast. But, we are now able to manage situations that are complex in both data and task aspects. Suddenly we have
the mechanisms to do it.
AB You also mentioned processing power as being a key factor?
WV	Yes, GPUs [Graphical Processing Units] have made a significant contribution to this. But also the widespread use of Cloud Computing, which
allows businesses to manage big data without the associated infrastructure investments and risks.
Lots of companies have invested in data infrastructure, either their
own or through the cloud, and AI allows them to utilise this to the full.
AB Do you think the hype helps or hinders AI?
WV	AI is certainly hyped, but it is also productive. It leads to new innovations
and forces businesses to discuss big questions. Hype shouldn’t only be
seen in the negative.
AB What do you see as the most valuable use cases for AI?
WV	AI can offer the opportunity to deal with ‘messy’ data – data that is fuzzier,
such as photographs, natural language speech and unstructured documents.

The Academic’s View

27

Big Data promised to deliver all of this but ultimately it didn’t because it
required armies of statisticians. Robotic Process Automation is great for
dealing with processes that are clean and structured – as soon as it becomes
messier then AI can step in and overcome that.
AB What do you think businesses need to do if they are to exploit AI to the full?
WV	They first need to look at the business case, and not the technology. But
they also need to consider the governance and to understand how they
manage the inherent risks. Humans have an ethical compass which, of
course, AI doesn’t. It will need some level of oversight because of the lack
of transparency with the AI models. They must also make sure the system
is not biased – it’s very easy to bake the bias into the company’s business.
And these risks scale very quickly because the underlying data is so much
bigger.
AB What do you see for the future for AI?
WV	I think we will see even better algorithms to deal with the bigger volumes
of data. But, of course, AI is a victim of its own success – there is an
inherent disillusionment as each exciting step forward then becomes
common place. You just need to look at advances like [Apple’s] Siri and
[Amazon’s] Alexa. The term AI will eventually become redundant and
people will cease to be questioning about it.
But until that day companies will invest in AI wisely as well as inappropriately. There is so much value to be gained that companies should
think deeply about their approach, take expert advice and make sure they
are not left behind.

3
AI Capabilities Framework

Introduction
It’s impossible to get value out of something if it is not understood, unless it’s
by some happy accident. In the world of AI there are no happy accidents—
everything is designed with meticulous detail and with specific goals in mind.
Therefore, the only way to truly make the most of AI in business is by having
a reasonable understanding of it. The challenge, of course, is that is fiendishly
complicated and full of complex mathematics, and is certainly not within the
bounds of what an ‘ordinary’ business person would be expected to grasp.
My approach in this book is to understand AI from the perspective of what
it can do, that is, what are its capabilities in the context of real world problems
and opportunities? In order to do this, I have developed a framework that
‘boils down’ all of the complexity into eight capabilities. In theory, any AI
application should fit into one of these, which allows the AI layman to quickly
assess and apply that application into something that relates to their business.
Conversely, if you have a specific business challenge, the AI framework can be
used to identify the most appropriate AI capability (or capabilities) that could
address that need. Having said that, I’m sure any AI scientist or researcher
would be able to pick some holes in the framework. It’s certainly not bomb-­
proof: it is intended to be a useful tool for business executives to help them get
the most value possible out of AI.
There are many ways to look at AI, some of which I have already discussed in
earlier chapters: Supervised versus Unsupervised Learning; Machine Learning
versus Symbolic Learning; Structured Data versus Unstructured Data;
Augmentation versus Replacement. All of these are valid perspectives, and all can
be understood to a reasonable enough degree within the capability framework.
© The Author(s) 2018
A. Burgess, The Executive Guide to Artificial Intelligence,
https://doi.org/10.1007/978-3-319-63820-1_3

29

30

3 AI Capabilities Framework

So, with each of the capabilities I will explain whether it will require
supervised or unsupervised learning (or both), the type of AI that is usually
used to deliver the capability (e.g. machine learning), whether it processes
structured or unstructured data and, most importantly, how it benefits businesses. In subsequent chapters I will draw out use cases for each, grouped into
the key themes of enhancing customer service, business process optimisation
and enhanced decision making.
Another perspective for AI, which I introduced in Chap. 1, is really the
meta-view, and probably the most important to understand first: these are the
three basic things that AI tries to achieve: capturing information; determining
what is happening and understanding why it is happening. Each of my eight
capabilities fits into one of these ‘AI Objectives’ (Fig. 3.1):
Capturing information is something that our brain does very well but
machines have historically struggled with. For example, the ability to recognise
a face is something that has evolved in humans since our earliest existence—it
is a skill that allows us to avoid danger and create relationships, and therefore
a great deal of brain power and capacity is used to work on this problem. For
machines, the process is very complex, requiring huge volumes of training
data and fast computer processors, but today we have this capability on our
desktop computers and even mobile phones. It may not be very fast or that
accurate yet but AI has been fundamental in achieving this.
Most of the examples of capturing information are where unstructured
data (such as a picture of a face) is turned into structured data (the person’s
name). Capturing information is also relevant to structured data, and AI
comes into its own when that data is big. Again, the human brain is very good
at identifying patterns in data (the impact of a manager on the success of a
football team, for example) but when there are hundreds of variables and millions of data points we lose sight of the forest and can only see the trees. AI is
able to find patterns, or clusters, of data that would be invisible to a human.
These clusters have the ability to provide insights within the data that have
real value to a business. For example, AI can find patterns between purchases
and customer demographics that a human would either take years to unearth,
or never have thought of in the first place. (This associated idea of the AI
being ‘naive’ to the data is something that I will return to later, but is an
important concept to bear in mind.)

CAPTURE INFORMATION

Fig. 3.1 AI objectives

WHAT IS HAPPENING?

WHY IS IT HAPPENING?

Introduction

31

The next AI Purpose is where it tries to determine what is happening, and
is usually a consequence of where information has already been captured by
AI. For example, we could have used Speech Recognition to extract (from a
sound file or a live conversation) the words that someone was speaking, but at
that point we would only know what all the individual words were rather than
what the person was actually trying to say. This is where Natural Language
Understanding (NLU) comes in—it takes the words and tries to determine
the meaning or intent of the complete sentences. So, we have gone from a
digital stream of sound to a set of words (e.g. ‘I’, ‘want’, ‘to’, ‘cancel’, ‘my’,
‘direct’, ‘debit’, ‘for’, ‘mortgage’, ‘protection’) to working out that this person
‘wants to cancel her Direct Debit for her mortgage protection insurance’.
We can then use other capabilities in this category to take that request further. For example, we could use an Optimisation approach to help the customer understand that if they were to cancel that Direct Debit they would
also probably need to cancel the insurance policy that it is associated with.
And then we could use a Prediction capability to work out that this customer
may be about to leave the bank and go to a competitor (because the AI has
determined from lots of other similar interactions that cancelling Direct
Debits is an indicator of potential customer churn).
So, from a simple interaction with a customer we are able to use different
AI ‘capturing information’ and ‘what is happening’ capabilities to build up a
useful picture as well as satisfying the customer’s request. But, in the example
above, although the AI has been able to identify this customer as a ‘churn risk’
it really doesn’t understand what that means. All it has done is correlate one
set of data (customer requests) with another (customers who leave) and then
apply that to a new data point (our customer’s Direct Debit cancellation
request). To the AI system the data might as well be about ice cream flavours
and the weather. It has no idea of the concept of Direct Debits just as much as
it has no idea of the concepts of banks or customers. The AI capabilities that
we have today and will have in the near future (and maybe even ever) do not
have the ability to understand. It is really important to be able to differentiate
between the very narrow things that different AIs can do (usually better than
humans) and the general intelligence that comes with understanding and
relating different concepts—something our brain is brilliant at doing.
Now that we have the three AI purposes defined, and we understand that
only the first two are relevant to us today, we can start to look in more detail
at each of the eight AI capabilities in the Framework.

32

3 AI Capabilities Framework

Image Recognition
One of the most active areas of research in AI at the moment is that of Image
Recognition. This is a prime example of where the four key drivers that I
described in Chap. 2 have aligned to catalyse the technology. Image
Recognition is based on Machine Learning and requires thousands or millions of sample images for training; therefore, it also needs lots of storage for
all the data, and fast computers to process it all. Connectivity is also important to enable access to the data sets, many of which are publicly available (in
Chap. 2 I mentioned image data sets for handwritten numbers and faces, but
there are many more including aerial images, cityscapes and animals).
Images, of course, fall under the category of unstructured data. So, what
sort of applications would you use Image Recognition for? There are three
main types:
Probably the most popular application is identifying what is in the image,
sometimes referred to as tagging. I’ve used this example a few times already—
find out if the picture contains a dog or a cat. It is often used to moderate
photographs by trying to identify pornographic or abusive images, but can
also be used to group photos with similar tags (‘these are all pictures taken at
the beach’). This photo tagging is a prime example of supervised learning—
the AI is trained on thousands, or millions, of tagged photographs—which is
why companies that have access to large volumes of images, such as Google
and Facebook, have some of the most advanced systems in this area.
Another use of Image Recognition is to find images that are similar to other
images. Google’s Reverse Image Search is a popular use of this approach; you
simply upload a picture and it will search for pictures that look similar to your
original (this technique is often used to identify the use of news photos that
have deliberately been used out of context). Unlike photo tagging, this is
mainly an example of unsupervised learning; the AI doesn’t need to know
what is in the picture, just that it looks like another picture. (A simple way to
think about this is to understand that the AI converts the image file into a
long series of numbers—it then looks for other images that have a similar
series of numbers.)
The final application for Image Recognition is to find differences between
images. The most common, and beneficial, use of this is in medical imaging.
AI systems are used to look at scans of parts of the body and identify any
anomalies, such as cancerous cells. IBM’s Watson has been a pioneer in this
space and is used to support radiologists in their work. The approach uses
supervised learning to tag, for example, X-rays as healthy or unhealthy. Based

Speech Recognition

33

on the algorithmic model that is built up, new images can be evaluated by the
AI to determine whether there is a risk to the patient. Interestingly, it has been
reported that Watson is more accurate at spotting melanomas than when it is
done manually (Watson has 95% accuracy compared to an average of between
75% and 84% for humans).
Image Recognition, probably out of all the capabilities I describe in this
chapter, is the most data hungry. Images are inherently unstructured and very
variable and therefore require very large amounts of data to train them effectively. Pinterest, the website that allows users to create ‘boards’ of their favourite images, has used all of the data from their hundreds of millions of users to
help develop their systems further (it does a good job of finding similar images
to ones that you have posted, even if they are not tagged and are pretty
abstract) but also to develop new AI applications such as Pinterest Lens which
enables you to point your phone camera at an object and it will return pictures of objects that are visually similar, have related ideas or are the same
object in other environments.
Other applications of Image Recognition are not so altruistic. A website in
Russia, where privacy laws are more lax than most Western countries, enables
its users to identify people in the street using their phone camera. Find Face,
as the website is called, exploits the fact that all 410 million profile pictures on
the country’s most popular social media site, VK, are made public by default.
Therefore, it is able to match, to an apparent accuracy of 70%, a face that you
point your camera at with a VK profile. The intended, and rather bawdy, uses
of this application are clear from the pictures of women on the website’s
homepage, but it does demonstrate how well image recognition technology
can work if it has a large enough training set to work from.
Image recognition is still at a relatively immature stage—many useful
things are possible now, but the potential is much greater. The use of still and
moving images in our daily lives and in the business world is increasing exponentially and therefore the ability to index and extract meaningful data from
these will be an increasingly important application.

Speech Recognition
Speech recognition, sometimes referred to as speech-to-text, is usually the first
stage in a string of AI capabilities where the user is providing instructions by
voice. It takes the sounds, whether live or recorded, and encodes them into
text words and sentences. At this point other AI capabilities such as NLU
would be needed to determine the meaning of the encoded sentences.

34

3 AI Capabilities Framework

Speech recognition has benefited enormously from the development of
DNNs although some of the more ‘traditional’ AI approaches (usually using
something called a Hidden Markov Model or HMM) are still widely used
today, mainly because they are much more efficient at modelling longer sections of speech. Just like for the images, the input data is unstructured, with
the technology using supervised learning to match the encoded words with
tagged training data (as such there are a number of publicly available speech-­
related training sets).
There are many challenges to an efficient and accurate speech recognition
system, as most readers will have experienced themselves whilst trying to get
their smartphone to understand a voice command from them. One of the
main challenges is the quality of the input—this could be because of a noisy
environment or because the voice is on the other end of a telephone line—
accuracy on the phone drops from a best-in-class Word Error Rate (WER) of
around 7% to more than 16%. (Human accuracy is estimated to be around
4% WER.)
Other challenges include obvious things like different languages and
regional accents but a key consideration is the size of the vocabulary that will
be required. For a very specific task, say checking your bank balance, the
vocabulary will be very small, perhaps ten or so words. For systems that are
expected to answer a wide range of questions, such as Amazon’s Alexa, then
the vocabulary will be much larger, and therefore present a greater challenge
for the AI.
A wider vocabulary requirement also means that context becomes more
challenging to determine. Context is important in speech recognition because
it provides clues to what words are expected to be spoken. For example, if we
hear the word ‘environment’ we would need to understand its context to
determine whether the speaker meant ‘our immediate surroundings’ or ‘the
natural world’. DNNs, and a particular type of DNN called a Recurrent
Neural Network, are very good at looking backwards and forwards through
the sentence to constantly refine the probability that it is one meaning or
another.
It is worth pointing out that speech recognition is a slightly different concept to voice recognition. Voice recognition is used to identify people from
their voices and not to necessarily recognise the words they are saying. But
many of the concepts used in the two applications are similar.
Speech recognition, and its cousin NLU, are seeing plenty of development
activity as people become more used to the technology through their smartphones, and therefore more comfortable and confident with talking to
machines. As a ‘user interface’, speech recognition will likely become the
dominant input method for many automated processes.

Search

35

Search
I use the word ‘Search’ in a specific sense here (another common term for this
is ‘Information Extraction’). What this AI capability is all about is the extraction of structured data from unstructured text. So, just as the Image
Recognition and Speech Recognition capabilities worked their ‘magic’ on pictures and sounds, so the Search capability does with documents.
There are strong elements of Natural Language Processing (NLP) used in
this capability to extract the words and sentences from the passages of text, but
I prefer to use the term ‘Search’ because it better describes the overall capability. Later on in this chapter you will hear about Natural Language Understanding,
which is generally described as a sub-set of NLP, but, in my mind, is a capability in itself which complements the outputs of Speech Recognition and Search.
AI Search, which is almost exclusively a supervised learning approach,
works on both unstructured and semi-structured data. By unstructured data
I mean something like a free-form email or a report. Semi-structured documents have some level of consistency between different instances but will
have enough variability to make it extremely difficult for a logic-based system
(such as a Robotic Process Automation [RPA] robot) to process.
A good example of a semi-structured document is an invoice. An invoice
will generally have the same information as another invoice, such as the supplier’s name, the date and the total value. But one may have a VAT amount on
it, another may have a sub-total and a total, another may have the address
written in the top right-hand corner rather than the top left, another may
have the supplier’s name written slightly different and another may have the
date written in a different format.
The traditional way of getting the information off the invoice and into the
appropriate ‘system of record’ would be to use an Optical Character
Recognition (OCR) system and a template for each different version of the
invoices (which could run into the hundreds) so that it knew where to find
each piece of data. An AI system, once trained on a sample of invoices, is able
to cope with all of the variabilities I have described. If the address is in a different place it will be found; if there is usually a VAT line and it’s not there
then that doesn’t matter; if the date is in a different format it will be recognised (and converted into the standard format).
Interestingly, the AI works in the opposite way to the template approach—as
more and more ‘versions’ of the document are found, rather than becoming
more uncertain and having to create more and more templates, the AI become
more confident and is able to cope with even more variability. This is because it
is essentially matching the ‘patterns’ of one document with the learned patterns
of the documents used in training.

36

3 AI Capabilities Framework

In the case of unstructured text such as a free-form email, the AI can do two
things. The first is to categorise the text by matching the patterns of words
with what it has already learnt. For example, if you gave the AI a random news
article it could work out whether that article was about politics, business,
sport and so on, as long as it had been trained on many other tagged news
articles. It would do more than simply look for the word ‘football’ to identify
a sports article (there are many business articles about football, for example)
but would instead look at the article holistically, creating an algorithmic
model that represents ‘sports articles’. Any new article with a similar model,
or pattern, would likely be about sports.
The second main thing that an AI search capability can do is to extract
‘named entities’. A named entity could be a Proper Noun—a place or a person, for example—or even a date or quantity. So, in the passage of text,
“Andrew Burgess, who lives in London, wrote his second book in 2017 which
was published by Palgrave Macmillan”, the named entities would be ‘Andrew
Burgess’, ‘London’, ‘second’, ‘2017’ and ‘Palgrave Macmillan’.
There are specific Named Entity Recognition (NER) algorithms that carry
out this task, but they all need to be trained and tuned to make them as accurate as possible. The best NER systems for English currently produce near-­
human performance. Accuracy for NER can be improved by training them on
specific domains of information, so one could be trained on legal documents
and another on medical documents, for example.
The categorisation and entity extraction tasks can be combined to enable
free-form, that is, unstructured, text to be ‘read’ and categorised, and then
have all the meta-data extracted. This would mean that, for example, a customer who emailed a request to a company could have that email categorised
so that it can be automatically forwarded to the right person in the organisation along with all the relevant meta-data that has been extracted. This meta-­
data can be automatically entered into the company’s Case Management
System so that the customer service agent has all the information available
when the case is received.
Search, or Information Extraction, is probably one of the most developed
capabilities in the Framework. There are established software vendors with
relatively mature products, as well as many start-ups all busy building
­applications in this space. As you will find in subsequent chapters, the principal attraction of this capability right now is that it provides a useful complement to RPA; the robots need structured data as their inputs and AI Search
can turn unstructured text into structured data, thus opening up many more
processes that can be automated through RPA.

Clustering

37

Clustering
All the capabilities I have described so far work on transforming unstructured
data (images, sounds, text) into structured data. In contrast, Clustering, our
fourth capability, works on structured data, and looks for patterns and clusters
of similar data, within that data; that is, it is a ‘classifier’. The other aspects of
this capability that are different from the others is that it can (but by no means
has to) learn unsupervised, and it takes a very statistical approach. But just
like the others, it still requires large amounts of data to be truly valuable.
Clustering is usually the first part of a series of stages that usually end with
a prediction, for example extracting insights from new data based on alignment with the original patterns, or identifying anomalous new data where it
doesn’t match the expected patterns. In order to be able to make those predictions or identify the anomalies, the patterns/clusters must first be discovered.
In its simplest form, this type of AI uses statistical methods to find the ‘line
of best fit’ for all the data (in mathematical terms it is minimising the square
of the distance from all the points to the line). It is then just a case of making
those statistical methods more and more complex in order to cope with more
and more features of the data. If there isn’t enough data, then the solution can
suffer from what is called ‘over-fitting’—this is where a theoretical line of best
fit is calculated but it bears little resemblance to any real trends in the data.
The more data you have, the more confident you can be in the patterns that
are found.
A good example of the use of clustering is to be able to identify similar
groups of consumers within customer buying-behaviour data. Humans will
usually be able to identify patterns within small data sets and will often use
past experience to help them shape those patterns. But where there are thousands or millions of data points with multiple characteristics and features,
humans will find it impossible to process all that information, and the AI will
come into its own.
The benefit of the approach, apart from the ability to use sheer computing
power to analyse data quickly, is that the AI is effectively naive to the data. It
is only looking for patterns in numbers, and those numbers could relate to
anything, including a person’s height, their salary, their eye colour, their postcode, their gender, likes, dislikes and previous buying history. If you were in
charge of the loyalty card data for a retailer, you may not have realised that
there is a correlation between eye colour and propensity to buy yoghurt (I’m
making this up) but, if it exists, the AI will find it. No one would have thought
of trying to match these two features and so, if they were using a traditional

38

3 AI Capabilities Framework

Business Information tool, wouldn’t have thought to ask the question in the
first place. It is this insight that AI can bring that provides much of its value.
Just like the other capabilities in the ‘Capturing Information’ group, the
Clustering approaches are reasonably mature and are being used in business—
it is the basis of the field of Predictive Analytics. You only have to buy anything online to see how you have been pattern-matched against other
consumers and then presented with an offer to buy something that they have
bought in the past. Or you may have received a special offer from your mobile
phone provider because some of your usage patterns and behaviours have
indicated you may be a ‘flight risk’ and they want to keep you as a customer.
Or you may have received a call from your bank because of some of your
spending behaviour doesn’t fit into the normal patterns raising the possibility
that you may have had your credit card skimmed.
It is this Clustering capability that benefits more than any of the others
from the sheer propensity of data in the world today.

Natural Language Understanding
Now we are ready to move on to the next type of AI objective—‘determining
what is happening’—and its first capability, Natural Language Understanding,
or NLU.
NLU is an important part of the AI world because it provides some meaning to all the text we are bombarded with, without having to resort to humans
to read it all. It acts as a ‘translator’ between humans and machines, with the
machine having to do the hard work. NLU is closely related to the Search
capability I discussed earlier: some people would actually group Search under
the wider banner of NLP, but I prefer to split them out as they are generally
used for different objectives in business. The same can be said of Speech
Recognition, but again, using these capabilities in the context of the business
setting, they provide different objectives. NLU is still, to a certain degree,
turning the unstructured data of a sentence into the structured data of an
intent, but its primary purposes are to work out what is the right structure
(syntactic analysis) and what is the meaning (semantic analysis) of the words
and sentences.
NLU has a special place in the history of AI because it forms the basis of
the Turing Test. This is the test that English polymath Alan Turing devised in
the 1950s to define whether a machine could be classed as artificially intelligent or not. In the test, an evaluator would have a typed conversation with a
computer and a human, both hidden behind screens—the computer would

Natural Language Understanding

39

pass the test if the evaluator could not tell which of the conversations was with
a computer or a human. The Turing Test is now an object of some debate—it
is not only highly influential, particularly when discussing the philosophy of
AI, but also highly criticised. A number of systems, most notably ‘Eugene
Goostman’, a Russian chatbot, have claimed to have passed the test, but the
real question, now that AI has developed so much further, is whether it is a
valid test of intelligence in the first place.
NLU uses supervised learning with machine learning in order to create a
model of the input text. This model is probabilistic, meaning that it can make
‘softer’ decisions on what the words mean. NLU is a very complex research
field and I won’t begin to go into the details here, suffice it to say that there are
many challenges, such as being able to cope with synonymy (different words
that have similar meanings) and polysemy (words that have several meanings).
The phrase ‘Time flies like an arrow. Fruit flies like a banana.’ perfectly sums
up the challenge that NLU researchers face.
We all see NLU in action most days. Siri or Cortana or Alexa all have
usable natural language interfaces. The majority of the time they will understand the words you are saying (using Speech Recognition) and convert that
into an intent. When these systems don’t perform well it is either because the
words weren’t heard correctly in the first place (this is the most common reason) or the question that has been asked doesn’t make sense. They will be able
to understand different versions of the same question (“What was the football
score?”, “Who won the football match?”, “Please can you tell me the result of
the football?”, etc.) but if you ask “Score?” they will struggle to come up with
a relevant answer. Most systems will have default answers if they can’t work
out what you want.
Chatbots, where the questions and answers are typed rather than spoken,
use NLU but without the ‘risk’ of Speech Recognition getting the word interpretation wrong. They therefore tend to be the simpler and preferred choice
when businesses are looking to create NLU interfaces with their customers,
but the use of speech recognition shouldn’t be discounted if the environment
and benefits are good enough.
And, of course, these personal assistants can talk or type back to you, but
the answers tend to be stock phrases. For more bespoke responses a specific
sub-set of NLP called Natural Language Generation (NLG) is required. NLG,
which can be considered as the ‘opposite’ of NLU, is probably the hardest part
of NLP, which is why it has only recently become commercially available.
Rather than the limited set of response phrases used in the majority of chatbots, creating whole new phrases and even articles is possible with
NLG. Applications include creating hyper-local weather forecasts from

40

3 AI Capabilities Framework

weather data, or financial reports from company’s earnings data or stock market data. Each of these provides a short narrative in a natural language that
can be hard to differentiate from one generated by a person.
NLU is also used to try and understand the emotion behind the sentences,
an area known as sentiment analysis. In its simplest form sentiment analysis
looks for ‘polarity’, that is, is the text positive, negative or neutral. Beyond
that it will look for the type of emotion that is being expressed—is the person
happy, sad, calm or angry, for example. The AI will, of course, look for specific
words within the text but it will try and put these in context, as well as try to
identify sarcasm and other tricky idiosyncrasies of language.
Sentiment analysis is already being used widely to assess online text from
customers, such as through Twitter and TripAdvisor. This allows businesses to
evaluate how their products or services are being perceived in the market
without the expense of focus groups or surveys, and allows them to respond
to potential issues in real time. Specific models, with the associated terminology and jargon, are usually generated for particular needs; understanding how
customers perceive a hotel room will be quite different to how they feel about
their mobile phone.
Another common application for natural language technologies is machine
translation. This is where a computer will translate a phrase from one language to another. It is a complex problem to solve but recent developments in
deep learning have made this much more reliable and usable. The key to success, as with most AI challenges, is the availability of data, and companies like
Google have used the transcriptions from the European Parliament, where the
proceedings are translated by humans between 24 different languages, as its
training set. That means that translating from, say, English to French is now
highly accurate, but for less common languages, such as Khmer, the official
language of Cambodia, translations are generally done through an intermediary stage (usually the English language). Chinese is notoriously difficult for
machines to translate, mainly because it is more of a challenge to edit sentences as a whole—one must sometimes edit arbitrary sets of characters, leading to incorrect outcomes.
The natural evolution of machine translation is real-time translation of
speech (epitomised by the Babel Fish in The Hitchhiker’s Guide to the Galaxy).
Already smartphone apps are able to read and translate signs (using Image
Recognition, NLU and machine translation) and can act as interpreters
between two people talking to each other in different languages. Currently
these capabilities are less reliable than typed translations and only work for
common language pairs, but the potential uses for this technology in a business
environment are huge.

Optimisation

41

NLU is leading the charge for AI in business: it allows us to communicate
with computers in the way we feel most comfortable with—we don’t all have
to be computer wizards to be able to work them effectively. But NLU is inherently a difficult technical challenge and will always be compared with, and
evaluated against, humans (unlike other AI capabilities such as Clustering
which can easily surpass human performance). Therefore, until it can be said
to be indistinguishable from us (i.e. truly pass the Turing Test) NLU will
always be open to criticism. Applied in the right environment and with the
right processes though, it can deliver substantial benefits to businesses.

Optimisation
In all the capabilities I have discussed so far, we have been manipulating data
to transform it from one form to another (images to descriptions, sounds to
words, words to meanings, text to information and big data to meta-data).
But even though the transformed data is much more useful to us than the
source data was, we haven’t yet done anything really meaningful with it. That’s
where the AI capability of Optimisation comes in.
Optimisation is at the heart of what people generally think AI does. It is the
closest analogy we get to mimicking human thought processes without having
to call on true cognitive Understanding.
I use the title ‘Optimisation’ for brevity, but it actually includes problem
solving and planning, which makes it quite a broad subject that has a lot of
science behind it all. A broad definition of this capability would be that if you
know a set of possible initial states as well as your desired goal and a description of all the possible actions to get you there, then the AI can define a solution that will achieve the goal using an optimum sequence of actions from any
of the initial states.
Historically, as I discussed at the start of Chap. 2, optimisation and problem solving were achieved through ‘expert systems’, which were really nothing
more than decision trees which had to be designed and configured up-front
by humans. With the advent of machine learning, much of that design and
configuration is carried out by the AI itself using an extreme version of trial-­
and-­error. (There is still, by the way, a place for knowledge-based AI systems
in the world today, as I shall discuss later.)
A useful way to understand the Optimisation capability is by looking at where
AIs are taught to play computer games. Scientists and researchers use computer
games as a kind of test and benchmark for how clever their systems are: can this

42

3 AI Capabilities Framework

particular AI beat the best human at playing that particular game? But it also
provides a very useful way of explaining what the different flavours are.
So, for example, to play chess and beat Gary Kasparov required brute
force—these systems ‘simply’ analysed every possible move much further
ahead than a human can do. That is more logic than it is AI. To learn and win
at Space Invaders or Breakout, however, needs DNNs and what is called ‘reinforcement learning’ (more of which later). From only being given the objective to maximise the score, and no other information on how to play, these AIs
perceive the state of the screen, trying lots of different approaches until the
score increases (shooting aliens, not getting bombed, etc.). Then, through
reinforcement, they start to develop strategies (such as getting the ball behind
the wall in Breakout) that get higher scores. It’s a great example of self-­learning,
but in a relatively simple environment.
In order to learn the Chinese game of Go, things are a lot trickier. Until
very recently, a computer being able to play Go was a Holy Grail for AI
researchers, mainly because the game relies more on ‘intuition’ than logic, but
also because the number of possible move combinations is trillions of times
greater than a game of chess (there are more possible Go moves than there are
atoms in the universe). In 2016, an AI designed by DeepMind, a British company owned by Google, beat the best Go player in Europe four games to one.
AlphaGo, as the system is called, uses the same concept of reinforcement
learning, this time based on studying 30 million moves from human Go
matches. It then plays different versions of itself thousands of times to work
out the best strategies, which it uses to create long-range plans for the actual
games. Apparently, during the second match, the human player, Lee Sedol,
‘never felt once in control’ (a general fear of AI that is shared by many people,
it must be admitted). Since then, AlphaGo has beaten the world’s best Go
player, Ke Jie, three games to zero.
So, the key characteristic of the Optimisation capability is that there is a
goal to be achieved—an idea that needs to be reasoned, a problem to be
solved or a plan to be made. At its simplest form this is done through iterative
trial-and error—small changes are made to the environment or small actions
are taken, after which the situation is evaluated to see whether it is closer to
the goal or not. If it is closer it carries on and makes further changes, if not it
tries something different.
As I’ve hinted with the gaming examples above, there are some subtleties to
this overly simplified description, which all come together to provide the AI
Optimisation capability. I’ve described a number of the most common
approaches below, most of which are inherently inter-connected.

Optimisation

43

Cognitive reasoning systems are the modern equivalent of the old Expert
Systems (which is why some people are reluctant to describe them as true AI
today). They work by creating a ‘knowledge map’ of the specific domain that
is being modelled, which means that they do not require any data to be trained
on, just access to human subject matter experts. These knowledge maps connect concepts (e.g. food and person) with instances (e.g. chips and Andrew
Burgess) and relationships (e.g. favourite food). Different relationships can
have different weights or probabilities, depending on their likelihood, which
means that the system can be interrogated to make recommendations. Unlike
decision trees, the system can start from any point on the map as long as a goal
has been defined (‘what is Andrew’s favourite food?’, ‘which person likes
chips?’) and can handle much greater complexity: its ‘clever’ bit is in the way
it comes to a recommendation using the shortest route through the map
rather than asking a linear series of questions. Compared to machine learning
approaches, cognitive reasoning systems have the major advantage that the
recommendation that has been made is fully traceable—there is none of the
‘black box’ characteristics that machine learning systems suffer from. This
makes these sorts of systems ideal for regulated industries that ‘need to show
their working’ (Fig. 3.2).
Beyond Cognitive Reasoning, the majority of AI optimisation approaches
are based around algorithms. A key AI concept involves the idea of breaking
a big problem down into much smaller problems and then remembering the
solutions for each one. This approach, known as Dynamic Programming, is
Place
located in

studied at
Establishment

Person
awarded

studied
specialises in

Qualification
Subject

Fig. 3.2 Knowledge map

type

44

3 AI Capabilities Framework

epitomised in the Coin Change Problem, which asks: how can a given amount
of money be made with the least number of coins of given denominations? If
my coins are in units of 1,4,5,15,20 and the total amount I need to make is
23, then a system which looked at it as a series of independent, ‘blind’ decisions would start with the biggest coin first and then add smaller coins until
the total was reached: this would be 20+1+1+1, so four coins. A dynamic
system though would have broken the problem down into smaller problems,
and stored each of the optimal solutions. It would have come up with an
answer of 15+4+4, so just three coins.
The system works by looking ahead at lots of different ways that the end
objective could turn out to be based on, analysing a large sample of the different individual steps that can be taken. Those sample actions that lead to, or
get close to, the desired goal are remembered as favourable and therefore more
likely to be chosen.
A common AI approach that exploits this method is called the Monte
Carlo Tree Search (MCTS). When, for example, playing a game it can ‘play
out’ lots of different versions of potential moves of both players, with each
move creating a branching tree of child moves. Using back propagation (a
form of feedback loop), those moves that achieve the desired goal have the
connections between each of the nodes that led to that result strengthened so
that they are more likely to be played.
These MCTS-like approaches do have a couple of drawbacks: they need to
balance the efficiency of a small sample size of moves with the breadth of covering as many moves as possible (this leads to some randomisation being built
into the choice of child nodes), and they can be slow to converge on to the
optimal solution.
Reinforcement Learning, an area of AI that is seeing a great deal of activity
at the moment, helps to alleviate some of these inefficiencies. Reinforcement
learning is really an extension of dynamic programming (it is sometimes
referred to as ‘approximate dynamic programming’) and is used where the
problems are much more complex than, for example, the coin change problem
or simple board games, and where there are more unknowns about the states
of each move that is taken. Reinforcement learning uses extreme ­trial-­and-­error
to update its ‘experience’ and can then use that machine-learned experience to
determine the most optimal step to take next, that is, the one that will get it
closer to achieving the goal. This approach is slightly different to the supervised learning approaches I described in Chap. 1 in that the ‘right answer’ is
never given in the training, nor are errors explicitly corrected: the machine
learns all of this itself through the trial-and-error iterations.

Optimisation

45

A tactic that is usually employed in reinforcement learning is to have the AI
systems play against each other. DeepMind’s AlphaGo, once it had been
trained on human games, then played itself (or, to be exact, two slightly different versions of AlphaGo played each other) over thousands of games in
order to refine its game even further. If it had only learnt from human games
it could only ever be as good as humans, whereas if it plays itself, it is possible
for it to surpass human skill levels. This combative approach, known as
Generative Adversarial Networks, as well as allowing the algorithms to tune
themselves, is a great way to run simulations to test the viability of AI systems
and to run different war-gaming scenarios.
Another ‘weakness’ in AI Optimisation approaches has been that the systems tend to look at the short-term gains at the expense of the longer-term
strategy. Recent advances in AI have seen a number of different systems combined to provide both these perspectives, especially where the domain is particularly complex: a ‘policy’ algorithm will look at the next best move, whilst
a ‘value’ algorithm will look at how the problem, or game, might finish up.
The two algorithms can then work together to provide the best outcome.
Facebook has already been able to train chatbot agents to negotiate for
simple items, in some cases doing it as well as humans. Facebook Artificial
Intelligence Research first used supervised learning to train the agents on a
large number of human-to-human negotiation scripts. This stage helped them
imitate the actions of the humans (mapping between language and meaning),
but didn’t explicitly help them achieve their objectives. For this they used
reinforcement learning, where two AI agents would ‘practise’ negotiating with
each other (interestingly they had to fix the language model in place at this
stage because they found that, if the agents were allowed to continue learning
the language elements, they would start to create their own private language
between them). At the end of each negotiation, the agent would be ‘rewarded’
based on the deal it had managed to achieve. This was then fed back through
the model to capture the learning, and make the agent a better negotiator.
Following training, it could then negotiate for the same types of items with
humans, matching the capabilities of other human negotiators.
Similar ideas of combining a number of AI techniques were used when an
AI system, called Libratus, was able to beat experienced players of poker—this
system used three different types of AI: the first relied on reinforcement learning to teach itself the game of poker from scratch; a second system focused on
the end game, leaving the first system to concentrate on the immediate next
moves; and, because some of the human players were able to detect trends in
how the machine bet, a third system looked for these patterns overnight and
introduced additional randomness to hide its tracks.

46

3 AI Capabilities Framework

Optimisation AI can be applied to many situations where there is a specific
goal to be achieved, such as winning a hand of poker or negotiating for items.
It can, as I mentioned earlier, provide performance that exceeds human capabilities. Other typical uses for the Optimisation capability include route planning, designing rota for shift workers (such as nurses) and making
recommendations.
All the approaches described in this section are just a sample of how AI can
attempt to solve problems, but should give you a feel for the general strategies
that can be employed. At the heart of it is the idea that big decisions can be
broken down into many small decisions, and that these can then be optimised
using trial-and-error so that a defined goal can be achieved, or a specific
reward maximised.

Prediction
Prediction employs one of the core ideas of AI in that it uses lots of historical
data in order to match a new piece of data to an identified group. Thus, prediction generally follows on from the Clustering capability described earlier in
the chapter.
I’ve already mentioned one of the more common uses of prediction, that of
recommending related online purchases (‘you bought this book therefore you
will probably like this other book’). So, in this case, what the retailer calls a
‘recommendation’ is actually a prediction they make in order to sell you more
stuff.
Some decisions can be described as predictions—if you apply for a loan
that is assessed by a machine, then the AI will try and predict whether you will
default on the loan or not. It will try and match your profile (age, salary, regular spending, other loans, etc.) with other customers that have similar profiles.
If those other matching customers generally default on their loans then you
will likely be refused the credit line.
Prediction can also follow on from the Search capability. Because Search
can look for patterns within text, it can match these with patterns from some
pre-defined requirements. For example, CVs (aka résumés) can be matched
with job descriptions in order to predict good candidates for the role.
The prediction capability differs from Optimisation in that it doesn’t have
a specific goal to achieve. There are no steps to determine to achieve a specific
objective—we are ‘just’ matching a new data point with our historical data.
At a simple level, predictions can be made from just a few ‘features’—these are
the specific characteristics that are being measured, such as the number of bedrooms

Prediction

47

in a house and the size of the garden. If you had a table of this data and the actual
price of each of the houses, then you could use this capability to predict the price of
another house if you knew the number of bedrooms and size of the garden. And
you wouldn’t necessarily need AI to do this, or even a computer.
But there are generally more factors (or features) which determine the price
of a house than just these two things. You might also consider how many stories
it has, whether it is detached, semi-detached or terraced, whether it has parking,
a utility room, a swimming pool, its postcode and so on. All this makes it much
more difficult to predict the house value without using some AI ‘magic’.
Clearly, the more features that are considered, the more training data needs
to be used, so that as many ‘variations’ of all the features can be captured and
built into the model. This model will have considered all the different influences (or ‘weights’) that each of the features has on the house price. By entering in the values of a selection of features of a different house, the model fits
(using Regression Analysis in this case) those features as closely as possible and
can therefore predict the house price. It is normal for the system to give a
figure for its level of confidence (i.e. the probability that it is correct).
Once the number of features is in hundreds or thousands then things become
more complex—this just means that more algorithms, more training data and
more computing power are needed. At the extreme end of this, the systems that
are used to predict our weather are some of the most powerful in the world.
An important aspect of the Prediction capability is to remember that it is
essentially naive. By this I mean that it is only manipulating numbers and has
no underlying understanding of what that data actually means. Those house
features could be car features or weather features or people features—to the
computer they are just numbers.
This naivety, whilst seeming like a virtue, is also one of Prediction’s biggest
challenges: that of unintended bias. Normally AI is lauded because it is not
influenced by the prejudices that humans inherently suffer from (there are
many studies that show that, although people may claim not to be biased in,
say, recruiting staff, there is usually some sort of unconscious bias involved).
But the unbiasedness of AI is only as good as the data that is used to train it—if
the training set contains CVs and hiring decisions that have been made by
humans, the biases that were in the original data will be ‘trained into’ the AI.
Another key issue for AI Prediction is the opaqueness of the decision-­making
process. When an AI makes a prediction, for example that a person is more than
likely to default on a loan and therefore their application should be rejected, that
prediction is based on all the training data. The training data has been used by the
machine to build an algorithmic model that it then uses to predict new cases from.
For the more complex algorithmic models this is just a matrix of numbers that

48

3 AI Capabilities Framework

makes no sense to a human trying to read it, and therefore the reason (or, more
likely, reasons) that the person had their loan refused is not easily available to scrutinise. (Simpler algorithmic models, such as Classification and Regression Trees,
do provide some transparency and are therefore more popular). Of course, we
don’t need to understand the workings of every prediction that is made (it would
be interesting to know why the house price predictor came up with that valuation,
but it is the actual house price we are most concerned with), but some industries,
especially those that are regulated, will require it. Also, if you are the person that
has been refused the loan, you should have the right to know why.
AI is being used to predict many things, including, in some US courts, a
defendant’s risk profile. When the consequences of those predictions could
have an influence on whether someone is found innocent or guilty, then the
challenges I have described above become very serious indeed. As Melvin
Kranzberg’s first law of technology succinctly states: technology is neither
good nor bad; nor is it neutral. These challenges of ‘algorithmic transparency’,
naivety and unintended bias will be discussed in more detail in Chap. 8.
AI Prediction is one of the most active areas in the field at the moment.
Where there are lots of good data, predictions can generally be made from
that data. That doesn’t necessarily mean that predictions need to be made or
that they will be useful in any way, but there are lots of cases where this can be
very beneficial indeed, including predicting valuations, yields, customer
churn, preventative maintenance requirements and demand for a product.

Understanding
I include this section on AI Understanding in the book only as a way to
describe what is not currently available in the AI world to businesses, or to
anyone outside of a research lab. It should be seen as a countermeasure to all
of the hype that can be put out by over-excited marketing departments.
By ‘understanding’, I am generally referring to the ability of a machine to
have conscious awareness of what it is doing or thinking (or to act like it
does—see next paragraph). This implies it can understand the intent and
motivations of people, rather than just blindly crunching numbers about
them. It is usually described by the concept of Artificial General Intelligence
(AGI), where the AI is able to mimic all of the capabilities of the human
brain, not just the very narrow ones we have discussed so far.
There is an interesting subtlety to the description of AGI. John Searle, a
philosopher, differentiated ‘strong AI’ and ‘weak AI’. A strong AI system can
think and have a mind, whereas a weak AI system can (only) act like it thinks

Understanding

49

and has a mind. The former assumes that there is something special about the
machine that goes beyond the abilities that we can test for. Ray Kurzweil, the
futurologist, simply describes strong AI as when the computer acts as though
it has a mind, irrespective of whether it actually does. For our purposes, we
will stick to this definition as we are dealing with practicalities rather than
philosophical debate.
So, what sort of tests would we use to claim strong AI? I have already mentioned the Turing Test as a (limited) test of AI. Other tests that people have
proposed are:
• The Coffee Test (Wozniak)—A machine is given the task of going into an
average home and figuring out how to make coffee. It has to find the coffee
machine, find the coffee, add water, find a mug and brew the coffee by
pushing the proper buttons.
• The Robot College Student Test (Goertzel)—A machine is given the task
of enrolling in a university, taking and passing the same classes that humans
would, and obtaining a degree.
• The Employment Test (Nilsson)—A machine is given the task of working
an economically important job, and must perform as well or better than
the level that humans perform at in the same job.
No AI systems are anywhere near passing these tests, and any systems that
are even close will be a combination of many different types of AI. As I’ve
described in the previous sections of this chapter, there are a number of different types of AI capabilities, each one very specialised in what it does. This
means, for example, an AI capability that is being used to recognise images
will be useless at processing language. This is the concept of Artificial Narrow
Intelligence (ANI). Even within the capability groups I have described there
is little or no crossover between specific uses—if I have a system that extracts
data from invoices, it will not be able to do the same for remittance advices,
without training the system from scratch. The same could even be true if I
wanted to take a trained invoice extractor from one business to the another—
there may be enough variation between businesses to mean that retraining
would be required.
Where our brain is so much cleverer than AI is where it is able to use different cognitive approaches and techniques in different situations and, importantly, take learnings from one situation and apply them to a completely
different one. For example, I may know that the value of a house generally
increases with the number of bedrooms it has, but I can then apply the same
concept to other objects (computers with more hard drive capacity cost more)

50

3 AI Capabilities Framework

but also know that this shouldn’t be applied to everything (cars with more
wheels are generally not more expensive). AI is not able to do this right now.
There is work being done to try and create AGI. At a national scale, there
is the Blue Brain Project in Switzerland (which aims to create a digital reconstruction of the brain by reverse-engineering mammalian brain circuitry), and
the BRAIN Project in the United States, which is also looking to model real
brains. Organisations such as OpenCog (an open-source AGI research platform), the Redwood Center for Theoretical Neuroscience and the Machine
Intelligence Research Institute are all at the forefront of researching various
aspects of AGI.
Interestingly, there has been some progress in getting neural networks to
remember what they have previously been taught. This means they could, in
theory, be able to use learnings from one task and apply it to a second task.
Although that may sound like a simple thing for a human to do, ‘catastrophic
forgetting’ (when new tasks are introduced, new adaptations overwrite the
knowledge that the system had previously acquired) is an inherent flaw in neural networks. DeepMind, the UK AI company owned by Google, is developing
an approach it calls Elastic Weight Consolidation which allows the algorithm to
learn a new task whilst retaining some of the knowledge from learning a previous task (they actually use different Atari computer games as their test tasks).
This is showing promise but is a long way from being put to practical use.
Despite the research and small steps that are being taken, the ability for a
computer to fundamentally understand what it is doing is still a long way off
(some argue that it could never be possible). Even using Searle’s definition of
weak AI, there are major hurdles to overcome. Some of these are very technical (such as the challenge of catastrophic forgetting) and some are simply
down to the fact that the necessary computing power doesn’t exist at the
moment. But, as described throughout this book, there are huge advancements being made in ANI that provide real benefits to people and businesses.
Some of these make our lives simpler whilst others are able to far exceed our
own competencies in specific tasks. By understanding each of these capabilities, and their respective limitations, we are able to benefit from AI technologies today and tomorrow.

Using the AI Framework
The AI Capability Framework is my attempt to bring some clarity and order
to the diverse, complex and often confusing field of AI. By ‘boiling it down’
into a set of discrete capabilities, AI hopefully becomes accessible to those

Using the AI Framework

51

who want to benefit from the technology but do not have the skills, ­experience
(or desire) to understand the technical aspects beyond the high-level appreciation I have provided in this chapter.
I have tried to differentiate each of the capabilities as much as possible, but
there are inevitably some overlaps between each of them, for example between
speech recognition and NLU, and clustering and prediction. The boundaries
between these are blurry; some people would describe both speech recognition and NLU as a sub-category of NLP but I think they sit better as discrete
capabilities; and some people might separate out planning and optimisation,
but I think they are close enough aligned to keep them as one. So, please don’t
get too hung up on some of the nuances—AI is a complex subject, with many
different viewpoints and opinions, and is constantly changing. The framework should be treated as your guide rather than a technical manual (Fig. 3.3).
So, the knowledge you now have should enable you to do three things:
• Identify the right AI capabilities for your business need. What are your
business objectives and could AI provide all or part of the solution? Are you
looking to capture information or understand what is happening, or both?
Do you want to replace existing capability (computer or human) with AI,
or do you want to augment their capabilities further? Which specific capabilities will you need to create a solution? Will that require a supervised
learning approach, and what data do I have available?

CAPTURE INFORMATION

structured
data

unstructured
data

Image Recognition

WHAT IS HAPPENING?

WHY IS IT HAPPENING?

Natural Language
Understanding

Speech Recognition

Search

Optimisation

Data Analysis /
Clustering

Prediction

Fig. 3.3 The AI framework

Understanding

52

3 AI Capabilities Framework

• Look beyond the hype. What capabilities do the AI vendors actually have,
rather than those they claim they have? How do these capabilities match
my requirements? Do their assertions about their product make sense? Are
there gaps that will need to be filled with different AI solutions or traditional technology? Will it be as easy to implement as they claim?
• Be realistic. What are the limitations of the AI capabilities that I might need?
Is AI the most appropriate solution for this or are there simpler or more effective solutions? Do I have the necessary data available to train the system
adequately? Will I need to bring in external support to help me get going?
Chapter 5 provides real examples of the different AI capabilities being used
in businesses today. They are focused on common strategic objectives (enhancing customer service, optimising business processes and generating insights)
and combine different capabilities to achieve these. Each of the examples references the capabilities that are being used and, where appropriate, what challenges and limitations were experienced by those responsible for delivering
the solutions.
Some of the approaches and examples that will be described required additional technologies (such as cloud computing and RPA) to enable them to
extract their full value. The following chapter outlines what these associated
technologies are and why they are such a good fit with AI.

The AI Start-Up’s View

53

The AI Start-Up’s View
This is an extract from an interview with Vasilis Tsolis, founder and CEO of
Cognitiv+, an AI start-up in the legal sector.
AB What got you into the world of AI in the first place?
VT	My personal background is not an obvious path for a co-founder of an AI
start-up. I am a chartered civil engineer who switched my career by
undertaking a law degree. After graduation, my career was moving
between law, engineering and commercial management in various sectors
including infrastructure, construction and energy. After a few years,
something became very clear to me: people spend substantial amounts of
time reading contracts, and most of the time there is a tremendous repetition of those tasks. And that is where the problem lies: rather than focus
on advanced solution discovery, professionals consume their daily lives
with data gathering. The solution was an obvious one; using AI will free
the professionals to focus on what they are good at. That is why we started
Cognitiv+ in 2015, which has been a fantastic trip so far.
AB What is it like building an AI start up at the moment?
VT	This is a period where people apply AI or attempt to use AI in almost all
aspects of our lives, where creativity sometimes trumps technology. There
are major projects in process, such as autonomous cars and it very clear
now to everyone that AI is here to stay. It is a very exciting period, but
there are quite a lot of challenges towards the road of maturity.
Working with a number of professional clients, we see numerous
opportunities and dozens of use cases, some possible now, some reachable
within the mid-term future, some look close to science fiction.
Technology advancements advertise extraordinary abilities and this goes
throughout the food chain, where clients are constantly bombarded with
innovative ideas where the success rate is variable. This could be because they
are early in their roadmap and suffer teething problems, or because of a lack
of technical or subject matter depth which is sometimes not an easy fix.
As with any early adoption, it is the clients and investors that get the
pressure to isolate the true technologists and innovators with clear business skills, and some in-house knowledge on how technology works will
definitely bring certainty and a quick return on investment.
AB So what value do your customers get out of your software?
VT	Reading text takes time, even Romans were complaining about this and
AI can help that. Cognitiv+ targets legal text, contracts and the regulation
corpus.
Using AI should be seen as an extrapolation of automation, a technology that improves our lives, makes things faster, simpler and enables us to
explore tasks in a way that was not possible before. For example, clients
can use our software to analyse contractual risk on their whole portfolio

54

3 AI Capabilities Framework

within hours in a way that was never possible before. Time-consuming
tasks can be delegated and free up professionals’ schedules, where technology can provide them with holistic views of their third-party risk to
levels that have been unprecedented.
AB	What do customers or potential customers need to focus on if they are
going to get the maximum value from AI?
VT	While many companies hit the ground by starting to write code, there is a
lot of preparatory work that projects need to undertake and ensure success.
The first focus will be the data. Data scientists require massive data sets
and the new techniques, supported by hardware innovation, command
new levels of data quantities. The good news is that we all produce more
data every day, but this is not necessarily for all the aspects of professional
services. Some data custodians can move faster to this gold rush because
they have better access to data from others.
Quantity is not enough though—it is also the quality that will determine if the data can be used. But this is a synthetic challenge that can be
tackled only if it is isolated and understood on a case-by-case basis.
Second, teams need to make an early decision on what success would
look like, making business and technology targets that are reachable.
The fact that algorithms provide answers that are probabilistic and not
deterministic makes the way we manage AI projects very different from
any previous IT transformation projects companies have participated in.
Thirdly, an important focus would be the talent and skills diversity that
a team has. A sensible balance between business subject matter experts,
data scientists and coders would determine how successful a project will be.
AB How do you see the market developing over the next 5 years or so?
VT	While I can predict where we are heading in the next 5 months, it is practically impossible to hit anything close to a 5 year prediction. The reason
is simple—all the parameters are rapidly changing: the type of data we are
surrounded with, the size of data sets, the maturity of the algorithms,
hardware improvement, the list goes on.
But here is one prediction: as we have started using various NLP, NLU
and machine learning techniques to interpret text and documents, it
seems that certain types of text, certain sources, can be analysed better
than others. People will pick this up and will write in a way that can be
processed and summarised by machines. Why do that? It is for the same
reason where we think twice what tags we are using for our blog articles
as this will enable Search Engine Optimisation (SEO) bots to categorise
our text accordingly and disseminate it to the right channels.
Typically, we write in a way that our audience will understand us—my
take is that we will continue to do that—the difference is that in the
crowd or readers will also be some bots. What might this look like?
Perhaps, simpler and shorter phrases where the object and subjects are
clear and NLP algorithms can consume them much better?

4
Associated Technologies

Introduction
Artificial Intelligence can do a lot of things, but it can’t do everything. Quite
a lot of the time implementing a stand-alone AI solution will satisfy whatever
objectives you are looking to achieve. Sometimes AI will rely on other technologies to make it work well, and other times it will complement them so
that they both work better. This chapter covers some of the associated technologies that anybody looking to implement AI will need to consider.
Some of these technologies are software-based, such as RPA. This is a relatively new technology that automates rules-based processes but struggles to
handle unstructured data and any decision making. Cloud Computing, a
combination of software and hardware capabilities, is a key enabler of AI, and
many of the AI applications that we are seeing now would not be possible
without it.
There are also some hardware technologies that enable, and can be exploited
by, AI. Physical robots (as opposed to the software-based RPA) can be made
more intelligent with the application of AI, and the Internet of Things (IoT)
can provide very useful data sources for AI systems. One of the technologies
that I have included in this chapter isn’t really a technology at all, but is powered by human beings: Crowd Sourcing enables data to be tagged and cleaned
in an efficient and flexible way, and therefore is extremely useful to AI
developers.
I have not included more common enterprise systems, such as Enterprise
Resource Planning (ERP) and Customer Relationship Management (CRM)
systems, in this summary. All could be considered sources of data, and some
© The Author(s) 2018
A. Burgess, The Executive Guide to Artificial Intelligence,
https://doi.org/10.1007/978-3-319-63820-1_4

55

56

4

Associated Technologies

claim to have AI capabilities built into them (e.g. email systems can be a good
data source and also use AI to identify spam messages), but overall, they have
only general associations with AI.

AI and Cloud
Cloud Computing is where a network of remote servers, hosted on the
Internet, are used to store, manage and process data, rather than the ‘traditional’ approach of using a local (on-premises) server or a personal computer
to do those tasks.
Because of these abilities to store, manage and process data away from the
user’s device (PC, mobile ‘phone, etc.) and instead on high-performance,
high-capacity specialised servers, Cloud Computing has become almost an
integral part of how AI systems operate today. As cloud technology itself
matures, the two will become inextricably linked—many people already talk
of ‘Cloud AI’ as the Next Big Thing.
One of the more straightforward applications of cloud combined with AI
is the making available of large, public data sets. Most AI developers, unless
they are working for large corporates, don’t have their own data sets to train
their systems, but instead rely on these public databases. As I mentioned in
the Big Data section of Chap. 2, there exist quite a few of these datasets, covering subjects such as Computer Vision, Natural Language, Speech,
Recommendation Systems, Networks and Graphs, and Geospatial Data.
But the cloud provides more than just access to data—more often it will
actually process the data as well. We can experience this as consumers every
time we use a service like Amazon’s Echo. Although there is a small amount of
processing power on the device itself (mainly to recognise the ‘wake word’),
the processing of the words into meaning (using Speech Recognition and
NLU) is done by software on Amazon’s own servers. And, of course, executing
the actual command is also done on the cloud, perhaps sending the instruction back to your home to, say, turn the kitchen light on.
On an enterprise scale, even more of the processing work can be done on the
cloud, on huge server farms full of powerful machines. The biggest challenge
to this model though is that the data generally needs to sit on the cloud as well.
For enterprises with huge amounts of data (perhaps petabytes’ worth) then
transferring all of this to the cloud can be impractical. As a solution to very
long upload times, Amazon actually offers a huge truck (called the Amazon
Snowmobile) that contains 100 petabytes of computer storage—the truck is
driven to your data centre and plugged in so that the data can be uploaded to

AI and Cloud

57

the truck. It then drives back to the Amazon Data Centre and downloads it all.
But, as network speeds improve, there will be less and less need for these sort
of physical solutions.
Another challenge to Cloud AI is that it can be perceived as risky to store
data off-premises, especially if the data is confidential, such as a bank’s customers’ details. The security aspects can be covered by the type of cloud service that is procured—the best suppliers can now offer security provisions that
are as good as, or better than, hosted or on-premises solutions. The UK
Government in fact released guidance at the start of 2017 that states it is possible for public sector organisations to safely put highly personal and sensitive
data into the public cloud.
The cost, though, of setting up an operation that can store, manage and
process AI data effectively, securely and economically means that the market
for Cloud AI is currently dominated by a small number of suppliers, namely
Amazon, Google and Microsoft. These companies actual offer a complete set
of AI services, including access to ready-to-use algorithms.
Generally, the AI Cloud offerings are made up of four main areas (I’ve used
Amazon’s model as a basis for this description):
• Infrastructure—this consists of all the virtual servers and GPUs (the processor chips) that are required to house the applications that train and run
the AI systems
• Frameworks—these are the AI development services that are used to build
bespoke AI systems, and tend to be used by researchers and data scientists.
They could include pre-installed and configured frameworks such as
ApacheMXNet, TensorFlow and Caffe.
• Platforms—these would be used by AI developers who have their own data
sets but do not have access to algorithms. They would need to be able to
deploy and manage the AI training as well as host the models.
• Services—for those who don’t have access to data or algorithms, the AI
Services offer pre-trained AI algorithms. This is the simplest approach for
accessing specific AI capabilities with a minimum knowledge of how they
work technically.
The ability to tap into relatively complex, pre-trained algorithms is a boon
for anyone wanting to build AI capability into their applications. For example, if you want to build an AI application that has some NLU capability
(a chatbot, for example) then you could use Amazon’s Lex algorithm,
Microsoft’s Linguistic Analysis or Google’s (wonderfully named) Parsey
McParseFace. Each of these is a simple Application Programming Interface

58

4 Associated Technologies

(API), which means that they can be ‘called’ by sending them specific data.
They will then return a result to you, which can be read by your application.
One of the interesting things about these services is that they are either free
or cheap. Microsoft’s Language Understanding Intelligent Service (LUIS) offering is currently based on a threshold of API calls per month—below this
threshold it is free to use, and every thousand calls above that is charged at just
a few cents. Other algorithms are charged as a monthly subscription.
Many businesses today are taking advantage of Cloud AI rather than building their own capabilities. For example: a brewer in Oregon, USA, is using
Cloud AI to control their brewing processes; a public TV company uses it to
identify and tag people that appear on its programs; schools are using it to
predict student churn and an FMCG company uses it to analyse job
applications.
As I’ve discussed in Chap. 2, one of the motivations for these companies to
offer their AI technologies for virtually nothing is that they get access to more
and more data, which in many ways is the currency of AI. But despite the
‘greedy corporate’ overtones, Cloud AI does have the feeling of democratisation about it, where more and more people have simple and cheap access to
these very clever technologies.

AI and Robotic Process Automation
Robotic Process Automation (or RPA) describes a relatively new type of software that replicates the transactional, rules-based work that a human being
might do.
For clarity, it is important to differentiate between RPA and the traditional
IT systems. RPA—at its most basic level—utilises technology to replace a
series of human actions (which is where the ‘robot’ terminology comes in).
Correspondingly, not all technologies provide automation, and replacing a
single human action with technology (e.g. a mathematical equation in a
spreadsheet) is not considered RPA. Similarly, much automation is already
embedded into software systems (e.g. linking customer information across
finance and procurement functions), but since it is part of the normal feature
functionality of a system, it is generally not considered as RPA, but simply a
more powerful system(s).
In an ideal world, all transactional work processes would be carried out
by large, all-encompassing IT systems, without a single human being
involved. In reality, whilst many systems can automate a large part of specific processes and functions, they tend to be siloed or only deal with one

AI and Robotic Process Automation

59

part of the ­end-to-­end process (think of an online loan application process
that will have to get data from, and input data to, a web browser, a CRM
system, a credit checking system, a finance system, a KYC system, an address
look-up system and probably one or two spreadsheets). In addition, many
businesses now have multiple systems that have been acquired as point solutions or simply through numerous mergers and acquisitions. The default
‘integration’ across all of these systems, tying the whole end-to-end process
together, has traditionally been the human being. More often than not,
these human beings are part of an outsourced service.
Robotic Process Automation can replace nearly all the transactional work
that the human does, at a much lower cost (as much as 50% lower). The RPA
systems replicate (using simple process mapping tools) the rules-based work
that the human does (i.e. interfacing at the ‘presentation layer’) which means
that there is no need to change any of the underlying systems. A single process
could be automated and be delivering value within a matter of weeks. The
‘robot’ can then continue processing that activity 24 hours a day, 7 days a
week, 52 weeks a year if required, with every action being completely auditable. If the process changes, the robot only needs to be retrained once, rather
than having to retrain a complete team of people.
So, to give a simple example, if a law firm is managing a property portfolio
on behalf of a client, they would be expected to carry out Land Registry
checks at some point. This is commonly a paralegal role that might involve
the person getting a request from a lawyer, or from the client directly, probably via a template, an email or a workflow system. The person would read the
relevant information off the form, log into the Land Registry site, enter the
information into the site and read the results that came back from the search.
They would then transpose that information back onto the form and respond
to the initial request. It is possible for this whole process to be handled by a
software ‘robot’ without the need for any human intervention.
Although this is a very simple example, it does demonstrate some of the
benefits of RPA:
• the cost of the robot is a fraction of the cost of the human (between a third
and a tenth);
• the robot works in exactly the same way as the human would, so no IT or
process changes are required;
• once trained, the robot will do the process exactly the same way 100% of
the time;
• every step that the robot takes is logged, providing full auditability;

60

4 Associated Technologies

• the robot can carry out the process in the middle of the night or over a
weekend if necessary and
• the robot will never be sick, need a holiday or ask for a pay rise.
What this means is that wherever there are processes that are rules-based,
repeatable and use (or could use) IT systems, the person doing that process
can be replaced by a software robot. Some further examples of processes that
can be automated are:
•
•
•
•
•
•

Employee on-boarding
Invoice processing
Payments
Conveyancing processing
Benefit entitlement checks
IT Service Desk requests

These are just a small sample of the sorts of processes that can be automated
through RPA. Any comprehensive review of the processes carried out in a
back office environment can identify a large number of candidates for automation. Some examples of where RPA has demonstrated particular benefits
include:
• O2 replacing 45 offshore employees, costing a total of $1.35m a year, with
ten software robots, costing $100,000. Example processes included the
provisioning of new SIM cards. The telecoms firm then spent its savings of
$1.25m on hiring 12 new people to do more innovative work locally at its
headquarters.
• Barclays Bank have seen a £175 million per annum reduction in bad debt
provision and over 120 Full Time Equivalents (FTE) saved. Example processes include:
–– Automated Fraudulent Account Closure Process—Rapid closure of
compromised accounts
–– Automated Branch Risk Monitoring Process—Collation and monitoring of branch network operational risk indicators
–– Personal Loan Application Opening—Automation of processes for
new loan applications.
• Co-operative Banking Group has automated over 130 processes with
robotic automation including complex CHAPs processing, VISA
­charge-­back processing and many back office processes to support sales and
general administration.

AI and Robotic Process Automation

61

As well as delivering cost savings, RPA is having a huge impact on the way
companies are organising their resources: shared service centres are ripe for
large-scale automation, and outsourced processes are being brought back in-­
house (on-shore) because the costs and risks are much more favourable after
automation. (This, by the way, is creating a huge threat to the viability of
Business Process Outsourcing (BPO) providers.
In implementing RPA, there are a number of aspects that need to be considered. Firstly, there is a need to understand whether the robots will be running ‘assisted’ or ‘unassisted’. An assisted robot generally works on parts of
processes and will be triggered to run by a human. In a contact centre, for
example, a customer service agent could take a call from a customer who
wishes to change their address. Once the call is complete, the agent can trigger
the robot to carry out the changes, which may be required across a number of
different systems. Meanwhile the human agent can get on with taking another
call. Unassisted robots work autonomously, triggered by a specific schedule
(e.g. every Monday morning at 8) or an alert (a backlog queue is over a certain
threshold). They will generally cover whole processes and are therefore more
efficient than unassisted robots. Different RPA software packages are suited to
these two different scenarios in different ways.
Once the type of robot has been selected, the candidate processes for automation can be considered. There are a number of characteristics that make for
a good candidate process:
• Rules-based, predictable, replicable—the process needs to be mapped and
configured in the RPA software; therefore, it has to be definable down to
key-stroke level.
• High volume, scalable—in most cases a process that is high volume (e.g.
happens many time a day) will be preferable because it will deliver a better
return on the investment.
• Relies on multiple systems—RPA comes into its own when the process has
to access a number of systems, as this is where people are generally employed
to integrate and move data around between them.
• Where poor quality results in high risk or cost—the exception to the high-­
volume criteria is where a low-volume process might have a large risk associated with it, such as in making payments, and where compliance and
accuracy are the main concerns.
These criteria provide plenty of opportunities for RPA in most large businesses. But they do have some limitations. The reason that RPA software is of
particular interest to exponents of AI is that, although the RPA software is

62

4 Associated Technologies

really clever in the way that it manages processes, the robots are effectively
‘dumb’; they will do exactly what they are told to do, with unwavering compliance. In many situations that is a good thing, but there are situations where
there is ambiguity, either in the information coming in, or in the need to
make a judgement. This is where AI comes to the fore.
One of the biggest constraints of automating processes through RPA is that
the robots require structured data as an input. This could be, for example, a
spreadsheet, a web form or a database. The robot needs to know precisely
where the required data is, and if it’s not in the expected place then the process
will come abruptly to a halt. Artificial Intelligence, and particularly the Search
capability, provides the ability to have unstructured, or semi-structured,
source data transformed into structured data that the robots can then
process.
Examples of semi-structured data would include invoices or remittance
advices—the information on the document is generally the same (supplier
name, date, address, value, VAT amount, etc.) but can vary considerably in
format and position on the page. As described in the previous chapter, AI
Search is able to extract the meta-data from the document and paste it into
the system-of-record even though every version of it may look slightly different. Once in the main system, the robots can use the data for subsequent
automated processing.
The robots can even use the AI output as trigger for them to run. For
example, a legal contract can be considered a semi-structured document (it
has some common information such as name of parties, termination date,
limits of liability, etc.). An AI Search capability can extract this meta-data for
all of a business’s contracts so that they can manage their total risk portfolio.
RPA robots could be triggered if there was a change in regulation and all contracts of a specific type (e.g. all those under English and Welsh Law) needed
to be updated.
Another area where RPA falls down is where judgement is required as part
of the process. For example, in processing a loan application, much of the
initial stage (such as taking the information from the applicant’s web form
submission and populating the CRM system, the loan system and carrying
out a credit check) can be automated through RPA. But at some point there
needs to be a decision made whether to approve the loan or not. If the decision is relatively simple this can still be handled by RPA—it would be a case
of applying scores to specific criteria with certain weightings, and then
­checking whether the total score was below or above a threshold. But for more
complex decisions, or where it seems like ‘judgement’ is required, then the AI
prediction capability can be used. (This could either be through a Cognitive
Reasoning engine or using a Machine Learning approach.)

AI and Robotic Process Automation

63

Thus, using a combination of RPA and AI allows many processes to be
automated end to end, which inherently provides greater efficiency benefits
than would partly automated ones.
Conversely, RPA can also help AI automation efforts. As mentioned above,
robots are very good at extracting and collating data from many different
sources. Therefore, RPA can be used as a ‘data supplier’ for AI systems. This
could also include manipulation of the data (e.g. remapping fields) as well as
identifying any unusable (dirty) data.
So far I have mainly focused on automating business processes, but RPA
can also be used to automate IT processes as well. The concept for automating
IT is exactly the same as for business processes: replace human staff doing
rules-based work with software agents. Many of the tasks carried out by an IT
Service Desk can be automated; common examples include password resets
and provisioning of additional software on a user’s desktop. One user of RPA
reported a reduction from 6 minutes to 50 seconds for the average incident
execution time for their Service Desk.
RPA can also work autonomously on infrastructure components, triggered
by system alerts: a robot could, for example, reboot a server as soon as it
received an alert that the server was no longer responding. Just like business
process automation, the robots can access almost any other system, and there
will be no disruption or changes required to those underlying systems. RPA in
this scenario can be considered a ‘meta-manager’ that sits across all the monitoring and management systems.
For IT automation, AI can enhance the capabilities of RPA. By using
Optimisation and Prediction capabilities they can be trained from run-books and
other sources, and will continue to learn from ‘watching’ human engineers. The
AI systems can also be used to proactively monitor the IT environment and its
current state in order to identify trends as well as any changes to the environment
(a new virtual server, for example) and then adjust their plans accordingly.
Companies that have implemented a combination of RPA and AI solutions
have seen, for example, 56% of incidents being resolved without any human
intervention, and a 60% reduction in the meantime to resolution.
So, RPA provides a useful complementary technology to AI. In addition to
AI enabling more processes, and more of the process, to be automated, RPA also
aids the data collation efforts that AI often needs. The two technologies also
work well together when there are larger transformational objectives in mind.
Enabling a self-service capability in a business can be a good way to improve
customer service at the same time as reducing costs. This could be achieved
through a combination of AI systems managing the front-end customer engagement (e.g. through chatbots) and RPA managing the back-end processes.

64

4

Associated Technologies

AI and Robotics
One of the first practical applications of AI was in a physical robot called
SHAKEY, which was designed and built by Stanford Research Institute between
1966 and 1972. SHAKEY used computer vision and NLP to receive instructions which it could then break down into discrete tasks and move around a
room to complete the objective it had been set. Although we would now see it
as very rudimentary, it was, at the time, at the forefront of AI research.
Today we have robot vacuum cleaners costing a few hundred pounds that
can clean your house autonomously, and, although they can’t respond to voice
commands, this would be relatively easy to implement. Other examples of
physical robots that use AI are
• Autonomous vehicles—driverless cars and trucks use AI to interpret all the
information coming in from the vehicle’s sensors (such as using computer
vision to interpret the incoming LIDAR data, which is like RADAR but for
light) and then plan the appropriate actions to take. All of this needs to be
done in real time to ensure that the vehicle reacts quickly enough.
• Manufacturing robots—modern robots are much safer and easier to train
because they have AI embedded in them. Baxter, a robot designed by
Rodney Brooks’s ReThink Robotics, can work on a production line without
a cage because it has the ability to immediately stop if it is going to hit someone or something. It can be trained by simply moving the arms and body in
the required series of actions, rather than having to program each one.
• Care-bots—the use of robots to supplement or replace the human care of
sick or vulnerable people is a controversial one, but they can have some positive benefits. There are robots that help care for the elderly, either by providing ‘company’ (through Speech Recognition and NLU) or support through,
for example, helping them remember things (Optimisation). Other
AI-powered robots that are used in the medical field include the use of telepresence robots—these are mobile pods that can travel around hospitals but
are connected via video and sound to human doctors at a remote location.
• Service robots—Some retailers are starting to use mobile robots to greet
and serve customers. As with the hospital telepresence robots described
above, these are mobile units with computer vision, speech recognition,
NLU and optimisation capabilities built in. As well as being able to serve
customers in a shop, different versions of the robots can take the role of
waiters in restaurants or concierges in hotels. There are also interesting
examples where robots are learning skills through reading material on the

AI and the IoT

65

internet—there is one case where a robot has learnt to make pancakes by
reading articles on WikiHow.
• Swarm robots—these are a specific field of robotics where many small
machines work collaboratively together. They rely heavily on the AI
Optimisation capability, constantly evaluating the next best move for each
of the swarm robots so that they can achieve the shared goal. Generally
ground-based, but can be aerial or water-based, they are usually used in
environments that are difficult for humans to work in, such as disaster rescue missions or, more controversially, in warfare (think autonomous
armies). Autonomous vehicles will also be able to exploit swarm
intelligence.
AI is also being used in human-like ways to help robots learn. Researchers
have developed an approach where a humanoid robot learns to stand itself up
by ‘imagining’ what it is like to stand up. Essentially it will run a series of
simulations using a DNN, and then use a second system to analyse the feedback from the various sensors as it actually tries to stand up.
Cognitive Robotics, therefore, can be considered the physical embodiment
of AI. Using input data from many different types of sensors the robot uses a
combination of the Speech Recognition, Image Recognition, NLU and
Optimisation capabilities to determine the most appropriate response or
action. And, because it is AI, the system can self-learn, becoming more effective the more it is used.
Of course, physical cognitive robots also stir in us the fear that they will eventually be able to ‘take over’. Software-based AI is not going to be a risk to the
human race if it is stuck on a server; we can always pull the plug out. But physical
robots could potentially have the ability to overpower us, especially if they are
able to build better versions of themselves. There are already examples of cognitive robots being able to do this, but I refer you back to my earlier comments in
the section on AI Understanding about the distinction between Artificial Narrow
Intelligence and Artificial General Intelligence. There is, of course, an existential
risk there; it’s just that it won’t need to concern us for a very long time.

AI and the IoT
The Internet of Things, or IoT, refers to simple physical devices that are connected to the internet. IoT devices include webcams, smart lights, smart thermostats, wearables and environmental sensors. Many people believe that IoT
is, along with AI, one of the major technology trends of the decade.

66

4

Associated Technologies

There are billions of IoT devices in the world today, each one generating
data or reacting to data, effectively creating and consuming big data on a
grand scale, which is why AI has such a symbiotic relationship with the IoT.
IoT devices are being used in business today to
• manage preventative maintenance programs by analysing data from sensors
embedded into the assets (e.g. escalators, elevators and lighting)
• manage supply chains by monitoring movement of products
• conserve energy and water usage of machines by analysing and predicting
demand (smart meters, which monitor energy usage every 15 minutes, are
already common in many homes)
• improve customer experience by providing personalised content based on
IoT data
• increase crop yields by analysing field sensors to provide precise feeding
and watering programs
• alleviate parking problems by matching empty spaces with cars and their
drivers
• keep us fitter by monitoring and analysing our steps and exercise patterns
through wearable devices
One of the headline successes of ‘IoT plus AI’ was when Google announced
that they had been able to reduce the energy used for cooling in one of their
data centres by 40%. By using DeepMind’s AI technology, they were able to
analyse data from many different types of sensors (such as thermometers,
server fan speeds, and even whether windows were open or not), precisely
predicting demand so that they could instruct the various machines to work
at optimum levels.
Another area where IoT and AI are key enablers is in the development of
Smart Cities. IoT devices are used to track and extract data for transport,
waste management, law enforcement and energy usage. This data is analysed
and turned into useful information by AI Clustering, Optimisation and
Prediction systems, and then made available to the relevant authorities, the
city’s citizens and other machines. For example, sensors in street lights and
urban furniture are able to measure footfall, noise levels and air pollution—
this data is then used to prioritise the delivery of other services.
The biggest challenge to the proliferation of the IoT is the poor security
that has been associated with it. Many devices do not have the ability to
change passwords, and have only basic authentication. There have been recent
cases of baby monitors allowing strangers to monitor their camera feed,
internet-­connected cars’ entertainment systems and central locking systems

AI and Crowd Sourcing

67

being taken over by hackers, and, probably most worrying of all, medical
devices being compromised so that a hacker could send fatal doses of medicines to drug infusion pumps. The good news is that these high-profile cases
has meant that IoT security is very high on many technology company’s agendas and there is now plenty of activity to fix all the issues. But if you are thinking of implementing an IoT strategy, do keep security at the top of your
considerations.
As more and more IoT devices are used (it is estimated that there will be 50
billion devices by 2020), there will be more need for AI to make sense of all
the data that is generated by them. The analysis that can be carried out can
then be actioned by other IoT devices such as actuators and lights. As in the
Smart Cities example, the real value will be the sharing and collaboration of
this data across organisations, people and other machines.

AI and Crowd Sourcing
Sometimes AI simply isn’t up to the job. Sometimes you will need to pull
humans into the loop to help complete the process.
A classic example of when this sort of situation can arise is when the source
documents for an automated process are handwritten. We already know that AI
is very capable of extracting structured data out of unstructured documents but
this really applies only when the unstructured documents are in an electronic
format in the first place. Optical Character Recognition is able to take typed
documents such as PDFs and convert them into an electronic format, but if the
source is handwritten the challenge becomes a whole lot more difficult.
One solution to this is to use a Crowd Sourcing service. Crowd sourcing is
where large numbers of people are engaged to carry out small parts of a process (‘micro-tasks’). The engagement model is usually via the Internet, with
each person paid a specific rate for every time they complete one of the
micro-tasks.
In the example above, one document can be split into many different tasks
so that one person is sent the First Name to read, another is sent the Last
Name and a third the Social Security Number. Each person will look at the
image of the handwriting and respond with the text that it represents. Because
each person only sees a small part of the information, confidentiality is maintained. To increase accuracy, a single image can be sent to a number of people
so that only the most common answer is selected. Specialised software is used
to manage the interface between the customer and the crowd, including the
splitting of the documents into smaller parts.

68

4

Associated Technologies

A second role for crowd sourcing in the AI space is to provide additional
capability when the AI is not confident enough of its answers. This might be
because the problem is too complex, or it hasn’t seen that particular problem
before (Fig. 4.1).
In this case, the AI will send the problem to a person to answer. For example, some AI systems are used to moderate offensive images on social media
sites. If the AI is unsure whether a particular image is offensive or not, it will
ask a human for their opinion. This approach is usually called ‘Human In The
Loop’, or HITL.
An enhanced version of HITL is where the input from the human is used
to actively train the AI system. So, in the offensive image example, the human’s
choice (offensive or inoffensive) would be sent back to the AI so that it can
improve its learning and perform better in the future (Fig. 4.2).
A third use of crowd sourcing with AI is in the development of the training
data. As I discussed in the Technology Overview section of Chap. 1, supervised learning approaches require data sets that are appropriately tagged (dog
pictures tagged as dogs, cat pictures tagged as cats, etc.). Because of the large
data sets that are required, the job of tagging all these data points is very laborious. Crowd sourcing can be used to satisfy that requirement. Google itself
pays for tens of millions of man-hours collecting and labelling data that they
feed into their AI algorithms (Fig. 4.3).
The most popular general crowd sourcing site is Mechanical Turk (owned
by Amazon) but others, such as Crowd Flower, that focus on supporting the
AI community, are also available. Some also have an Impact Sourcing angle as

Artificial
Intelligence

Confident?

No

Human
Intervention

Fig. 4.1 Human in the loop

Yes

Output

AI and Crowd Sourcing

Artificial
Intelligence

Confident?

Yes

69

Output

No

Active Learning

Human
Intervention

Fig. 4.2 Training through a human in the loop

Tagging By
Humans

Training

Artificial
Intelligence

Fig. 4.3 Crowd-sourced data training

well, where the people doing the work are disadvantaged in some way and this
work gives them new opportunities.
So, despite the amazing things that AI can do, at some point it will probably still rely on people to do part of the job, whether that is training it or
helping it make decisions.

70

4

Associated Technologies

The Established AI Vendor’s View
This is an extract from an interview with Andrew Anderson, CEO of Celaton,
a leading AI software vendor.
AB What got you into the world of AI in the first place?
AA	The story began in 2002. I sold my previous software company to a much
larger company that pioneered the delivery of software as a service, at that
time they were described as an ‘application service provider’. As the VP of
product development, we went on to acquire, license and develop lots of
software applications that solved many different challenges faced by organisations, all delivered as a service. However, despite access to all this technology we still couldn’t solve the challenge of dealing with unstructured,
varied (human created) data that flowed into organisations every day. It
remained a manual, labour-­intensive task.
I saw an opportunity to create a solution to the challenge and so in
2004 I bought my company back, along with my development team and
set about trying to build an ‘as a service’ platform that could automate the
processing of unstructured, varied data. My optimistic view then hoped
that we would have a solution built within a year. The reality was that it
took over six years to create a platform that had the ability to learn, and
therefore enable us to describe what we had created as ‘artificially intelligent’. In fact, it wasn’t until we demonstrated it to analysts that they told
us what we had created.
We spent a few years (2011–2013) trying different approaches until we
realised that the greatest value we deliver was to large, ambitious organisations who deal with demanding consumers.
AB Why do you think AI is being so talked about right now?
AA	AI is being talked about more than ever because it promises so much.
Whether those promises are realised is yet to be seen. It’s like a wonder
drug that has huge potential but the enthusiasm needs to be tempered
with a little reality.
What is emerging now are specialist AIs that are good at specific tasks
rather than general-purpose AIs that are good at everything. Whilst there
are many case studies where specific or narrow AI is delivering real benefits this is being mixed up with noise about general AI.
There’s sanity in numbers and it takes time and effort to educate and
convince enough organisations that they should try something but each
new client means a new case study and each case study leads to new customers. These are the realities behind the hype, but the media often prefer
to hear the more exciting stories of potential rather than reality.

The Established AI Vendor’s View

Automation has been around since the wheel was invented and the
world has continued to invent and innovate and that’s why we see this
continuous improvement in technology. The increased power of technology enables it to achieve outcomes that were previously out of reach.
Automation of any kind is also accelerating and with the emerging of
AI it must seem that everything can be automated. What was previously
considered the domain of human effort is now being eroded and this creates negative media.
In summary, I think it’s being talked about more for its potential than
its reality, although the reality (i.e. the case studies) are helping to fuel the
future potential because people can actually see the future emerging.
AB What value do your customers get out of your software?
AA	There’s a different answer depending on the customer challenge. By
understanding what customers are saying, then reacting and responding
to them enables our customers to deliver better service faster, with fewer
people. In summary, it enables them to achieve competitive advantage,
compliance and improve their financial performance.
AB	What do customers or potential customers need to focus on if they are
going to get the maximum value from AI?
AA	I think the key is not to focus on the technology but to understand the
problem you want to solve. Many people (within organisations) have
trouble understanding their problem and therefore they struggle to identify the technology that might help to solve it.
This is where experienced consultancy is important. Understand the
problem, select and apply the most appropriate technology and then share
the story.
AI is often considered to be the solution to all problems like a magic
drug. The reality is there is no magic drug but there are lots of drugs that
solve different problems. The key is to talk with someone who can help to
understand the problem and prescribe the right ‘drug’.
AB Do you think the current hype is sustainable?
AA	I don’t think that we are witnessing anything different with this hype. The
industry does tend to get ahead of itself because hype serves it well, creating awareness that helps to capture the interest of customers. There’s sanity in numbers.
The difference this time is that technology innovation is moving
quicker and it’s not so long after the thought has been considered that the
invention can be demonstrated.

71

72

4 Associated Technologies

Regardless of how fast things move, it’s the human that is the limiting
factor. The bigger the problem that it solves, the more likely it is to be
adopted by humans and therefore be successful.
AB How do you see the market developing over the next 5 years or so?
AA	Hype will lead to reality, but by solving problems that really exist. There
are a couple of areas that are particularly relevant:
• 	Natural interfaces. We’re seeing this already with the likes of Amazon
Echo and IoT appliances. There is some way to go but they will become
more natural and to such an extent that humans will have to think less
about how they communicate with technology.
•	Broader AI. We’re currently seeing specialist AIs that are good at doing
specific things – these technologies will consolidate into broader solutions that are able to do lots of things well because they’re made up of
lots of vertical AIs.
And I particularly see some of the greatest and most profound use of
AI in medicine, energy and, subject to world events, in the military too.

5
AI in Action

Introduction
Up to this point in the book I have been pretty theoretical, explaining the capabilities of AI and the types of technologies it requires to work. In this chapter I
change focus to look at how AI is actually being used in businesses—real use
cases of AI adding value and changing the way companies do business.
I have split the chapter into different themes which look at specific aspects
of business and the ways that AI adds value: enhancing customer service,
optimising processes and generating insights. There are overlaps between each
of these areas (there are, for example, insights you can generate about your
customers) but it gives a general, useful framework to describe how AI is
being used in business.
In the chapter following this I describe the different types of benefits that
should be considered in an AI Business Case. These can be broadly mapped
against the themes:
• Enhancing customer service leads to revenue generation and customer
satisfaction
• Optimising processes leads to cost reduction, cost avoidance and
compliance
• Generating insights leads to risk mitigation, loss mitigation and revenue
leakage mitigation
Each of these themes will also draw upon different capabilities from the AI
Framework. All will use at least one of these capabilities, whilst others will
exploit a few.
© The Author(s) 2018
A. Burgess, The Executive Guide to Artificial Intelligence,
https://doi.org/10.1007/978-3-319-63820-1_5

73

74

5 AI in Action

I have tried to take examples from a wide range of industries, and have
deliberately not grouped the case studies into specific sectors. This is so you
are not tempted to skip over sectors that are not immediately relevant to
you—I am a firm believer that very different industry sectors can still learn
from each other. Even if you work in the retail sector, for example, you may
still get a spark of an idea from something in the utilities sector.

How AI Is Enhancing Customer Service
The ‘front office’ is one of the more active areas when it comes to implementing AI. This is because there is generally plenty of customer data available to
work off, but it is also down to the proliferation of chatbots.
Chatbots come in all shapes and sizes, which is a rather polite way of saying that there are really good chatbots but also very bad ones. Chatbots, which
aim to have a natural conversation with a customer through a typed interface,
use NLU as their key AI capability. This means that, in theory, a customer
who wants to change their address on the system (this is their ‘intent’ in
AI-speak) can ask the chatbot in any way that they want (‘I’m moving to a
new house’, ‘I’ve got a new address’, ‘I’m not living at the same place any
more’, ‘My postcode has changed’, etc.) and it will still understand the underlying intent.
In reality, this is a big challenge for NLU, and most chatbots use a ‘dictionary’ of equivalent phrases to refer to (e.g. if the input is ‘I’ve got a new address’
or ‘I’m moving to a new house’, then the intent is ‘Customer wants to change
their address’). This, of course adds to the complexity of designing the chatbot, as every alternative phrase has to be identified and entered.
Some chatbots make extensive use of multiple-choice questions. So, instead
of relying on understanding what the customer has typed, the chatbot will ask
a question with a limited number of answers (e.g. Yes/No, Ask for a Balance/
Make a Payment/Change Your Address, etc.). This can be more efficient and
accurate but doesn’t make for a natural conversation.
Another challenge for chatbots is how they move through the conversation.
The simplest and most common approach is a decision tree, where each question will branch off to a new question depending on the answer. If the process
being replicated is complex, this can lead to an over-complicated and burdensome decision tree. A better approach here is to have a cognitive reasoning
engine do all of the ‘thinking’ whilst the chatbot gets on with the conversation. This provides much more flexibility in how the chatbot deals with the
flow of the conversation. Cognitive reasoning systems were described in the
Optimisation section of Chap. 3.

How AI Is Enhancing Customer Service

75

The purest AI approach to chatbots is to train them on thousands of
human-to-human chat conversations, where each interaction has been tagged
with the intent and whether it was productive or not. These chatbot systems
effectively learn the knowledge map (a.k.a. ontology) through those historical
interactions. The challenge with these systems is that there needs to be plenty
of training data available, and they tend to very expensive to implement.
With chatbots, it is very much ‘horses for courses’. A simple, ‘free’ chatbot
system will be fine for very simple and non-critical interactions with customers, but because of the reputational risk involved, implementing chatbots is
worth doing only if you do it right. The better the chatbot system, the better
it will cope with the challenge.
In 2017 Royal Bank of Scotland introduced a chatbot to help answer a
limited number of questions from customers. The chatbot, called ‘Luvo’, uses
IBM Watson’s Conversation capability to interact with customers who are
using the bank’s website or app. It was trialled for nearly a year with internal
staff who manage relationships with SMEs before being released to a small
number of external customers. At the time of writing, it is able to answer just
ten defined questions, such as ‘I’ve lost my bank card’, ‘I’ve locked my PIN’
and ‘I’d like to order a card-reader’.
Luvo demonstrates the cautious approach that enterprises are taking to
implementing chatbots. Over time though, as the system learns from each
interaction it has, the bank will allow it to handle more complex questions,
build in greater personalisation and use predictive analytics to identify issues
for customers and then to recommend the most appropriate action. The key
objective is to free up more time for the human customer service agents to
handle the trickier questions that customers may have.
In a similar two-step implementation, SEB, one of Sweden’s largest banks,
deployed IPsoft’s Amelia software first in their internal IT Service Desk, and
then to their one million customers. During the first three weeks of the trial,
the chatbot-cum-avatar helped to answer the staff’s IT issues, resolving around
50% of the 4000 queries without human interaction. During 2017, it was
rolled out to help manage interactions with the bank’s customers (and named
‘Aida’). Three processes were initially chosen as candidates: providing information on how to become a customer, ordering an electronic ID and explaining how to do cross-border payments.
Although Amelia learns from historical interactions and uses sentiment analysis, it is also able to follow defined workflow paths to ensure compliance with
banking regulations. Whereas most chatbots can be described as probabilistic,
providing likely results, Amelia can be considered a deterministic system because,
once it has worked out the intent, it can, where possible, carry out the required

76

5 AI in Action

actions on the enterprise systems (in a similar way that RPA tools would). IPsoft
is a US-based company, and this was their first non-­English deployment of their
solution.
From an infrastructure point of view, SEB decided to install the Amelia technology on their own servers rather than as a cloud solution. This was due to
concerns over the compliance and legal concerns that a cloud deployment raises.
For IPsoft’s financial services customers, this is the most common approach.
The deployment of Amelia is very much aligned with the bank’s overall
strategy, which includes the line “SEB will focus on providing a leading customer experience, investing in digital interfaces and automated processes”.
The US flower delivery service 1-800-Flowers deployed a simple chatbot
that allowed customers to place orders through Facebook Messenger. It is a
linear, decision tree system with an NLU chat interface that took around three
months to develop and test. This makes it rather constrained in what it can
do, but additional functionality and complexity is being rolled out. Following
the initial two months of operation, 70% of the orders being placed through
the chatbot were from new customers, and were dominated by ‘millennials’
who tend to already be heavy users of Facebook Messenger.
As well as taking orders, the chatbot can direct a customer to a human
agent, of which there may be up to 3,500 at any one time. They have also
implemented an integration with Amazon Alexa, and a concierge service that
is powered by IBM Watson. Together, these digital initiatives have attracted
‘tens of thousands’ of new customers to the brand. They also provide up-to-­
the-minute behaviour data for the company which can then influence real-­
time marketing activity such as promotions.
Another company that has used IBM Watson to enhance customer engagement is Staples, the stationery retailer. They have implemented a range of
different ways for people to buy their products as easily as possible, including
email, Slack, mobile app and (honestly) a big red button. The button is similar to Amazon’s Alexa in that it can understand voice commands (although it
does need a physical press to activate and deactivate it). The mobile app can
also understand voice commands as well as being able to identify products
from photographs taken. All these channels make the buying process as
­frictionless as possible for the customer, and therefore has a direct and positive
impact on revenue for the retailer.
As well as chatbots, recommendation engines are another common AI technology that is used to enhance customer service (and, of course, drive revenue).
Amazon and Netflix have the most well-known recommendation engines, and
these are deeply embedded in the normal workflow of how the customer engages
with the companies. All the required data—the individual customer’s buying

How AI Is Enhancing Customer Service

77

and browsing behaviours and the historical data of all the other customers’
behaviours—are available through those normal interactions; that is, there is no
additional work that the customer has to do to enrich the data set.
In some cases, the recommendation engines will require additional information from the customer and/or the business in order to work effectively. North
Face, a clothing retailer, has implemented a recommendation engine for their
customers who want to buy jackets. Based on IBM Watson this solution, called
XPS, uses a chatbot interface to ask a series of refining questions so that it can
match the customers’ requirements with the product line. According to North
Face, 60% of the users clicked through to the recommended product.
Another clothing retailer, Stitch Fix, uses a slightly different approach by
deliberately including humans in the loop. Its business model involves recommending new clothes to their customers based on a selection of information
and data that the customer provides (measurements, style survey results,
Pinterest boards, etc.). All this structured and unstructured data is digested,
interpreted and collated by the AI solution, which sends the summary, plus
anything that is more nuanced (such as free-form notes written by the customer) to one of the company’s 2,800 work-from-home specialist human
agents, who then select five pieces of clothing for the customer to try.
This is a good example of where the AI is augmenting the skills and experience of the human staff, making them better at their jobs as well as being
more efficient. Having humans in the loop (HITL is the acronym for this)
also makes experimenting easier, as any errors can quickly be corrected by the
staff. To test for bias, the system varies the amount and type of data that it
shows to a stylist—it can then determine how much influence a particular
feature, say a picture of the customer or their address, can make on the stylist’s
decisions. On top of all this, the data that they gather about all of their customers can also be used to predict (and influence?) general fashion trends.
Other customer service AI solutions don’t rely on chatbots or recommendation engines to provide benefits. Clydesdale and Yorkshire Banking
Group (CYBG) is a medium-size bank in the United Kingdom, having to
compete with the ‘Big Four’ of Barclays, HSBC, Lloyds and RBS. Their digital strategy includes a new current account, savings account and app package,
called ‘B’. It uses AI to help manage the customer’s money: it allows you to
open account, and, once opened, will learn the patterns of usage so that it can
predict if you might run out of funds in your accounts, and suggest ways to
avoid unnecessary bank charges. The bank claims that an account can be
opened in 11 minutes. Clearly, customers avoiding bank charges will result in
lower revenue, but the hope is that it will attract enough new customers that
this revenue loss will easily be outweighed by the benefits of having the additional funds.

78

5 AI in Action

Virgin Trains ‘delay/repay’ process has been automated through the application of AI. After implementing Celaton’s inSTREAM AI software to categorise inbound emails (see next section) the train operating company used the
same software to provide a human-free interface for customers to automatically claim refunds for delayed trains.
The Asian life insurance provider AIA has implemented a wide range of AI
initiatives that all impact how they engage with their customers, including
insight generation from prospective customers, enhanced assessment of customer needs, 24×7 chatbots for online enquiries, inbound call handling by
NLU-based systems, enhanced compliance of sales, personalised pricing,
dynamic underwriting and an augmented advice and recommendation engine.
Some companies are building AI into the core of their customer applications. Under Armour, a sports clothing company that has a portfolio of fitness apps, uses AI in one of those apps to provide training programs and
recommendations to the users. The AI takes data from a variety of sources,
including the users’ other apps, nutritional databases, physiological data,
behavioural data and results from other users with similar profiles and objectives. It then provides personalised nutrition and training advice that also
takes into account the time of day and the weather.
Other examples of where AI can enhance customer service are: optimising
the pricing of time-critical products and services such as event ticketing; optimising the scheduling of real-time services such as delivery and departure
times; creating personalised loyalty programs, promotional offers and financial products; identifying candidate patients for drug trials and predicting
health prognoses and recommending treatments for patients.
There is clearly a balance between using AI to serve customers better, and using
AI to gather information about customers. In some of the cases I have described,
the same AI capability will do both. As I have reiterated a number of times, for
anyone looking to implement a ‘front office’ AI solution, it is important to first
understand your objectives. The danger of trying to ask your AI to do too much
is that you compromise on one objective or the other. I discuss approaches and
examples of how to generate customer insights later in this chapter.

How AI Is Optimising Processes
If the previous section was all about enhancing the front office, then this section focuses on what happens behind the scenes in the back office operations.
The benefits here are centred around, although not exclusively, reducing cost
(directly or through avoidance) and improving compliance.

How AI Is Optimising Processes

79

One of the key capabilities exploited in the back office is being able to
transform unstructured data into structured data, that is, from the Image
Recognition, Voice Recognition and Search capabilities.
Tesco have implemented a number of AI-powered solutions to improve the
productivity of their stores. They are using image recognition systems to identify empty shelves on stores, called ‘gap scanning’. They are experimenting
with physical robots that travel down the aisles during quiet periods to film
the shelves so that they can measure stock availability and inform staff to
replenish them where necessary. Not only does this save the time of the staff
(who would normally have to use a diagram on a piece of card as the template)
but it also reduces revenue loss due to stock unavailability.
For their home delivery service Tesco have implemented optimisation systems that minimise the distance the picker walks around the store collecting
the goods, and also a system that maximises the productivity of the delivery
vans through effective route planning and scheduling. Interestingly, in a meta-­
example, they have used AI to help place the many products into the necessary taxonomy that AI systems use for classification.
AIA, the Asian insurance provider I mentioned in the previous section,
have also implemented initiatives in their servicing and claim functions,
including service augmentation, investment management, medical second
opinions, risk management, fraud management, claims adjudication and
health and wellness advice to their customers. AIA have been very clear that
transforming their business requires end-to-end transformation with AI at
the core.
Insurance claims processing is a prime candidate for enhancement through
AI and RPA. Davies Group, a UK-based insurance firm that provides a claims
processing service to their own insurance customers, has automated the input
of unstructured and semi-structured data (incoming claims, correspondence,
complaints, underwriters’ reports, cheques and all other documents relating
to insurance claims) so that it is directed into the right systems and queues.
Using Celaton’s inSTREAM solution, a team of four people process around
3,000 claims documents per day, 25% of which are paper. The tool processes
the scanned and electronic documents automatically, identifies claim information and other meta-data, and pastes the output into databases and document stores ready for processing by the claims handlers and systems (which
could, of course, be either humans or software robots). It also adds service
meta-data so the performance of the process can be measured end to end.
Some documents can be processed without any human intervention, and others need a glance from the human team to validate the AI’s decisions or fill in
missing details.

80

5

AI in Action

AI can also be used in the claims process to identify fraud (by looking for
anomalous behaviours), to help agents decide whether to accept a claim or
not (using cognitive reasoning engines) and by using image recognition to
assess damage. Ageas, a Belgian non-life insurance firm, has deployed
Tractable, an AI image recognition software, to help it assess motor claims. By
scanning images of damage to cars, the software is able to assess the level of
damage, from repairable to being a write-off. Through the process, it also
helps identify fraudulent claims.
Image recognition has also been used in other industry sectors. Axon,
which used to be known as Taser, the manufacturer of police equipment, are
now using AI to tag the thousands of hours of video that is recorded from the
body cameras that they manufacture. The video recordings were previously
available to police forces on subscription but, due to the sheer volume of
recordings, they were rarely used as evidence. Now that the film is being
tagged, the crime agencies will be able to automatically redact faces to protect
privacy, extract relevant information, identify objects and detect emotions on
people’s faces. The time saved, from having to manually write reports and tag
videos, will be huge, but the ability to make better use of the available evidence will likely have a bigger impact. For Axon, it will encourage more police
forces to buy their body cams.
Data from video feeds has been used as the core of the business model of
Nexar, an Israeli-based technology company. They are giving away for free
their dash cam app which helps drivers anticipate accidents and problems in
the road ahead. By using the data provided by the apps from their user base
(which is mandatory if you want to use the app) the system predicts the best
course of action, for example to brake hard if there is an accident up ahead.
This model relies heavily on having large amounts of data available, hence the
up-front giveaway. Rather controversially, the company can use the data for
‘any purpose’, including selling it to insurance companies and giving it to the
government. Whether this is a sustainable business model remains to be seen,
but it does demonstrate how business models are transforming to focus more
on the data than the service being provided to the customers.
A particularly worthwhile use of image recognition is in the medical
field. Probably the most widely reported has been the use of IBM’s Watson
platform to analyse medical images, such as X-rays and MRI scans, so that
it can help identify and diagnose cancer cells. The real benefit of this system is that it can cross-reference what it sees with the patient’s (unstructured) medical records and thousands of other images that it has learnt
from. This approach is already spotting characteristics of tumours that
medical staff hadn’t seen—they are finding that peripheral aspects of the

How AI Is Optimising Processes

81

tumour (e.g. those distributed around, rather than in, the tumour) have a
much greater influence on predicting malignancy than at first thought.
These systems are currently working in two hospitals: UC San Diego
Health and Baptist Health South Florida. (See the section on Generating
Insights for more examples of how AI is helping fight cancer.)
Deutsche Bank are using the speech recognition technology to listen to
recordings of their staff dealing with clients, in order to improve efficiency,
but, more importantly, ensure compliance to regulations. These conversations
are transcribed by the AI and can be searched for specific contextual terms.
This used to be the job of the bank’s auditors, and would require listening to
hours of taped recordings, but the AI system can do the same thing in minutes. The bank is also using other AI capabilities to help it identify potential
customers based on the large amount of public information that is available
on people.
The legal sector is currently going through a big transition period as it starts
to adopt new technologies, including AI. Inherently risk averse, and never an
industry to change quickly, the legal sector is starting to feel the benefit of AI
particularly through implementing Search and NLU technologies to help
make sense of contracts.
Legal contracts can be described as semi-structured documents—although
they follow some general rules (they usually include parties to the contract,
start and end dates, a value, termination clauses, etc.), they are verbose and
variable. There are now a number of software solutions on the market which,
together, work across the full contract lifecycle—some are able to search document repositories to identify where contracts are; others can ‘read through’ the
contracts, identify specific clauses and extract the relevant meta-­data (such as
the termination date and the limits of liability). Some have the capability to
compare the real contracts with a model precedent and show all instances
where there is a material difference between the two. All of this allows large
enterprises and law firms to manage their contracts much more effectively,
especially from a risk point of view.
McKesson, a $179 billion healthcare services and information technology
organisation, uses Seal’s Discovery and Analytics platform to identify all of
their procurement and sourcing contracts across the business (they employ
70,000 people) and store the documents in a specific contract repository.
When in the repository, the contracts could be searched easily and quickly,
saving some staff hours per day. The meta-data from the contract analysis
allowed them to identify potential obligation risks as well as revenue opportunities and savings from otherwise hidden unfavourable payment terms and
renewal terms.

82

5 AI in Action

A Magic Circle law firm, Slaughter & May, adopted Luminance AI software to help manage the hundreds of Merger and Acquisition (M&A) transactions they do a year. This area is a particular challenge because of the complexity
involved (thousands of documents, many jurisdictions) and the intensity of
the work—the law firm was concerned that some of the junior lawyers who
were tasked with managing the M&A ‘data room’ (the contract repository created specifically for the deal) would eventually burn out. Luminance is used to
cluster, sort and rank all of the documents in the data room, with each document assigned an anomaly score to show how it differs from the ideal model.
Whereas in normal due diligence exercises only around 10% of the documents
are analysed, due to the sheer volume, Luminance is able to analyse everything.
It takes about an hour for it to process 34,000 documents. Overall, it halves the
time taken for the whole document review process.
One of the benefits of having the AI systems carry out ‘knowledge based’
works such as this is that it allows lawyers to focus on complex, higher-value
work. Pinsent Masons, a London-based law firm, designed and built in-­
house a system, called TermFrame, which emulates the legal decision-making
workflow. The system provides guidance to lawyers as they work through different types of matters, and connects them to relevant templates, documents
and precedents at the right time. Taking away much of the lower-level thinking that this requires means that the lawyers can spend more time on the
value-add tasks.
As well as calling on Search and NLU capabilities, companies are also using
NLG solutions to optimise some of their processes. Using Arria’s NLG solution, the UK Meteorological Office can provide narratives that explain a
weather system and its development, with each one configured with different
audiences in mind. The Associated Press (AP), a news organisation, uses
Wordsmith to create publishable stories, written in the AP in-house style,
from companies’ earnings data. AP can produce 3,700 quarterly earnings stories which represents a 12-fold increase over its manual efforts.
Artificial intelligence can also be used to enhance the IT department as well
(‘Physician, heal thyself ’). There are a number of AI capabilities that can be
deployed, both in the support function and for infrastructure management.
Generally, AI brings a number of important aspects to the IT function: the AI
systems can be trained from run-books and other sources; they will continue
to learn from ‘watching’ human engineers; they can proactively monitor the
environment and the current state; they can identify trends in system failures
and they are able to cope with the inherent variability of IT issues.
We have already seen how chatbots can support customer engagement, but
they can also be used with staff as well, particularly where there are a large

How AI Is Optimising Processes

83

number of employees, or a large number of queries. Although chatbots can be
deployed on any type of service desk, including HR, they are most commonly
found in IT.
As in the SEB example discussed in the previous section, AI Service Desk
agents are able to receive customer enquiries and provide answers by understanding what a customer is looking for and asking the necessary questions to
clarify the issue. The more advanced systems will, if they cannot help the
customer themselves, raise the issue to a human agent, learning how to solve
the problem itself for future situations.
AI can also be used to manage a company’s IT infrastructure environment.
The AI systems tend to work as meta-managers, sitting above and connecting
all the various monitoring systems for the networks, servers, switches and so
on. They include both proactive and predictive monitoring (using AI) and
execution of the necessary fixes (using software robots). Key performance
indicators such as downtime can be significantly improved using these systems. And, by tackling many of the mundane tasks, the systems free up the IT
engineers to focus on higher-value, innovative work.
TeliaSonera, a European telecoms operator, implemented IPCentre to
help manage its infrastructure of 20 million managed objects, including
12,000 servers, and have subsequently seen cost savings of 30%. A New York-­
based investment firm used the same solution to help it fix failed fixed-­
income securities trades due to system issues. Eighty per cent of the failed
trades are now fixed without human intervention, and the average resolution
and fix time has been reduced by 93% (from 47 minutes to 4 minutes). This
has resulted in a staff reduction of 35%.
Google, a company with many very large data centres, turned to its AI
subsidiary, DeepMind, to try and reduce the cost of powering these facilities.
The AI worked out the most efficient methods of cooling by analysing data
from sensors amongst the server racks, including information on aspects such
as temperatures and pump speeds. From the application of the AI algorithms,
Google was able to reduce cooling energy demand by 40%, and to reduce the
overall energy cost of the initial data centre by 15%. The system now controls
around 120 variables in the data centres including fans, cooling systems as
well as windows.
Another example of where AI can optimise processes includes the ability to
model real-world scenarios on computers, with the most obvious example being
the prediction of long-term weather conditions. But this modelling approach
has been used to test physical robotic systems so that the designers can make
adjustments to their robots virtually without the fear of them falling over and
breaking. Facebook have created a specific version of Minecraft, a virtual-world

84

5 AI in Action

game, for this sort of purpose. AI designers can also use the environment to help
teach their AI algorithms how to navigate and interact with other agents.
In this section, we have seen examples of how AI can optimise the back office
processes of retailers, banks, insurance firms, law firms and telecoms companies,
across a wide range of functions, including claims management, compliance and
IT. Other use cases for AI in optimising processes include: optimising logistics
and distribution of inventory in warehouses and stores; real-time allocation of
resources to manufacturing processes; real-time routing of aircraft, trucks and so
on; optimising the blend and timing of raw materials in refining and optimising
labour staffing and resource allocation in order to reduce bottlenecks.

How AI Is Generating Insights
In the previous two sections I looked at how AI can enhance customer service
and optimise processes. But, in my mind, the biggest benefits from AI are
when it can deliver insight. This is where new sources of value are created from
the data that already exists, enabling better, more consistent and faster decisions to be made. AI can therefore enable a company to mitigate risks, reduce
unnecessary losses and minimise leakage of revenue.
One of the most effective uses of AI to date has been in the identification
of fraudulent activity in financial services. The benefit here is that there is
plenty of data to work from, especially in retail banking. PayPal process $235
billion in payments a year from four billion transactions by its more than 170
million customers. They monitor the customer transactions in real time (i.e.
less than a second) to identify any potentially fraudulent patterns—these ‘features’ (in AI-speak) have been identified from known patterns of previous
fraudulent transactions. According to PayPal, their fraud rate is just 0.32%,
compared to an industry average of 1.32%.
The American insurance provider USAA is similarly using AI to spot identity theft from its customers. It is able to identify patterns in behaviour that
don’t match the norm, even if it is happening for the first time. Axa, another
insurance firm, has used Google’s TensorFlow to predict which of its customers are likely to cause a ‘large-loss’ car accident (one that requires a pay-out
greater than $1,000). Normally up to 10% of Axa’s customers cause a car
accident every year, but about 1% are ‘large-loss’. Knowing which customers
are more likely to be in that 1% means that the company can optimise the
pricing of those policies.
The Axa R&D team initially tried a Decision Tree approach (called Random
Forest) to model the data but were able to achieve only 40% prediction
accuracy. Applying deep learning methods on 70 different input variables,

How AI Is Generating Insights

85

they were able to increase this to 78%. At the time of writing, this project is
still in its proof-of-concept (PoC) phase but Axa are hoping to expand the
scope to include real-time pricing at point of sale and to make the system
more transparent to scrutiny.
Another company that is using AI to optimise pricing in real time is Rue
La La, an online fashion sample sales company who offer extremely limited-­
time discounts on designer clothes and accessories. They have used machine
learning to model historical lost sales in order to determine pricing and predict demand for products that it has never sold before. Interestingly, from the
analysis they discovered that sales do not decrease when prices are increased
for medium and high price point products. They estimated that there was a
revenue increase of the test group of almost 10%.
Otto, a German online retailer, is using AI to minimise the number of
returns they get, which can cost the firm millions of euros a year. Their particular challenge was the fact that they knew that if they got the orders to their
customers within two days then they would be less likely to return them
(because it would be less probable for them to see the same product in another
store at a lower price). But, they also knew that their customers liked to get all
their purchases in one shipment, and, because Otto sources the clothes from
other brands this was not always easy to achieve.
Using a deep learning algorithm, Otto analysed around three billion historical transactions with 20 variables so that it could predict what customers
would likely buy at least a week in advance. Otto claims that it can predict
what will be sold within 30 days to an accuracy of 90%. This means it can
automate much of the purchasing by empowering the system to auto-order
around 200,000 items a month. Their surplus stock has reduced by a fifth and
returns have gone down by more than two million items a year.
Goldman Sachs, the global investment bank, has implemented a range of
automation technologies to both improve decision-making processes and also
to reduce headcount. They started by automating some of the simpler trades
that were done at the bank—in the year 2000 there were 600 equity traders
working in the New York headquarters; at the start of 2017 there were just
two. The average salary of an equity trader in the top 12 banks is $500,000,
so the savings have been significant. The majority of the work is carried out by
automated trading systems that are supported by 200 engineers (one third, or
9,000, of Goldman Sachs employees are computer engineers).
To automate more complex trades such as currency and credit, cleverer
algorithms are required. Where Goldman Sachs has automated currency trading, they have found that one computer engineer can replace four traders.
They are also looking at building a fully automated consumer lending platform that will consolidate credit card balances. This is an innovation that has

86

5 AI in Action

been incubated internally at the bank’s New York office by small ‘bubble’
teams. As Marty Chavez, the company’s deputy chief financial officer, says:
there is plenty of empty office space to use up.
The UK-headquartered bank HSBC has been running five proof-of-­
concepts (PoCs) that will utilise Google’s AI capabilities. Because their 30
million or so customers are engaging online much more, their data store has
increased from 56 petabytes in 2014 to more than 100 petabytes in 2017.
This means they are now able to mine much more value and insight from it.
One of the PoCs is to detect possible money laundering activity. Just as
with PayPal and USAA, they are looking for anomalous patterns that they can
then investigate with the relevant agencies to track down the culprits. As well
as improving detection rates, using the AI software (from Ayasdi) means that
there are less false-positive cases that have to be investigated, which saves the
time of expensive resources. In the case of HSBC, they managed to reduce the
number of investigations by 20% without reducing the number of cases
referred for further scrutiny. HSBC are also carrying out risk assessments
using Monte Carlo simulations (these were described in the Optimisation
section in Chap. 3). From these they will be better able to understand their
trading positions and the associated risks.
Some of the more contentious uses of AI in business are in recruitment and
policing. The main contention is how bias can be (or even perceived to be)
built into the training data set, which is then propagated into the
­decision-­making process. In Chap. 8, I will discuss this challenge and the
potential mitigating approaches.
The Police Force of Durham, a county in the North East of England, have
started using an AI system to predict whether suspects should be kept in custody or not (in case they might reoffend). In this case, there are lots of data
available, from previous cases of reoffenders and non-reoffenders over five years,
that can be used to predict future reoffenders. At the time of writing it would
seem that there still a number of issues with this approach, the biggest of which
is probably with the validity of the data—this comes only from the region’s own
records, so excludes any offending that the suspect may have done outside
Durham. Despite the data challenges, the system accurately forecast that a suspect was a low risk for recidivism 98% of the time, and 88% for whether they
were a high risk. This shows how the system deliberately errs on the side of
caution so that it doesn’t release suspects if its confidence is not high enough.
In recruitment, a number of firms are taking the plunge into using AI to
help short-list candidates (although there are only a few who will currently
admit to it). Alexander Mann, a Human Resource Outsourcing Provider,
initially automated some manual tasks such as interview scheduling and

How AI Is Generating Insights

87

authorising job offers. They have recently introduced some AI software
(Joberate) to help them find candidates that are good two-way matches with
the job requirement. The software analyses both the candidate’s CV and publicly available social media feeds to create profiles of the candidates.
But AI can mean more than just mitigating business risks. Insights from big
data analysis can also be used to help fight diseases, especially cancer. AI has
been applied to the molecular analysis of gene mutations across the whole
body—usually cancer treatment research relates to specific organs—which
means that treatments that have been developed for, say, breast cancer could be
used to treat colorectal cancer. Through this approach personalised treatments
become possible—statistically significant meta-studies have looked at how
much benefit there was in matching the molecular characteristics of the tumour
of a patient with their treatment. This matching resulted in tumours shrinking
by an average of 31% compared to 5% from a non-personalised approach.
AI is also being used to develop cancer drugs. A biotech firm, Berg, fed as
much data as it could gather on the biochemistry of cells into a supercomputer so that the AI could suggest a way of switching a cancerous cell back to
a healthy one. The results have so far been promising, leading to the development of a new drug, and, of course, those results have been fed back into the
AI to further refine the model.
The quality of healthcare can also be monitored using AI. The Care Quality
Commission (CQC), the organisation that oversees the quality of healthcare
delivered across the United Kingdom, implemented a system to process large
numbers of textual documents and understand the opinions and emotions
expressed within them. The CQC is now able to manage the influx of reports
with fewer people and, most importantly, can apply a consistent scoring
methodology. The use of sentiment analysis (which was described in the NLU
section of Chap. 3) is also used in business-to-consumer companies, such as
Farfetch, a UK online retailer, to understand in near real time what their
customers think of their products and services. The companies are then able
to respond quickly and get a better understanding of their customers’ needs.
Other use cases for AI in optimising processes include: predicting failure
for jet engines; predicting the risk of churn for individual customers; predicting localised sales and demand trends; assessing the credit risk of loan applications; predicting farming yields (using data from IoT sensors) and predicting
localised demand for energy.
From the above examples, it should be clear that AI has huge potential to
unearth hidden value from a company’s data, helping manage risks and make
better, more informed, decisions. There are certainly challenges around the
lack of transparency and unintended bias, but, managed correctly the data can
provide insights that would have been impossible for a human to discover.

88

5

AI in Action

The Established AI User’s View
This is an extract from an interview with John Sullivan, CIO of Virgin Trains
West Coast, one of the major train operating companies in the United
Kingdom.
AB Tell me how you first came across AI.
JS	Well, I actually studied artificial intelligence at college. At that time, it was
obvious to me how useful AI would be but it wasn’t yet very practical,
particularly in a business setting. But I did get a good understanding of
what it could do and why it is different from traditional systems.
AB And when did you start looking for AI opportunities at Virgin Trains?
JS	In my current role as CIO, I was interested in looking for applications for
AI that could solve some real issues. We had a customer relationship challenge in that we were getting lots of queries by email that took a long time
for people to deal with. Some of the same questions were coming up again
and again, day in, day out.
So we looked at implementing Celaton’s inSTREAM solution to
improve the service we were giving to the customers. By using AI we were
able to respond to the majority of these queries quicker and more consistently. It also made it more efficient for us as a business and the work
became more interesting for the staff.
AB What does the AI software do?
JS	Basically, it reads all of the incoming emails, which of course are all written
in free-form, and works out what the customer wants – it understands
whether their email is a question, a complaint or a compliment and sends
it to the right agent within the organisation to deal with. It also does a lot
of validation work at the same time – so if a customer is writing about a
specific train, the system will check that that train actually ran, and will
also see if there are any similar queries relating to that journey – that all
helps us prioritise the email.
We’ve managed to reduce the amount of effort required at that initial
stage by 85%, which far exceeded our expectations. All of the people that
were doing that very mundane work are now doing much more interesting
stuff dealing with the more challenging questions we get.
AB So, do humans still deal with the remaining 15%?
JS	That’s right. When we started off we tried to identify which types of queries the AI could respond to and which ones a human would need to do.
Because the system learns as it goes along, more and more of the work can
now be done by the AI. So the agents only work on the really tricky ones,

The Established AI User’s View

89

as well as over-seeing the work done by the AI. This oversight task is
important for Virgin as we want to get the tone of the responses just right
so that they align with the brand. We call it ‘bringing out the Richard
Branson’.
AB	So apart from improving customer service, what other benefits have you
seen?
JS	The really important thing for us is that it has allowed the business to be
scalable. We now have a core team of human agents that is stable, so that
any increase in volume can be handled by the AI. Email volumes are now
not an issue for us.
AB How did you approach the implementation?
JS	We found Celaton quite quickly and immediately got on well with them.
They came in and we did a prototype with the Customer Relationship
team. Doing ‘small speedboat’ projects always works best for me for trials.
It’s always better than trying to build an ocean liner, which we know takes
too long, and are difficult to stop!
AB What challenges were there in implementing the system?
JS	Change management is always going to be a challenge in projects like this.
The CS team, though, got on board really quickly because they actually
wanted it. Remember – this was a system that was going to take away all
of those mundane, repetitive queries for them. They couldn’t wait!
The other thing that can be tricky with an AI project is getting the data
right. We definitely needed to involve the IT team, but, because we had a
good vendor in, Celaton knew what the challenges would be so we could
prepare as much as possible. We relied on their resources quite a bit, as we
didn’t really have any AI capability in-house at the time, but they were
good people who could explain everything in a simple manner to our
people—they weren’t just a ‘technical’ supplier.
AB	Thanks John. And finally, what advice would you give anyone just starting
out on their AI journey?
JS	Well, the communications, which I just mentioned, are a really important
aspect. You need to be able to articulate what AI is and how it works –
don’t assume that even the CIO knows these things. If I were to do this
whole thing again I’d probably bring somebody in that could focus on
this. Almost like doing internal marketing for the project.
I also think it is vital to open your mind up to what AI can do—again
external input can be useful here. We have Innovation Days to look at
developments we are doing with the trains, and we really should try and
do the same sort of thing for AI as well. It’s all about trying to understand
the art of the possible.

6
Starting an AI Journey

Introduction
I am often asked by executives, “I need AI in my business—how can I implement it?”. This is, of course, the wrong question to ask. The much more apposite question is, “I have some big business objectives/challenges—how can AI
help me deliver or address them?” Ideally, executives should be pulling AI in
to where it is needed the most and will provide the greatest value. In reality,
though, it is usually a bit of push and pull—AI is such a new subject and has
such huge potential to disrupt whole business models that it would be foolish
to consider it only in the context of your existing business strategy.
So, if businesses are already implementing AI and deriving real value from
it, as we have seen in the case studies presented in this book, how did they
actually start their journey? How did they discover, plan and implement the
solutions? This chapter will cover the approach to creating a meaningful and
workable AI strategy.
I have deliberately blurred the lines between an ‘AI Strategy’ and an
‘Automation Strategy’. As is hopefully clear from Chap. 4 when I discussed
the associated technologies, AI is rarely the only answer. Usually other automation technologies are required, such as RPA, cloud and IoT. But with this
book’s focus on AI, I will be looking at the automation strategy very much
through the lens of AI, mentioning the other technologies only where I think
it is necessary for clarity.
If I could summarise the best approach to maximising value from AI it
would be: think first, try something out, then go large. It’s very tempting to
jump straight in and create a PoC or bring in some AI software, but creating
© The Author(s) 2018
A. Burgess, The Executive Guide to Artificial Intelligence,
https://doi.org/10.1007/978-3-319-63820-1_6

91

92

6 Starting an AI Journey

an automation or AI strategy, one that matches your ambitions, is based on
your business strategy and has explored the business at a high level for the
prime opportunities, will give you, by far, the best foundation to maximise
value.
Once you have the AI strategy sorted, then you can start trying things out.
This isn’t mandatory, but it usually helps to build knowledge, trust and
momentum within the business. Your initial steps could be through a pilot, a
PoC, a hard prototype or a purchased piece of software. I will explain each of
these and discuss their pros and cons in this chapter.
But once the initial efforts are in and proven, then you will need to look at
pushing on quickly. There can be, at this stage, what I call the AI Bermuda
Triangle—lots of really good ideas that were heading in the right direction
mysteriously disappear from the radar. This happens with other technology
programs as well—once the effort of getting that first piece of work off the
ground has been done, everyone sits back and becomes distracted by other
things, motivations dissipate and momentum is lost. Once this happens it is
very difficult to get it back on track again. Therefore, the time to really push
forward is as soon as that pilot or PoC or prototype is showing promise and
demonstrating value.
And that plan really needs to be big and bold if it is to escape from the AI
Bermuda Triangle. Creating an AI roadmap is a key part of the AI strategy—
after all, how are you going to go on an AI journey without a map to guide
you? But it’s more than just a list of projects—it needs to describe how AI will
be industrialised within the business, and I’ll cover those aspects in more
detail in the penultimate chapter.
So, armed with the knowledge of what AI is capable of and how other businesses are using it, it is now time to start your own AI journey.

Aligning with Business Strategy
In my work as a management consultant and AI specialist, I know that the
one single activity that will ensure that maximum value is extracted from any
AI program is to first create an automation strategy that aligns with the business strategy, but also one that challenges it.
To align the automation strategy with the business strategy it is necessary
to understand the benefits and value that will be derived from that overall
strategy. Business strategies, if they are written succinctly, will have a handful (at most) of strategic objectives—these might be things like ‘Reduce the

93

Aligning with Business Strategy

cost base’, ‘Reduce the exposure to internal risk’ or ‘Improve customer service CSAT score’.
Each of these strategic objectives would deliver benefits to the business:
‘Reducing the cost base’ would deliver lower costs through, for example, not
hiring any new staff, not taking on any additional office space or minimising
travel. ‘Reducing the exposure to risk’ might be delivered through reducing
the number of unnecessary errors made by Customer Service staff, or improving reporting and compliance. And ‘Improving Customer Service’ could be
fulfilled by improving the average handling time (AHT) of inbound queries,
reducing the unnecessary errors (again) and enabling 24×7 servicing.
The important thing from our perspective in creating an automation strategy is to understand its role in enabling some or all of these benefits.
In a hypothetical example: AI Search could be used to reduce AHT by
reading incoming documents; Optimisation tools could be deployed to provide knowledge support to the customer service agents; self-service could be
enabled through a number of AI and RPA tools so as to reduce the demand
on bringing new staff and renting more office space; Clustering and Search
capabilities could be used to offer deep reporting insights into the business
information; Search could be used to identify areas of non-compliance against
a constantly updating regulatory database and chatbots (using NLU) could be
used to provide a 24×7 service desk. RPA could also be brought in to eliminate errors. Each of these tools then becomes an enabler to the business strategy benefits (Fig. 6.1).
These strategic
objectives…

…need to deliver these
benefits…

No new hires

Back Office Automation

No new facilities

“Improve customer
experience”

Analytics & Reporting

Improve AHT

Customer Centricity

24x7 servicing
Reduction in errors

Fig. 6.1 Aligning with the business strategy

Plan

3

Improved reporting
Improve compliance

3rd Parties
ties

“Reduce risk
exposure”

Self Service

Improve AHT

Tech3

2

Minimise travel

Tech2

“Reduce cost base”

Tech1

1

…enabled by
automation:

94

6 Starting an AI Journey

So, before we have even started creating the automation strategy, we must
know what it should achieve. Then it is a case of working backwards to determine the capabilities, technologies, people, plan and so on that will deliver
those objectives.
Of course, there is nothing to stop you from implementing AI simply
because you want to implement AI. In fact, demonstrating to shareholders or
customers that you are an innovative, forward-thinking company because you
are using AI can be of great benefit, and much PR can be made of the fact.
But, generally, aligning your Automation Strategy firmly with the Business
Strategy will ensure that the benefits you are hoping to achieve will deliver
long-term, sustainable value to the business.

Understanding Your AI Ambitions
The second major aspect to consider before starting your AI journey is to
understand, as much as possible, where you want to end up. That may sound
obvious, or you may think why should you bother thinking about that now
when you haven’t even started anything yet. But, as with any journey, it is
crucially important to know your destination. In AI terms, you won’t be able
to know exactly where you will end up (it’s a voyage of discovery, after all) but
you should at least understand your initial AI ambitions.
These ambitions can range from anywhere between ‘we just want to say
we’ve got some AI’ to creating a completely new business model from it. There
are no right and wrong answers here, and your ambition may change along
the way, but knowing your overall aspirations now will mean that you can
certainly start on the right foot and be going in the right direction.
I would consider four degrees, or levels, of ambition, which I have called:
ticking the AI box, improving processes, transforming the business and creating a new business.
The first of these, ‘ticking the AI box’, is for those who just want to be able
to claim in their marketing material and to their customers that their business, service or product has AI ‘built in’. This approach is surprisingly common, but I’m not going to focus too much on this approach because it can be
covered by selecting the most appropriate elements from the other types of
approaches. To be honest, you could justifiably say that your business uses AI
now because you filter your emails for spam, or that you use Google Translate
occasionally. Many businesses that have deployed simple chatbots claim to be
‘powered by AI’, which is factually true but a little disingenuous.

Understanding Your AI Ambitions

95

The first serious step to extracting value from AI will be in improving the
processes that you already have, without necessarily changing the way the
function or the business works. AI is being used to make existing processes
more efficient and/or faster and/or more accurate. These are the most common uses of AI in business right now, and inherently the least risky; process
improvement is the approach that most executives will consider first when
deciding on their initial steps into the world of AI.
I’ve discussed case studies where processes can be streamlined through the
application of AI Search (extracting meta-data from unstructured documents,
for example) or by using big data to carry out predictive analytics (more
­accurate preventative maintenance schedules, for example). Other examples
of process improvements are where AI has helped to filter CVs for job candidates, or helping make more efficient and more accurate credit decisions.
Whilst making existing processes faster, better and cheaper can provide a
great deal of value to a business, there is arguably even more value available
through transforming the processes or the function. By transformation I
mean using AI to do things in a materially different way, or in a way that
wasn’t even possible before. Some of the examples of transformation that I
have discussed earlier include: analysing the customer sentiment from hundreds of thousands of interactions (NLU); predicting when a customer is
going to cancel their contract (Prediction); recommending relevant products
and services to customers (Clustering and Prediction); predicting demand for
your service (Clustering, Optimisation and Prediction); or modelling different risk scenarios (Optimisation).
One transformation that AI is particularly suited for is the enablement of
customer or employee self-service. The benefits of a self-service option are that
it can be made available twenty-four hours a day, seven days a week, and that
it is generally cheaper to run. It also gives the customers or employees a sense
of empowerment and control. AI can be used for the direct engagement
aspects of the process, employing chatbots and/or speech recognition capabilities to communicate with the person, and also for any decisions that need
to be made based on that communication (e.g. should this request for credit
be approved?) by using prediction or reasoning tools. (Also useful when creating a self-service capability is RPA, which is able to handle all of the rules-­
based processing and connect all of the necessary systems and data sources
together in a non-disruptive manner, as discussed in Chap. 4.)
The biggest impact that AI can make on a company is when a whole new
product, service or business can be created using the technology at its core.
Probably the most famous example of this is Uber which uses a number of

96

6 Starting an AI Journey

different AI technologies to deliver its ride-hailing service. For example, it
uses AI to provide some suggestions as to where you might want to go to,
based on your trip history and current location (e.g. I usually want to go
straight home after I am picked up from the pub). It will also help predict
how long it is going to take until the driver arrives to pick you up. Uber have
also analysed the time it takes to actually pick people up (i.e. from the car
arriving at the site to when it leaves again) to enable the app to suggest the
most efficient pickup spots. (After using third-party navigation apps, Uber
eventually developed its own navigation capability, although it still relies on
data from third parties.)
Other types of businesses have been created from a foundation of AI. (I’m
not considering AI software vendors or AI consultancies here as they will
inherently have AI at their core.) Powerful recommendation engines have
been used by companies such as Netflix and Pandora to transform their businesses, and Nest, a ‘smart thermostat’ that uses predictive AI to manage the
temperature of your house; without AI Nest would just be another thermostat. Other businesses that have initially used AI as a foundation for their core
business are exploiting the technology to create new revenue. Pinterest, the
website where people post interesting images from other sites, is a good example of this. They developed a very strong image recognition system so that
users could find, or be recommended, similar images. They are now developing apps based on the same technology which can automatically detect multiple objects in images and then find similar images of the objects on the
internet, including links to buy those objects (from which they are bound to
take some commission).
So, understanding your AI ambitions is an important early step in developing an AI strategy and subsequent AI capability. Those ambitions will guide
the first steps that you take and help steer you on the overall journey. But do
bear in mind that not every part of your business will want or need to go on
the same journey or at the same speed. Your front office might want to adopt
AI wholesale whilst the back office is happy to take things slowly. And each
department or function might be starting with different levels of automation
maturity. If there are functions that have already adopted some form of automation or have made efforts to organise their data, these will provide stronger
platforms to introduce AI than somewhere that still operates very manually or
doesn’t have a strong data governance.
The next section introduces the AI Maturity Matrix that can help you assess
both the current automation maturity as well as your ambitions for each significant area of your business.

Assessing Your AI Maturity

97

Assessing Your AI Maturity
A Maturity Matrix is a relatively simple concept that is a very useful on a
number of fronts. Not only does it encourage discussion and reflection during
the creation process but, once completed, it can also be used as a communication tool.
Maturity Matrices were originally developed by Carnegie Mellon University
to assess the maturity of IT development functions. They generally describe
five levels of maturity, from very immature (Level 1, or ‘Initial’) to world class
(Level 5, or ‘Optimising’). Each maturity level consists of related practices for
a predefined set of process areas that improve an organisation’s overall performance. Thus, an organisation achieves a new level of maturity when a system
of practices has been established or transformed to provide capabilities the
organisation did not have at the previous level. The method of transformation
will differ at each level, and requires capabilities established at earlier levels.
Consequently, each maturity level provides a foundation on which practices
at subsequent maturity levels can be built.
Capability Maturity Model (CMM) levels in the IT development world
can be formally assessed by approved consultancies, and many big IT companies wear their CMM Level 5 badges with pride. But not everyone should
want or need to reach Level 5—in many cases Level 3 (‘Defined’), where
processes are documented, standardised and integrated into an organisation,
is perfectly adequate for most businesses.
This idea of evaluating what is ‘good enough’ for your business, and each of
the major functions in it, is at the heart of developing the AI Maturity Matrix.
The methodology I describe in this section takes the concept of the Maturity
Matrix and applies it specifically to automation. As I mentioned when discussing the Automation Strategy generally, it is best to try and think holistically about automation but with a strong focus and lens on AI, just so that
related opportunities, or dependencies, are not missed along the way. So, in
the matrix we are considering ‘automation’ that includes the full range of AI
tools (chatbots, search, data analytics, optimisation engines, image classification, voice recognition, etc.) as well as RPA, robotics, IoT and Crowd Sourcing
(which are described in detail in Chap. 4).
The AI Maturity Matrix has six levels rather than the usual five—this is to
introduce a Level 0 where there is no automatson at all. Each of the six levels
is described thus:
Level 0: Manual Processing There is very little evidence of any IT automation
in the organisation. Only basic IT systems such as email and ‘office’ applications
are deployed. There are large numbers of people processing transactional work,

98

6 Starting an AI Journey

either in-house or through an outsource provider. Data is not considered an
asset and there is no formal governance in place to manage it. There are no significant projects looking at, or delivering, automation.
Level 1: Traditional IT-Enabled Automation The organisation has implemented task-specific IT applications for particular processes (e.g. an invoice
processing application to process invoices). There is no evidence of automation tools, and specifically AI nor RPA, having been deployed. Data is
­managed only to the extent that it needs to be to ensure the smooth running
of the organisation. There are still large numbers of people carrying out transactional and cross-system processing work.
Level 2: Isolated, Basic Automation Attempts Some individual teams have
used scripting or macros to automate some tasks within a process or application in isolated areas of the business. Minimal benefits are evidenced from
these efforts. There are still large numbers of people carrying out transactional
and cross-system processing work. Data is still managed on a nominal basis.
Level 3: Tactical Deployment of Individual Automation Tools Some functions have deployed individual automation tools such as AI and RPA to automate various processes. Some useful benefits have been identified. No
dedicated automation resources are used or have been set up. Some data is
managed so as to make it valuable for the automation efforts, but an
organisation-­wide data governance framework does not exist. In areas that
have experienced automation, the staff are working in different ways. A case
study or two from the organisation is known within the industry.
Level 4: Tactical Deployment of a Range of Automation Tools Functions or
divisions within the organisation have deployed a number of different automation tools, including AI, across a range of processes. A strong business case
exists that identifies sizeable benefits. Data is being managed proactively, with
some areas having introduced data management and retention policies. Staff
may have been redeployed as a result of the automation efforts, or they are
working in materially different ways. Some dedicated resources have been
organised into an automation operations team. The organisation has a reputation for implementing new technologies and being innovative.
Level 5: End-to-End Strategic Automation The organisation has implemented a strategic end-to-end process automation program with a range of
automation tools including AI. Significant benefits in cost, risk mitigation

Assessing Your AI Maturity

99

and customer service have been delivered. Data is treated as a valuable asset
and is managed through an organisation-wide data governance framework.
Many of the organisation’s staff are working in materially different ways than
before. An Automation Centre of Excellence has been established. The
­organisation is renowned for its innovative and forward-thinking culture.
In the above descriptions, I have used the word ‘organisation’ to describe
the scope of the work being evaluated—this could indeed be the whole organisation but could equally be applied to parts of it, such as the Finance
Department, the Customer Services Department or a Shared Service Centre.
When choosing the level of granularity to assess, it should be at the highest
point (i.e. the biggest scope) that has a significant difference in automation
approach and needs. Generally, this will be by function or division, but could
also be by geography.
The evaluation of maturity levels is a subjective one, assessed through interviews and the review of evidence. The approach should be one that is satisfactory to all people who are involved and must be as consistent as possible across
the different areas. Many people use a third-party consultancy to carry out the
evaluation to ensure consistency but also provide independence from any
internal politics there may be.
There can be some overlap between the different levels. For example, a
department could have tactically deployed a range of automation tools (Level
4) but also created an Automation Centre of Excellence (a Level 5 criteria). It
will be up to the person or team evaluating the levels to decide the most
appropriate level to assign it to. The important thing is to be consistent.
As well as assessing the current automation maturity, the ‘automation
ambition’ should also be agreed upon. As mentioned earlier, the target level
doesn’t necessarily have to be Level 5, and could be different for different areas
that are being assessed. There are many reasons why not all areas should strive
for Level 5, including the cost of implementation, there not being enough
relevant data, it not being aligned with the strategic objectives and it simply
not being appropriate (a human touch might be the best one in some cases).
A completed AI Maturity Matrix would therefore look something like the
following (Fig. 6.2):
It shows both the ‘as is’ level of maturity (in grey) and the agreed ‘automation ambition’ for each area (in black). Although it is a relatively simple chart,
it lays the foundations for the Automation Strategy and Roadmap. It also
provides a useful communication tool to describe, at a high level, what the
business is hoping to do with automation and AI.

100
Maturity Level
Process Area

6 Starting an AI Journey
0

1

2

3

4

5

Manual processing

Traditional ITenabled automation

Isolated, basic
automation attempts

Tactical deployment
of individual
automation tools

Tactical deployment
of a range of
automation tools

End-to-end strategic
automation

Customer Service
Risk Assessment
Operations
Finance
HR
ITSM

Fig. 6.2 AI maturity matrix

Levels 0, 1 and 2 do not involve the sort of automation that I am considering in this book. Level 3 is where AI (and RPA) are starting to be introduced,
generally focusing on one or two individual processes. This equates to the
‘ticking the AI box’ ambition discussed in the previous section. Level 4 applies
AI to a wide range of processes, applying it across a number of areas, although
still tactically. This level equates to the ‘process improvement’ ambition. Level
5, where AI is applied strategically across the business, is equivalent to the
‘transformation’ ambition.
The ambition I described of creating a new service line or business by
exploiting AI goes beyond the maturity matrix structure as it is not based on
an existing function or department. That level of ambition means starting
with a blank sheet of paper.
The AI Maturity Matrix then is a useful tool to start your AI journey. It
provides an opportunity to openly discuss the role and opportunities for AI
and automation across the business and provides a platform to develop a ‘‘heat
map’ of what those opportunities might be. The approach to creating an AI
heat map is described in the next section.

Creating Your AI Heat Map
Following the development of the AI Maturity Matrix, the next step in creating an AI Strategy is to build a ‘Heat Map’ of where the opportunities for
automation might be. Driven by the Business Strategy and the benefits that it
is hoping to achieve, the Heat Map provides a top-down perspective on areas
where AI is desirable, economically viable and/or technically feasible. It starts
to identify the types of AI capabilities that could be applied in each area in
order to realise the automation ambitions (and therefore contribute to delivering the business’s Strategic Objectives).

Creating Your AI Heat Map

101

The approach to creating an AI Heat Map is not too complicated as long as
you understand your business well enough, and, of course, have a sound grasp
of the AI Framework. It is intended as a starting point, a way to bring some
focus and logic to prioritise your initial AI efforts.
Firstly, you must decide on the overall scope of the Heat Map. It usually
makes sense to keep this consistent with the Maturity Matrix. So, if you originally evaluated the whole business split down into, say, five different business
areas, then use the same structure for the heat map.
Each area, then, is assessed in turn, identifying the opportunities and the
relevant AI capabilities required in each one. This is best done in two passes,
the first one identifying all opportunities without any judgement being
applied to them, rather like the brainstorming of ideas: all potential opportunities should be logged, dismissing none of them at this stage.
The opportunities are identified by considering a number of different criteria, each of which is described below. The opportunities are surfaced through
interviews with relevant managers in the areas being assessed, and should be
done by people who have a sound understanding of the AI capability framework as well as the AI technology market. Organisations without the appropriate resources in-house will tend to use third parties for this exercise.
• Alignment with strategic objectives—unless you are looking to implement AI for the sake of it, it is important to identify AI opportunities that
contribute in some way to achieving the Strategic Objectives. For example,
if your only strategic objective is to improve customer satisfaction, then the
focus for this exercise should be on those processes that can impact customer service, rather than, say, cost reduction.
• Addresses existing challenges—there may be opportunities for AI to solve
existing problems, such as inadequate management information, poor
compliance or high customer churn. These opportunities may or may not
be aligned with the strategic objectives but they should still be considered
for inclusion as they could certainly add value.
• Available data sources—because AI, in most of its forms, requires large
amounts of data to be most effective, a key consideration is to identify relevant data sources. Where there are very large data sets, then it is possible
that there is an opportunity for AI to extract some value from them.
Conversely where there is very little, or no data, it may preclude using AI.
(Some AI technologies, such as chatbots and cognitive reasoning engines,
both discussed in Chap. 3, do not require large data inputs, just captured
knowledge, to work effectively, so don’t simply equate low data volume to
no AI opportunities.)

102

6 Starting an AI Journey

• Available technology—understanding the AI capabilities and the relevant
tools that can deliver them is also important in order to identify AI opportunities. Whilst it is always best to ‘pull’ ideas in based on what the business
requires, with AI at the forefront of technology it is also valid to ‘push’ ideas
based on the technology that is available.

Customer Service
Risk Assessment
Operations
Finance
HR
IT
Fig. 6.3 AI heat map first pass

Crowd

IT automation

Image

Voice

Fraud

Risk

Process

Analytics

Chatbots

Function

Search

Automation Type

RPA

By this point you should have a range of different opportunities identified
in each of the respective areas you are investigating. For example, in Customer
Services, you could have identified the opportunity to provide an online self-­
service capability for the purchase of tickets for events. This could help deliver
your strategic objectives to improve the customer experience and to increase
ticket sales. In terms of the AI capabilities it could require Speech Recognition
(if you want to offer telephone access), NLU (for the chatbots) and
Optimisation (to guide the user through the purchasing process). It could also
require some RPA capability to actually execute the purchase. (I would also
include Crowd Sourcing, discussed in Chap. 4, as another capability that
should be considered in order to support any of the AI opportunities.) With
regard to data sources, there are pre-trained voice services available, and there
is a good source of knowledge available from the human customer service
agents that currently process these transactions (Fig. 6.3).
For each of these ideas, you should therefore understand what the opportunity is, how it links to strategic objectives (and/or addresses current issues) and
which AI capabilities you will require to realise it. Building each of these on top
of each other then starts to create a heat map of AI benefits and requirements.

Creating Your AI Heat Map

103

(I tend to use colours to identify the ‘hottest’ areas, but you can use numbers
or whatever approach you are used to.) Areas that are delivering the most benefit are clearly visible, as well as those AI capabilities which are going to be most
important to you (Fig. 6.4).
The second pass across the various opportunities is to filter them down to
those that are desirable, technically feasible and economically viable. If the
idea can’t satisfy all of these criteria then it is unlikely to be successful.
Desirability is a measure of how much the business wants and needs this
new idea, so it relates to how aligned it is with the strategic objectives and how
it addresses existing challenges. But it should also consider the opportunity
from the customer’s perspective (if it impacts them) and, culturally, how
acceptable the idea will be to the department or function that will be implementing it. In some cases, you will have to consider the personality of the
managers and staff of those areas as well and understand how supportive (or
defensive) they may be. Clearly, if an idea does not pass the desirability test (to
whatever threshold you think is appropriate for your organisation), then it
probably shouldn’t be progressed any further for now.
Technical feasibility will already have been assessed to a certain extent
during the first pass when the data sources and the availability of technology
were taken in to account. For the second pass this is done in more detail,
considering aspects such as the quality of the data, the processing power or
bandwidth that may be required, the maturity of the required technology and
the technical skills that will be required internally and externally to implement it. Other aspects such as regulatory restrictions, especially about the use
of data, may also have to be considered.
Testing the economic viability is the first look into what the business case
might be like. A comprehensive business case will be developed in time, but
at this stage the economic viability can be assessed at a relatively high level. It
should start to consider the financial benefits to the business which could be
achieved through cost reduction, risk mitigation, improved debt recovery,
new revenue generation or reduced revenue leakage. Estimates can be made as
to how the opportunity might improve Customer Satisfaction (CSAT) scores,
if relevant. Costs should also be assessed where they can be at this stage. These
could include license fees, IT infrastructure and professional services fees. It
may be difficult to assess the costs at this early stage; therefore, it could be
appropriate to score, out of 10, say, the magnitude of the costs, with anything
scoring above a certain threshold rejected (this obviously needs to be balanced
against the benefits that will be gained).
Each idea should be able to pass all three tests if it is to survive into the final
Heat Map. You can, of course, make exceptions, but be clear on why you are

Opportunity 1

Opportunity 2

Opportunity 1

Fig. 6.4 AI heat map

Finance

Operations
Opportunity 3

Opportunity 2

2

1

1

Customer Services
Opportunity 1

1

Organisation

AI Heat
Map
Maturity

Existing
Challenges

Strategic
Objectives

Strategic Objective 1

Challenge 3

Strategic Objective 3

Strategic Objective 2

Challenge 1
Challenge 2

Cost Reduction
Customer Service

Benefits

Compliance
Risk Mitigation
Loss Mitigation
Revenue Generation
Leakage Mitigation
Image
Voice

AI Capabilities

Search
NLP
Planning
Prediction
RPA
Crowd

104

6 Starting an AI Journey

Developing the AI Business Case

105

making those exceptions, and keep them in mind as you progress forward. Of
the rejected ideas, do keep them safe. Things change, especially around the
technological viability, so something that may not be suitable now may be just
right in the future.
So now you have an AI Heat Map which shows the main areas you will be
focusing on, the benefits they can bring and the capabilities required to
achieve them. Each opportunity, as well as the high-level view presented in
the Heat Map, should also have a corresponding passage of text (or a separate
slide) which provides more of the necessary details about what the opportunity actually entails and some explanation of the numbers.
As a summary, and for presentation purposes, the AI Heat Map can be
rolled up into an overview across the main areas and the organisation as a
whole.
Now that we have a good idea of the opportunities that form the basis of
the Automation Strategy, the next stage is to develop those Heat Map opportunities into a Business Case.

Developing the AI Business Case
Creating a Business Case for an AI project is, in many ways, the same as for
any technology project—there are benefits to be achieved and costs to incur.
But with AI the task is made more challenging because there tends to be more
unknowns to deal with. These unknowns can make calculating an ROI
(Return on Investment) an almost-futile exercise, and the organisation must
rely on an element of intuition as well as the numbers. This is especially true
when new approaches and new services are being developed through AI.
Luckily, the work done creating the AI Maturity Matrix and Heat Map
provides a good methodology for evaluating which are the AI opportunities
that will deliver the most value. Creating a meaningful business case for AI for
each of these is certainly feasible with a little thought.
We already have, from the AI Heat Map, a list of opportunities and a high-­
level assessment of which are the most promising. In my example in the previous section the opportunities which are showing as the ‘‘hottest’ (dark grey)
are the ones to prioritise, especially if they are strong across a number of different aspects (strategic alignment, solving current challenges, benefits types).
At this stage, you can easily apply some scoring mechanisms to make the
prioritisation process easier. For example, you could replace the shades or
colours with scores out of 3, and give weightings to each of the criteria, and
then add all these together to give a total score for each opportunity.

106

6

Starting an AI Journey

If you are feeling confident, or have a third-party supporting you in this
exercise, you could also introduce additional criteria under the broad heading
of ‘Ease of Implementation’. This would be influenced by the maturity of the
function (from the Maturity Matrix) but also considers the technical feasibility and desirability that we looked at when filtering the opportunities during
the second pass of creating the Heat Map. Scoring each opportunity in this
way, again with appropriate weightings, will give additional depth to your
decision-making process.
From this work, we now have a prioritised list of AI opportunities to consider: it’s time to start getting serious with a few of these. Whether you take
just the top scoring idea, or the top 3 or top 10 is entirely up to you, and will
be based on your overall AI ambitions and the time and resources you have
available. However many you take, it is likely that you will need to draft in
some specialist resources at this stage, either from your own organisation or
from third parties.
Different organisations have different ways of calculating business cases.
Some will focus on ROI, some Net Present Value, some internal rate of return
and others payback. I’m not going to explain each of these here as they are
pretty standard approaches, and I am assuming that you are familiar with the
one that is favoured by your own organisation. But all of these do require
calculating the benefits and the costs over time, and so I will provide an aide
memoire of what sorts of things you might need to include in each of these
areas.
The benefits that AI can deliver are significant and diverse, and can be split
into those that are ‘hard’, that is, can easily be equated to monetary value, and
those that are ‘soft’, that is, those that are more intangible and difficult to
quantify. It should also be noted that a single AI implementation may deliver
a range of different benefits, not just one specific type. This is an important
point—whilst you may implement an AI system to, say, reduce costs, there
could well be associated benefits in compliance or risk mitigation. For each
opportunity identified in the AI Heat Map, consider whether each of the
benefit types described below could be applicable.
The Hard Benefits can be categorised as follows:
• Cost reduction: this is the simplest benefit to quantify as there is usually a
base cost that can be compared to a future, lower cost. For AI, these are
generally the situations where the new system will replace the humans
doing a role or activities (remember the discussion in Chap. 1 around
replacement versus augmentation?). The AI search capability will replace
the activities of reading and extracting meta-data from documents—it will

Developing the AI Business Case

•

•

•

•

107

do it faster and more accurately than humans which means the benefits will
be magnified. Chatbots (using NLU) can also replace some of the work
done by human call centre agents, as can speech recognition when combined with a cognitive reasoning engine. NLG (a sub-set of NLP) can
replace business analysts in creating financial reports. More efficient route
planning for workers and/or vehicles reduces time and cost.
Cost avoidance: for businesses that are growing, cost avoidance is the more
acceptable face of cost reduction. Rather than recruiting additional people
to meet demand, AI (as well as other automation technologies such as
RPA) can be implemented as an alternative. The AI solutions will be similar to those for cost reduction.
Customer satisfaction: this is usually measured through a survey-based
satisfaction index such as CSAT or Net Promoter Score (the difference
between positive and negative customer feedback). Some businesses have
connected this to a monetary value, and others relate it directly to a manager’s own evaluation and bonus. For many it is a key performance indicator. AI has the opportunity of improving customer satisfaction through
being more responsive to queries, being more accurate in the responses,
providing richer information in the responses, reducing the friction in customer engagements and in making relevant recommendations. (Of course,
AI can also be used to automatically measure customer satisfaction itself
through sentiment analysis.)
Compliance: some people might challenge the fact that a self-learning system can improve compliance which is based on hard and fast rules, but
actually AI is very good at identifying non-compliance. Using NLU and
Search it can match policies and procedures against regulations and rules
and highlight the variances. Compliance benefits can be measured through
the potential cost of not being compliant, such as through fines, or loss of
business in specific markets. (RPA, by the way, is a strong driver of compliance as each automated process will be completed in exactly the same way
every time, with each step taken logged).
Risk mitigation: AI can monitor and identify areas of risk in areas where
it would be impossible, or very expensive, for humans to do it. The classic
example for AI here is fraud detection in high-volume transactions, such as
credit card payments. AI can, in some circumstances, make better risk-­
based decisions than humans, such as for credit approvals. AI can also contribute to the Know Your Customer (KYC) process in being able to validate
documents and data sources and help with the credit checking process
(RPA also helps here to access the necessary systems and run the process).

108

6

Starting an AI Journey

As with compliance, risk mitigation can be measured by the costs that have
been avoided, such as the loss through fraud or poor credit decisions.
• Loss mitigation: loss mitigation can be achieved through improved debt
recovery. Most of the heavy lifting here can be done through RPA by managing the timing and activities in the process, but AI can play a role, for
example, in generating appropriate letters or engaging with the debtors.
Loss mitigation can be measured through the increase in cash being
­recovered through the use of automation.
• Revenue leakage mitigation: this is usually achieved through reducing the
situations where revenue generating opportunities are lost, such as when
customers no longer use your services. Reducing customer churn through
the early identification (i.e. clustering) of the tell-tale behaviours can be
measured by calculating the revenue value of a customer. Other AI capabilities, such as NLU and prediction, can also contribute by keeping the
customer more engaged with your business.
• Revenue generation: this is the AI benefit that can probably deliver the
most in terms of monetary value. AI can help enable self-service capabilities (through Speech Recognition, NLU, Optimisation and RPA, for
example), identify cross-selling and up-selling opportunities (through
Clustering) and create new revenue generating opportunities either from
your existing business or through completely new services or products.
Revenue generation is easy to measure after the fact, although isolating the
specific contribution of the AI may be challenging and assumptions will
need to be made. Predicting revenue generation will need to use modelling
techniques, most of which should already be available in your business.
One thing to bear in mind with measuring automation benefits is that, in
some cases, staff productivity could actually be seen to decrease after implementation. This is not a bad thing, but just a reflection that the staff are now
handling the more complex cases whilst the technology is handling the simpler ones. Overall, the productivity will have improved, but if you just look at
the human staff they may well be taking longer to process the (more complicated) cases.
The softer benefits are inherently more difficult to put a monetary value to,
but can still be evaluated, even if the accuracy is less certain:
• Culture change: this is probably the most difficult thing to achieve and the
most difficult to measure, but can deliver significant value to an organisation. Of course, it depends on the type of culture you have already and

Developing the AI Business Case

•

•

•

•

109

what you would like to change it to, but AI can help embed a culture of
innovation as well as deliver customer centricity into the business. Actual
benefits will usually be associated with a wider change program and it will
be challenging to isolate these that are specifically due to AI, but, because
of the potential value, these should not be ignored.
Competitive advantage: this can deliver significant benefits if AI provides
a first-mover advantage, a new service line or access to new markets or customers. The benefits will be closely associated with (and usually accounted
for in) some of the ‘hard’ ones such as cost reduction, customer satisfaction
and revenue generation, but competitive advantage should always be
sought as a way to deliver a step-change in value.
Halo effect: AI, at least in the current decade, is a highly marketable aspect
to have within a business. The objective of being able to ‘tick the AI box’
can provide the opportunity to claim you are an ‘AI-enabled’ business
which could attract new customers to your product or service. Any amount
of AI implementation, if marketed correctly, can also have a positive impact
on shareholder value if it results in an increase in the share price of the
company.
Enabling other benefits: as well as providing direct benefits, AI can also
enable other, indirect benefits to be delivered. For example, implementing
AI can free up staff that can be deployed on higher-value activities, or data
that is generated from an automated process can be used to provide additional insights that can be exploited in another part of the business.
Implementing AI may, through eliminating much of the mundane work in
a department, also reduce employee churn.
Enabling digital transformation: transforming businesses to be more
‘digital’ is a common strategic objective for many businesses right now. AI
is clearly part of that transformation, and many of the benefits of a digital
business can be directly delivered, or indirectly enabled, through the implementation of AI.

An AI business case will include at least one of the above hard and soft
benefits I have listed and often quite a few of them. The individual classes of
benefits should be used as a starting point for your AI business case. For each
AI opportunity identified in the heat map, consider each type of benefit carefully so that all can be identified and none are forgotten or left out.
Once all the relevant calculations have been done, the benefits for each
opportunity can be summed across the departments and the organisation to
give a full picture of the value that AI can bring. Also, consider whether there

110

6

Starting an AI Journey

are any synergies between the different opportunities—for example, does having a number of AI initiatives going on enable a favourable cultural change in
that department?
The heat map, bolstered with the business case data, now provides a comprehensive prioritised list of AI opportunities within your business. But,
before you take the first step into prototyping and implementing the most
favourable candidates, you will need to understand what your ultimate roadmap looks like and how you will handle all the change management aspects
that go along with it. These two pieces of the AI strategy jigsaw are covered in
the next two sections.

Understanding Change Management
Even the simplest IT projects require change management. With automation
and AI, the challenges are magnified—not only are you going to fundamentally change the way people work, but you may also have to let people go. And
even if this is not your intention, the reputation of automation projects means
that staff will likely think it is the case.
You may, as part of your AI strategy, be creating a new product or service
that will be competing with your legacy products or services. That ‘healthy
competition’ may be good for business but is not necessarily good for staff
morale. And building a new capability will mean transforming your organisation—new jobs will be created and old jobs will disappear. Some are going to
be able to cope with the change, and some are not.
Right at the beginning of the book I talked about the differences between
AI replacing humans and AI augmenting them. Clearly an augmentation
approach will have less change management issues than a replacement one.
But anything that can demonstrate how AI will enrich the work that people
are doing (even if they have to do it differently) will help. Having greater
insights into their work, or customers, or suppliers, through AI analysing and
interpreting data, will help people become more successful in their jobs.
If automation is replacing tasks, it is often for the more transactional, tedious
ones. There is a useful phrase coined by Professors Willcocks and Lacity, which is
that “automation takes the robot out of the person” (Willcocks and Lacity 2016).
So, if your AI projects in any way relieve people of the mundane, repetitive tasks
that they have been doing for years, then this should be celebrated.
Despite the amount of change that AI inherently throws up, many of the
common change management practices are relevant to AI projects: communicate early and often; involve and empower employees as much as possible;

Understanding Change Management

111

look for short-term wins; consolidate gains to generate further momentum;
incentivise employees where appropriate; pace the change appropriately and
have honest reviews before the next one.
A challenge that automation projects in particular face is a post-prototype
slump. The initial build can become a victim of its own success—the relief of
actually creating something that works, even if it is not a complete solution,
can be so overwhelming that everyone then takes a step back, resulting in a
loss of focus and momentum.
To overcome this, it is first necessary to be aware of it, and then to prepare
for it. Schedule in review sessions for as soon as the prototype is due to be
completed. Have a Steering Group Meeting soon after, or get the CEO to
make a celebratory announcement about how this is ‘the end of the beginning’. If possible, have the next phase of the project, or another prototype,
start before the first project finishes, so that focus can quickly shift to that.
The worst thing you can do at this point is to pause for a few weeks or months.
Generally, a ‘big and bold’ approach to an AI program will deliver the most
success, as it avoids any potential slump periods, and helps maintain that all-­
important momentum. However, all organisations are different and an ‘under
the radar ‘approach may be the most suitable for you—just make sure that
your business case and roadmap (covered in the next section) reflect that.
As well as managing all the change within your organisation, you may have
to manage change for your customers as well. If your AI project is one that
impacts the way customers engage and interact with you, then you will need
to fully prepare them for that change. This will involve plenty of communication, supported by marketing efforts and relevant collateral such as emails,
website articles and frequently asked questions (FAQs).
If your AI project changes the way that you use your customers’ data, then
you will need to make it clear to them what the impact of this is. Any sensitivities around data privacy, in particular, will need be made explicit. You would be
well advised to seek legal advice for any changes to your customers’ data usage.
This isn’t a book about how to manage change (there are plenty of those
about already), and you will likely already have your own preferred methodologies to tackle this. This is only to say that you should not neglect change
management in any way—implementing AI brings about its own unique
challenges, especially when some of the changes in an organisation can be
seismic in nature. These challenges need to be understood and carefully managed, and built into your AI roadmap.

112

6 Starting an AI Journey

Developing Your AI Roadmap
The AI Roadmap provides a medium- to long-term plan to realise your AI
Strategy. It can be a relatively simple affair that draws heavily from the AI
Maturity Assessment and the AI Heat Map.
The Maturity Assessment will provide you with your starting point. Those
areas that are immature when it comes to automation will require more effort
and time to complete. Your Change Management Assessment will also direct
you to those areas that might need additional support along the way.
At the core of the Roadmap will be the opportunities identified through
the AI Heat Map and subsequent Business Case. But rather than having a
series of individual opportunities being worked through, it is good practice to
group them into common themes. These themes can represent a number of
different projects, or work streams, ideally centred around common AI tools.
For example, a process excellence stream might focus on improving the accuracy of processes and reducing the AHT, whilst an analytics and reporting
stream would focus on delivering compliance and improved reporting. Each
stream, using a range of different technologies, relates specifically back to a
number of benefits that are outcomes of the business strategy.
The AI Roadmap should be kept at a relatively high level, identifying the
specific project work streams and associated activities (change management,
program management, governance, etc.). For each work stream there should
then be a much more detailed project plan which identifies dependencies and
responsibilities (Fig. 6.1).
It is likely that you will want to build one or two prototypes of your most
favourable candidate opportunities, and these should be built into the roadmap, ideally with their own detailed project plans. More details on the
approach to prototyping are provided in the next chapter. The prototype
builds will test and validate assumptions but also provide important momentum for your AI program.
The key considerations when building a roadmap include: your ultimate
ambition with regard to automation (is this a game changer for your business or just something you want to try out in certain areas); whether you
want to evaluate the whole business first and then implement AI or try and
get some quick momentum by implementing in some specific areas first;
and how much change your organisation can bear at any one time (see previous section).

 

Creating Your AI Strategy

113

It is important to remember that your AI Roadmap will be an important
communication tool. It should set out clearly what you are going to do and
when, so that you can use it to explain how it will deliver your AI strategy, and
ultimately contribute to your overall Business Strategy.

Creating Your AI Strategy
Your AI Strategy is now effectively complete. You will have looked at your
Business Strategy and worked out what aspects of that could be satisfied by
the application of AI technologies. You will have determined how ambitious
you want to be with your AI program—whether it is just to be able to say
you have some AI, or to improve some processes, or to transform parts of
your business. You may have even decided to create a new business by
exploiting AI.
You will then have looked at your organisation and assessed its level of AI
maturity. Some functions or departments could still be working manually,
whilst others could have started some automation projects and others might
even have some AI projects in place already. For each of these areas you will
have worked out their AI ambition, and thus determined the size of the gap
that you have to work with.
The AI Heat Map is the tool that helps identify how you are going to bridge
that gap. Based on your knowledge of the AI Capabilities Framework you will
have identified a number of different AI opportunities in each of the areas,
assessing and prioritising their potential benefits, challenges and their alignment with the Business Strategy.
From the information in the AI Heat Map you will have developed a high-­
level business case that sets out both the hard and soft benefits that you expect
to achieve.
And finally, you will have considered the change management activities
that are necessary to ensure the success of the AI program, and built them into
an AI roadmap that sets out your medium- and long-term plans of how you
will realise the AI strategy.
It is now time to start building something.

114

6 Starting an AI Journey

The Cognitive Reasoning Vendor’s View
This is an extract from an interview with Matthew Buskell, Head of Customer
and Partner Engagement at Rainbird, a leading vendor of Cognitive Reasoning
software.
AB: Why do you think AI is being so talked about right now?
MB:	I think it’s a combination of forces coming together and driving the interest. Specifically Outsourcing, Cloud Computing, Big Data and, of course,
Hollywood.
The major driving force for outsourcing, realistically, is economic.
However, regulation has increased and margins have continued to reduce.
So, when outsourcing does not give a deep enough return we need to look
elsewhere. In this case AI is attractive because it allows us to amplify the
productivity of the staff we already have in a way that is not possible by
outsourcing alone.
Cloud computing is a big part of the reason AI has become economically viable. That and the algorithms that have been optimised and refined
over the past 30 years. To illustrate this let me tell you about Ian. In 1994
Ian was doing undergraduate studies in software engineering and AI with
me at Birmingham. He decided to code a PhD thesis on linguistics.
When he ran his program the processing power required was so great that
it ground the entire Sun Hyperspark network of servers and workstations
to a halt for the entire campus. So he was banned from using the computers outside of the hours of 10 pm–6 am. As a result, we did not see Ian for
nearly a year. Today we call his field of research Natural Language
Understanding and you can buy it for a fraction of a cent per hit from
several cloud providers.
Big Data has made AI interesting. For some types of AI like Machine
Learning to work well they love large amounts of data. Whilst it’s true
most businesses were collecting data 20 years ago, until recently they had
not put the big data platforms in place that would allow AI systems to
access them. Now that they do, AI is well placed to take advantage of it.
Hollywood might seem a weird one to add, but that fact is as humans
we have been fascinated by the idea computers could think since Ada
Lovelace described the first computer in the eighteenth century. TV programs like Channel 4’s Humans in the UK and movies like Ex-Machina
fuel that excitement. It also sells news articles and ad-words on blogs.
There is a famous, but anonymous quote that sums up Hollywood’s influence on AI: “Artificial Intelligence is the study of how to make real computers act like the ones in the movies.”

The Cognitive Reasoning Vendor’s View

115

This also poses a big challenge for the AI industry today. The fact is we
are way off simulating a human mind, so the expectation of AI is so high
it is bound to lead to some disappointments along the way.
AB: What value do your customers get out of your software?
MB:	The Rainbird software is about mimicking human decisioning, judgement and recommendations. It’s different to other AI in that its ‘people
down’ rather than ‘data up’. Meaning we take a subject matter expert and
sit them down with Rainbird so they can teach it.
When you stop for a moment and think about what you can do if you
mimic human expertise and decisioning the list is endless. We have been
involved with projects from Abbey Road studios trying to mimic the
expertise of their (rather ageing) sound engineers, through to large banks
and tax consultancies trying to reduce headcount by encoding the best
employees’ expertise into Rainbird.
On the whole, however, we have noticed two areas where we see significant value:
1. If you can take expertise and create a solution that allows customers to self-­serve you can drive some staggering efficiencies. At the
moment, this is being manifested as a ‘chatbot’. However, long
term, we believe companies will have a ‘cognitive platform’ and the
UI may or may not be a chatbot.
2. If you can take knowledge that was in a process and move it further upstream the impact to the downstream processes is dramatic.
In one instance, we moved the liability resolution in a motor claim
so that it could be handled by the call centre agent. This removed
the need for over 15 downstream processes.
AB:	What do customers or potential customers need to focus on if they are
going to get the maximum value from AI?
MB:	I have noticed a lot of companies rushing into AI proof-of-concepts without spending enough time working though the customer journey or the
business case.
There are some customers that are approaching this correctly. For
example, I was with a client the other day in Scotland that was taking a
much more Design Thinking approach to the development of a PoC.
When we started the process we all had the idea we could get up and running quickly with a simple bot that could answer FAQs. When we actually looked at the questions and followed the conversation with a real
agent it quickly became clear that the general advice would only last a few
seconds before it became specific. At that point, you needed a human in

116

6 Starting an AI Journey

the loop, so the result of putting AI in place is that you would have paid
twice—once for the virtual agent and secondly for the human. This is a
Bad Idea. We were actually able to find the value by going deeper with the
virtual agent and make it capable of completing a full interaction with a
smaller number of questions.
AB: Do you think the current hype around AI is sustainable?
MB:	Personally, I don’t think the hype is helpful, so I welcome it slowing down
or becoming more grounded. However, I think unfortunately it will continue, mainly because we need AI to work. Without it we have to face
some pretty stark economic realities.
AB: How do you see the market developing over the next 5 years or so?
MB:	The market is very fragmented at the moment and the use cases for the
technology are very broad. So, I think it will do what software innovation
has always done: all these companies will consolidate so they can offer an
end-to-end solution and the use cases that work well will survive and the
rest will die out.
The final thought I would leave you with is from John McCarthy, who
is considered the father of AI: “As soon as it works it’s no longer called AI
anymore.”

Reference
Willcocks LP, Lacity MC (2016) Service automation: robots and the future of work.
Steve Brookes Publishing Warwickshire, UK.

7
AI Prototyping

Introduction
Creating your first AI build, however small, is a key milestone for any AI program. After all of the hard work developing the AI Strategy, this is the point
at which things really start happening, and people in your organisation can
actually see tangible outputs. Getting this stage right, therefore, is vitally
important.
There are a number of different approaches to this, which I cover in some
detail in the ‘Creating Your First Build’ section of this chapter, and they each
have different names and acronyms. For simplicity, I have used the word ‘prototype’ as the general term for all of these.
Prototyping can be done once the AI Strategy is complete, or can be done
to validate some of the assumptions during the development of the strategy.
This may be influenced by how important those assumptions are (if they are
fundamental to the success of the program, then test them early) and how
much stakeholder buy-in and momentum you need (a successful prototype is
a great way to engage with stakeholders).
Some of the decisions you make at the prototyping stage may have an
impact on the rest of the program. The most important of these is your technology strategy—do you build your AI solutions yourself, buy them from a
vendor or build them on top of an established AI platform, or use a mixture
of all of these? It is this question which I will address first.

© The Author(s) 2018
A. Burgess, The Executive Guide to Artificial Intelligence,
https://doi.org/10.1007/978-3-319-63820-1_7

117

118

7 AI Prototyping

Build Versus Buy Versus Platform
Some of the business case costs that were discussed in the previous chapter
will have required some assumptions on how you intend to create your AI
solutions. Generally, there are three main approaches to take: off-the-shelf AI
software, an AI platform or a bespoke AI build. Your final solution may
involve any of these approaches, but the core functionality is likely to initially
focus on one of these:
• Off-the-shelf AI Software is generally the simplest approach, since most of
the hard work in designing the system has already been done by the vendor.
As I’ve described in the earlier chapters, there is a plethora of AI vendors,
each of which can provide a specific capability that can be used as a stand-­
alone application or as part of a wider solution.
As well as all the design, testing, debugging and so on having already
been done, implementing off-the-shelf software means that the product
will be supported by the vendor, and they are likely to have either their own
implementation resources, or partners trained up to support implementation. You will still need to exert effort to identify, clean and make available
the required data, as well as providing subject matter expertise regarding
your own processes (so ‘off-the-shelf ’ is a bit of a misnomer), but generally
this is the ‘path of least resistance’ for implementing an AI capability.
The biggest disadvantage of using a software ‘package’ is that its capabilities may not align well enough with your objectives and required functionality. You may want to implement, for example, a system to analyse the
sentiment of your customers’ feedback. One vendor may provide a system
that analyses the data exactly as you would have expected but doesn’t provide any useful reporting tools. Another may be less effective at the analysis
but has a best-in-class reporting suite.
• There are also a number of risks in relying on a software vendor to provide
your AI capability. Because of the excessive hype around AI right now
(remember Chap. 1?) there are quite a few vendors that are little more than
an idea and a website. The challenge is that many of the AI software vendors are start-ups or young companies, so it can be a difficult task to identify those with viable and stable products. Your due diligence should be
extremely thorough so that it filters out those who do not yet have enough
experience or capability. Using AI experts at this selection stage can save
considerable pain later on in the project. (You may, though, wish to deliberately go with a software vendor that does not yet have a fully ­commercialised
system if their technology could potentially provide you with a distinct

Build Versus Buy Versus Platform

119

competitive advantage, or where you want to help them shape their offering, so that it delivers some specific advantage to you, whether that be
functionality, exclusivity or price.)
The commercial models for AI software can be quite varied. The most
common, and simplest, is an annual subscription that includes the necessary licenses as well as the support and maintenance of the software. It will
usually include hosting of the software by the vendor (or an agent of the
vendor) if it is a software-as-a-service model. These AI software packages
are fully formed, with a user interface for training and using the product.
For other types of AI software where the software is accessed through an
API (e.g. to determine the sentiment of a particular passage of text) a ‘payas-­you-go’ model is used—the user pays per API call made. There are usually bands of pricing depending on the volume of calls made.
• AI Platforms are the middle ground between an off-the-shelf-package and
a bespoke build. They are provided by the large tech companies such as
IBM, Google, Microsoft and Amazon, some large outsourcing providers,
such as Infosys and Wipro, as well as specific platform vendors such as
H20, Dataiku and RapidMiner.
In the section on ‘Cloud Computing’ in Chap. 4 I described Amazon’s
‘stack’ of AI services, which is typical of a platform offering. It provides
ready-made, ready-trained algorithms for those with simple and well-­
defined requirements (such as generic text-to-speech translation); untrained
but ready-made algorithms for those that need to add their own specific
practices, nuances and data into the model; and a set of AI tools for
researchers and data scientists to build their algorithms from the ground
up. (The platform provider can also usually provide the infrastructure facilities as well.) Of course, an organisation may want to use a range of these
services for the different capabilities that their solution requires.
The platform approach may make sense if your organisation already has
a relationship with one of the tech giants, especially if there is already a
strong connection with cloud services. You may find, though, that you
have to make some compromises on certain aspects where a platform provider’s specific AI capability may not be as strong as a stand-alone one. Of
course, you can always mix and match between the two types, depending
on how rigid your partnership model is.
Outsourcing providers also provide AI platforms for their clients—for
example Infosys have Nia and Wipro have Holmes. They usually combine
a data platform (e.g. for data analytics), a knowledge platform (to represent
and process knowledge) and an automation platform which brings these
together with RPA. If your outsourcing provider does have a robust AI

120

7 AI Prototyping

platform then it is certainly worth looking at it—you will probably have to
use their resources for the implementation, but it does provide a relatively
simple route assuming that their capabilities match your requirements.
As well as the platforms provided by the tech giants and outsourcing
providers, there are also development platforms available from AI-only
vendors, such as H20, Dataiku and RapidMiner (these are sometimes
referred to as Data Science Platforms). These provide a solid foundation for
developing AI applications with your own resources or using third parties.
They consist of a number of different tools and application building blocks
that can be integrated with each other whilst providing a consistent look
and feel. This approach will give you the greatest flexibility and control out
of all the platform models—the tools are generally more advanced than
those provided by the bigger platforms. You will need to take a more handson approach to the design and implementation though—the main users of
these systems will be the data scientists.
The pricing models vary between the different providers. For commercial developments, some offer a revenue share model where they will charge
an ongoing percentage (usually around 30%) of any subsequent revenue
generated from the application built on their platform. The most common
approach is to charge per API call—this means that every time the application that you have developed uses a specific piece of functionality of the
platform (e.g. converting a piece of text to speech) there is a small charge.
Commonly there is a threshold below which the API calls are free or at a
fixed price.
In the platforms space, there is some overlap between each of the categories I have defined—IBM, for example, could easily fit into all three—but
this does provide a useful framework for selecting which one (or ones) may
be most appropriate for your AI strategy.
• Building bespoke AI applications will provide you with the greatest level
of flexibility and control; however, it is unlikely that you will want to take
this approach across all the opportunities that you have identified. Bespoke
AI development, just as for any bespoke IT development, can be great for
providing you with exactly what you need but can create related issues
around change management and support. Therefore, for the majority of
enterprises, bespoke AI development should be used only when absolutely
necessary, that is, for complex, very large data problems, or when creating
a completely new product or service that requires technological competitive advantage. (It may make sense to use bespoke development for creating
the initial builds—see next section for details—but subsequently moving
onto a vendor or platform approach.)

Creating Your First Build

121

A bespoke build will also require effort to be put into designing and
building the user interface for the AI system. This can be good in that you
can develop something that suits your users perfectly, but can be a challenge to get it just right. Many excellent and clever AI systems have failed
because the user interface simply wasn’t good enough.
For a bespoke development, you will need your own highly capable data
scientists and developers or (more likely) to bring in specialist AI consultancies that already have these resources. The commercial model is usually
based around Time & Materials, although Fixed Price and Risk/Reward
can also be used when the requirements are well defined.
You may want to use a combination of all the above approaches. If your
AI strategy is ambitious and will impact many areas of the business, then
you will probably want to consider the AI platforms as your prime approach,
making use of off-the-shelf packages and bespoke build where appropriate.
If you only want to target one or two AI opportunities, then it may make
more sense to consider just the off-the-shelf solutions. A strategic program
to create a brand-new product or service may require bespoke development
to ensure that it delivers enough competitive advantage.
So, now that you have an idea of the development approach, or approaches,
that you are going to use, it is time to actually build something.

Creating Your First Build
At this point in your AI journey you will have a list of prioritised opportunities, each with an estimate of the value that it could deliver, as well as an idea
of how they could be implemented (build, buy or platform-based). For the
next stage, we need to test some of our assumptions and build some momentum for the program, which means we have to start building something.
The scale of this initial build stage will depend very much on your intent.
There are five generally accepted approaches to this, each of which focuses on
satisfying slightly different objectives. You won’t need to understand the
detailed intricacies of each of these approaches (your internal development
team or external partner will have the necessary knowledge) but it is important to know which is the most appropriate approach for your project so that
you don’t waste valuable resources and time.
There are some overlaps between the different approaches, and you could
legitimately choose specific aspects from each of them. You may also want to
use a number of these approaches in turn, as they each focus on different
objectives to be achieved.

122

7 AI Prototyping

• Proof of Concept (PoC): a PoC is a software build that will be discarded
once it has been able to demonstrate that the concept that is being tested
has been proven. It can be of any scale, but generally focuses on the barest
functionality whilst ignoring any complexities such as security or process
exceptions (this is called ‘keeping to the happy path’). PoCs will not go into
live production, and will use non-live ‘dummy data’. PoCs are usually carried out to test the core assumption (e.g. we can automate this process
using this data set) but are also useful to manage stakeholder expectations—seeing a system in the flesh early on often provides an extra boost of
momentum for the project as a whole.
• Prototyping: this is a broad term that covers the building of specific capability in order to test the viability of that capability. Prototypes may focus
on the user interface (usually called horizontal prototypes) or specific functionality or system requirements (vertical prototypes). One project may
have a number of different prototypes, each testing for a different thing. It
is rare for a prototype to resemble the final product. As with PoCs, prototypes are generally discarded once they have done their job, but some can
be incorporated into the final product (this is called ‘evolutionary prototyping’)—it is important to know at the start whether the prototype will be
retained as it will need to be built more robustly than a throwaway one.
• Minimum Viable Product (MVP): the proper definition of an MVP is
one where the build, which contains the minimum level of functionality
for it to be viable, is released to its customers or users. The feedback from
these early adopters is then used to shape and enhance the subsequent
builds. The term MVP has, though, become a catch-all description for any
sort of early build and therefore tends to lose focus as to what it might be
looking to achieve. For a build that is not intended to be released, then a
RAT (see below) is generally the better approach. For a build that will go
live, then a Pilot (again, below) is usually preferred. MVPs can have their
place in the development cycle, but only once their objectives have been
well defined.
• Riskiest Assumption Test (RAT): RATs are a relatively new approach that
focus on testing only the riskiest assumption that requires validation. This
is in contrast to the MVP which is less specific about what it is trying to
achieve. Also unlike MVPs, RATs are not intended to go into production.
A RAT will seek to build the smallest experiment possible to test the riskiest assumption that you have. As with prototypes (RATs are really a specific
type of prototype) a number of different tests are carried out to validate
different assumptions—once the riskiest assumption has been tested and
validated, then the next RAT will focus on the next riskiest assumption and

Understanding Data Training

123

so on. Understanding what the riskiest assumption actually is requires
some detailed thought—you will need to identify all of the assumptions
that you have made so far that need to be true for the AI opportunity to
exist. Ask yourself if these assumptions are based on other assumptions,
then try to identify the root assumptions and then find the riskiest one of
those. Then it is a case of working out how you can test that assumption
with the minimum amount of code. RATs are not the easiest approach to
take but can deliver the best long-term value for your project.
• Pilot: a Pilot, like an MVP, will be used in a production environment.
Unlike an MVP, it should be fully formed and complete. It is considered as
a test because it focuses on automating a process or activity that is relatively
small and low risk to the organisation. Pilots take longer to build and are
more expensive than any of the options above, but have the advantage of
not ‘wasting’ time and effort (although focused prototyping is rarely a
waste of time). The Pilot needs to achieve all of things that are expected of
a PoC and MVP as well as being fully functional; therefore, the candidate
process needs to be chosen well—if it is too complex (e.g. it involves lots of
systems or has many exceptions) then it will take too long (and cost too
much money) to build, and crucial momentum will be lost. A successful
Pilot will provide a huge confidence boost to the project as well as deliver
tangible value.
As with the wider automation strategy, when selecting the most appropriate approach for that initial build always bear in mind what your end goal
is—ask yourself whether you need to build any sort of prototype at all (the
answer is probably yes) but also which approach is going to provide you with
the highest chances of success. Speed of implementation and cost will also be
big factors in the choice of your approach. As a default answer, the RAT is
likely to be the best approach for the majority of cases, but do consider all
possibilities.

Understanding Data Training
One aspect where AI projects are generally trickier than ‘normal’ IT projects
is with the dependency on data, and this challenge is particularly acute during
the prototyping stage.
Although it is possible to cut down on the functionality and usability of the
system being built in order to test assumptions, it is rarely possible to do the
same with the training data. Of course, not all AI systems rely on large data

124

7

AI Prototyping

sets—cognitive reasoning engines and expert systems require knowledge
rather than data, but for machine learning applications in particular, data
availability is paramount.
There will be a minimum amount of data that will produce a viable outcome, and this can sometimes mean that the prototype will have to rely on
nearly all the data available, rather than a small sub-set. If your complete data
set contained, say, 10 million records, then a sample size of 10% may still be
workable, but if your complete data set was, say, 100,000 records then
(depending on what you intend to do with it, of course) a sub-set may not
give you any meaningful results.
The more complex the model you want to create (i.e. one with many different parameters), the more training data you will need. If the data is insufficient, your AI model could suffer from what data scientists call
over-fitting—this is where the algorithm models the fit of the data too well
and doesn’t generalise enough (it ‘memorises’ the training data rather than
‘learning’ to generalise from the trend). To understand whether your model is
exhibiting over-fitting and correct for it there are various statistical methods,
such as regularisation and cross-validation, that can be exploited, but usually
at the expense of efficacy.
A good portion of your training data must also be used for testing. This
validation set should, like the training set, be ‘known’; that is, you should
know what output you are expecting from it (‘this is a picture of a cat’) but
you don’t reveal this to the algorithm. Because AI outputs are probabilistic the
answers won’t always be correct, so you need to understand and agree what
level of confidence you are comfortable with. There is no hard and fast rule
about how much of your training data should be retained to do this testing
but 30% is a good starting point. Factors such as the size of the complete set,
the richness of the data and risks of getting the wrong answer will impact this
figure.
You may not be able to get around the fact that, in order to train and test
your prototype system effectively, you need to use nearly all of the data you
have available. If data availability is going to be a challenge then one or more
RATs, as described in the previous section, may be the best way forward at this
point.
Once you have a good understanding of how much data you will need for
the prototypes, you will need to acquire the data, (probably) clean it and
(maybe) tag it. I can’t stress enough how important the quality of the data is
to the value that the AI system will deliver.
The most likely scenario is that the data will be sourced from within your
own organisation—it may be customer records, transaction logs, image libraries

Understanding Data Training

125

and so on. You may also want to acquire data from external sources such as the
open-source libraries I discussed in the section on big data in Chap. 2. This
data will come ready-tagged, but there may also be publicly available data sets
(such as traffic movements) that will require tagging before they become useful
for training your algorithm.
The cleanliness, or accuracy, of the data will clearly have an impact on the
efficacy of your AI system. The old adage of ‘garbage in, garbage out’ is particularly relevant here. Common sense and system testing will determine
whether your data is of sufficient quality to provide useful results.
For training data that needs tagging, such as images or video, the most
efficient way is through crowd sourcing (see Chap. 4 for how crowd sourcing
works). Crowd sourcing firms, such as CrowdFlower, have many human
agents at hand to complete micro-tasks—each are sent items of the data and
simply tag them with the required information (this is a picture of a cat, this
is a picture of a dog, etc.). The fully tagged data set is then returned to you so
that you can train the model.
One risk with crowd sourcing is that any biases in the humans doing the
tagging will eventually be reflected in the AI model. As a rather silly example,
if the human agents doing the tagging were all very short, and they were asked
to tag pictures of people as to whether they were tall or not, then it may be
that a higher proportion than expected would be tagged as tall. That ‘smallness bias’ would be carried through into the model and skew the results of the
new, unseen data. Most good crowd sourcing firms though will ensure that
bias doesn’t enter the data in the first place, but can also correct for any bias
that may exist.
Another, rather innovative, approach to data acquisition and training is to
use virtual worlds or computer games instead of ‘real life’. I mentioned previously that Microsoft have created a specific version of Minecraft which can be
used to train AI agents, and some software companies are using games like
Grand Theft Auto to help train their image recognition software that is used
in driverless cars. So, rather than having to drive around many different environments for hours and hours filming and then tagging objects, all of it was
done in a fraction of the time in the virtual world of the computer game.
So, with your AI Strategy complete, and your first build underway (whichever way you have decided to approach it), it is time to think about the longer
term, and how you can start to industrialise your new-found AI capability.
But not before getting a good understanding of all the risks that AI can raise,
and how you might mitigate these, and that is the subject of the next
chapter.

126

7

AI Prototyping

The AI Consultant’s View
This is an extract from an interview with Gerard Frith, co-founder of Matter
and Sensai.ai, AI Strategy and Prototyping consultancies.
AB: What got you into the world of AI in the first place?
GF:	I have been fascinated with the mind since I was small child. I read Jung
and Freud in my early teens, Godel Escher Bach when I was 18. I initially
studied Psychology at university, but in my first year it was my Cognitive
Science module that grabbed my attention most. I became obsessed! At
the end of my first year, I decided to start over, and signed up for an AI
degree instead. That was 1991.
Once my degree was done, I found out that the rest of the world was
rather less interested in AI than I was. During the AI winter of the 1990s, I
mostly had to pretend my degree was just an exotic computer science degree.
In 2013, the rest of the world was finally catching on, and I established
Matter AI, one of the first AI consultancies anywhere. I sold Matter in
late 2016.
My major focus at the moment is on the use of AI to create better and
deeper customer relationships.
AB: Why do you think AI is being so talked about right now?
GF:	There are two key reasons: firstly, AI tech has recently become practically
useful. That’s mostly due to the vast amounts of data that the digital age
has exponentially spewed, and to the massive advances in available processing power. The theoretical advances in AI that underpin much of
what is being used in business today were developed years, and sometimes decades, ago.
The second reason is that AI has always been exciting, scary, troubling
and challenging in a way that no other technology is or could be. If you take
away the breathless discussions of how AI will take all our jobs, then you are
left with a pretty appropriate level of coverage of a disruptive technology.
The creation of genuine AI could easily claim to be the most significant scientific achievement in history. It has the potential too to be one of
the most significant social and spiritual achievements.
AB: What value do your customers get out of your services?
GF: My customers say they get three main things from working with us:
• Expert guidance. I’ve been a developer, a CEO, and a management
consultant, so I know the technology, I know the market and I
understand what a company needs to be successful.
• I’m a disruptor. I like to challenge the current equilibrium and bring
new thinking to situations that allows organisations to reinvent
their value propositions and the way they deliver them to market.

The AI Consultant’s View

127

• Because of my breadth of experience, I can provide the glue that
brings together tech and business strategy, and then how to execute
those strategies in effective ways.
AB:	What do customers or potential customers need to focus on if they are
going to get the maximum value from AI?
GF:	The same things they should always focus on when assessing new technology—how can I create greater value for my customers using this technology? What competitive advantage could this give me? The critical
thing to do is to move fast. It’s getting easier for disruptors to change
industry dynamics quickly, and with less capital. But do treat AI as an
early stage technology too. Invest seed capital in experiments, but get
those experiments live quickly and then iterate.
AB: Do you think the current hype is sustainable?
GF:	It’s inherent in human systems that hype is never sustainable. Our biology seeks novelty and in doing so makes cycles of hype and deflation
inevitable. That said, I do believe that AI is worthy of tags like the ‘4th
Industrial Revolution’ in ways that, say, the 2nd and 3rd ones weren’t. In
this sense it’s just a continuation of the impact of technology development in general, but its impact will be quicker, but only in the same way
that the internet’s impact was quicker than that of TV, and the impact of
roads occurred faster than the impact of the printing press.
Beyond its vast (but normal) economic impact though, I think it will
actually change everything for humans. It is not like any other technology
in terms of how it affects how humans will think about themselves.
Currently even professors of neuroscience who argue that the mind is
nothing greater than a complicated computer hold on privately to the
idea that they have free will and a ‘soul’. AI will challenge those ideas in
quite new and scary ways.
AB: How do you see the market developing over the next 5 years or so?
GF:	The next five years will be very exploratory. AI is a general-purpose technology and will thrust out in multiple directions at once. We are seeing
new products appear weekly covering a plethora of use cases, with even
the biggest tech players only covering relatively small areas. Companies
will need to look at gluing together a range of best of breed providers,
often with a strong vertical industry focus.
We’ll undoubtedly see a lot of acquisitions, but only slow consolidation as tech companies focus on land grab. We’ll also begin to see some
industries such as Law begin to be taken over by software. In general,
industrial boundaries will become even more porous as data drives the
discovery of new horizontal processes that can cross the traditional
verticals.

8
What Could Possibly Go Wrong?

Introduction
It is important to provide a very balanced view on the benefits and risks of
AI. Throughout the book, I’ve talked about some of the unique challenges
that need to be faced when implementing AI, but it is worth drawing these
out into a separate chapter, as I have done here, to make sure they get the
attention they need.
In the chapter I’ve tried to cover the ‘localised’ challenges—those that any
company will have to consider when implementing AI—but also the more
general ones that will affect the way that we work and how we will live our lives.

The Challenge of Poor Data
Data quality is traditionally measured by the ‘4 Cs’: Clean, Correct, Consistent
and Complete. But in the world of AI and Big Data we need different rules to
guide us. Data quality can be thought of as a combination of data veracity (its
accuracy) and data fidelity (its quality in the context of how it is being used).
Data veracity is less important in the world of Big Data because any small
errors will be drowned out by the sheer volume of correct data. The model
that is created from the algorithm will look at overall trends and fitness, so a
few outliers will not significantly impact the outcome. (If there are fundamental errors in the data, such as the decimal point being one place out on every
data point, then, of course, it will have a material impact, but in this section
I am only considering small issues in data quality.)
© The Author(s) 2018
A. Burgess, The Executive Guide to Artificial Intelligence,
https://doi.org/10.1007/978-3-319-63820-1_8

129

130

8 What Could Possibly Go Wrong?

With data veracity, whether the data is being used for AI or not, there is
always a balance to be made between accuracy and cost. The trick is to understand what level of accuracy is ‘good enough’. With AI applications, that
threshold is generally lower than for more traditional computing applications
because of the generalised modelling effect.
Data fidelity, on the other hand, can have a serious impact on the results.
This is where the data is inappropriate for the task for which the AI has been
set. In Chap. 5 I wrote about one example of AI being used to predict recidivism of people who have been charged with criminal offences by the Durham
Police Force. I noted that the data used for the predictions was sourced only
from the region itself, and therefore excluded any relevant activity the subject
might have taken part in outside of Durham County. So, a career criminal
from Leeds (just down the A1 main road from Durham) who had just committed his first offence in Durham may have been allowed to walk out of the
police station because he was considered a low risk. In this case, the data that
the AI system is using is correct but not wholly appropriate.
Another, slightly different, reason that the source data would not be appropriate is if it contained unintended bias. This is a specific problem with AI
systems and is covered in detail in a later section in this chapter.
Related to data quality is the question of whether more data always means
better results. Generally, with machine learning applications this is true; however, there are some subtleties to consider. There are two, almost conflicting,
perspectives.
The first, called ‘high variance’, is where the model is too complex for the
amount of data, for example where there are many features compared to the
size of the data. This can lead to what is called over-fitting and would cause
spurious results to be returned. In this case, the data scientist could reduce the
number of features, but the best solution, and the one that seems most like
common sense, is to give the model more data.
But for the opposite case, where the model is too simple to provide meaningful results, called ‘high bias’, adding more data to the training set will not
deliver any improvement to the outcome. In this case adding more features,
or making the available data a higher quality, both in veracity and in fidelity
(such as data cleansing and outlier removal) will be the most effective
approaches to improving the outcome.
Generally, it is the role of data scientists to ensure that the data is big
enough, relevant and appropriate for the problem at hand. Although my
police example above is relatively simple, some aspects of data science can
seem more like an art form than a science. In most cases there is no clear
‘right’ and ‘wrong’ answer—the data scientist must use creativity and ­inference

The Challenge of Poor Data

131

to identify the best sources and mix of data. This particular skill set is one of
the reasons that the best data scientists can demand such high fees at the
moment.
Once poor data quality has been identified as an issue then there are a number of approaches to try and fix (or enhance) it. Some of these could be the
traditional methods of data cleansing; crunching through the data to identify
anomalies to constraints and then correcting them using a workflow system.
Crowd Sourcing, which I described in Chap. 4, can also be used as a cost-­
efficient way to cleanse or enhance data sets, particularly with regard to ensuring labels are correct in training data sets. Tasks, usually split down into small
units, are assigned to a distributed team of humans, who will process each one
in turn.
RPA can also be a useful tool to help cleanse data where large enterprise
workflow systems are not available. The robots will need to be configured to
identify all the constraints they need to look for. Once they have found the
data errors they can then validate those with other systems to determine what
the correct answer should be, and then make that correction. RPA can be a
good solution for ‘one off’ cleansing exercises because it is relatively easy to
configure. It can, of course, be used in combination with crowd sourcing.
And, with only a slight element of irony, AI systems themselves can be used
to cleanse data for AI systems. Some vendors now have solutions which can
carry out Field Standardisation (e.g. making two different data sets consistent), Ontology Mapping (e.g. extracting product characteristics),
De-duplication (e.g. removing entries with the same content) and Content
Consistency (e.g. identify and matching abbreviation forms). Generally, they
work by using machine learning to analyse the structure of the data model to
then determine the kind of errors such a model is likely to generate.
Another area where AI could be used to improve the quality of data, and
even generate new data, is in the use of Generative Adversarial Networks
(GANs). This is where one AI system judges the outputs from another system.
For example, the originating AI could generate a picture of what it thinks of
as a cat. The second system then tries to work out what the picture is of. The
assessment, i.e. how close is this to looking like a picture of a cat, is then used
to guide the first system to try again and create a better picture. GANs are
relatively new, and this approach is still at the experimental stage, so I will save
a detailed description of how this works to the final chapter.
As should be clear from much of this book, AI (in most cases) feeds off
data—without the data then there is little value that AI can add. But, as we’ve
seen in this section, more data does not always mean better results (especially
in high bias model scenarios). And it’s not just about volume of data but the

132

8 What Could Possibly Go Wrong?

quality. The veracity and fidelity of the data can have a huge impact on the
performance and effectiveness of the outcomes. Any data-dependent AI project must therefore involve careful planning, ensuring the data is the appropriate size and complexity for the problem that is being addressed.

Understanding the Lack of Transparency
The reason Machine Learning is called Machine Learning is, rather obviously,
that it is the machine, or computer, that does the learning. The computer does
all the hard work in ‘programming’ the model that is going to make the predictions. All it needs is the data to be trained on: feed it the data and it will
come up with the model. The human (the data scientist) has to choose the
right algorithms in the first place and make sure the data is appropriate and
clean (see previous section), but the model that is created is essentially the
work of the machine.
I labour this point somewhat because it is an important, inherent aspect of
machine learning that is responsible for one of its biggest drawbacks—the
lack of transparency. I can ask a trained AI system to, for example, approve a
credit lending decision, or to recommend whether a candidate gets short-­
listed for a job, but I can’t easily ask it how it came to that decision. The model
has looked at the many different features of that customer or that candidate
and, based on the many different features of the training set, has determined
the answer to be, in these examples, yes or no. But which features were influential in that decision, and which ones were irrelevant? We will never really
know. It’s a bit like trying to recreate an egg from an omelette.
Which, of course, can be a problem if you need to explain to that candidate
why they weren’t short-listed, or why you didn’t give that customer the loan.
And if you are in a regulated industry, then you will be expected to be able to
provide those answers as a matter of course.
There are three general approaches to tackling AI opaqueness. The first one
is to not use machine learning in the first place. In Chap. 3, in the section
concerning the optimisation capability, I described Expert Systems and
Cognitive Reasoning Systems, which are particular flavours of AI that work
on the principle of a human subject matter expert creating the original model
rather than it being created using data. (For this reason, some AI experts claim
that expert systems should no longer be described as AI.) The ‘knowledge
map’ that is created from these approaches can be interrogated, by a person or
via a chatbot, to access the information. In some of the more advanced
­systems, the queries can be automatically routed through the map using the
specified weightings that have been applied to each node and connector.

Understanding the Lack of Transparency

133

This all means that the decision that comes out at the end can be traced
back through the map, making it fully auditable. So, using one of my examples above, the failed credit application could be traced to, say, 75% of the
decision due to salary, 40% due to postcode, 23% due to age and so on. And,
once it has been designed, assuming the processes don’t change, it will remain
consistent in its decisions.
The challenge in using expert systems is that they can get very complex very
quickly. The more features that the system has, the more complicated the
model becomes to define. If the processes do change, the model then has to
be changed as well to reflect them, and that can be a laborious task.
Another approach for simple problems is to use self-learning decision trees,
usually called Classification and Regression Trees. From a set of input data,
these CART algorithms create a decision tree that can then be interrogated to
understand which features have been the most influential. CART alogorithms
generally provide the best balance between effectiveness and transparency.
For complex systems with lots of data and many features, an ‘ensemble’
approach that calls on many different algorithms to find the most popular
answer (Random Forest is the most common ensemble approach), will prove
the better choice, but it will have that challenge around transparency. The
most common approach to tackling this is to try and reverse-engineer the
decision by changing just one variable at a time. Returning to my customer
who has had his loan rejected, we could repeat his case but changing each of
the features that he has fed into the system (his salary, his postcode, his age,
etc.). This trial-and-error approach will then be able to indicate (but not necessarily isolate) which of the features had the most influence on the decision.
The problem with this approach is that, with many features, this can be a
long process. Carrying out the analysis on a number of test cases and publishing these first is a good way to avoid analysing every rejected case. The ‘model
cases’ could demonstrate, for example, how, with everything else being equal,
age impacts the decision-making process. But you might have to do that for lots
of different model cases since the influence of age might be different in different
postcodes or for different salary levels. And, if the system self-learns as it goes
along, then the test model will need to be refreshed at regular intervals.
The level of testing required will be very dependent on the type of problem
that is being solved. The more sensitive the problem (medical imaging, for
example), the more robust the testing will need to be.
A disgruntled customer who has just been rejected for that loan may, then,
be satisfied with this demonstration of the ‘logic’ used to make that decision.
But would that satisfy the industry regulators? Again, each use case will have
different requirements for testing, but it could be the case that being able to
demonstrate a fully tested system, and the influences that different features

134

8 What Could Possibly Go Wrong?

have, will be enough for some regulators. In the legal sector where e-­Discovery
software is used to automatically process thousands or even millions of documents to identify those that are relevant or privileged in a litigation case, the law
firm using the software must be able to demonstrate a robust use of the solution, with due process applied at every stage. Once the court is satisfied that this
is the case, then the results from the analysis can be used as valid evidence.
Artificial Intelligence is a new and fast developing technology—regulators
are generally slow and considered in their approach, and will need time to catch
up. Putting the risk of fraud to one side, once AI is more generally accepted in
society it may be the case that customers, and therefore regulators, take a more
relaxed approach to the transparency question. Until then, AI-adopting companies need to ensure that they have the answer when the question is raised.

The Challenge of Unintended Bias
A common fallacy about AI systems is that they are inherently unbiased: being
machines they will surely lack the emotional influences that inflict humans in
our decision making that lead to bias, whether intended or not. But that is
just a fallacy. The core of the problem is that the data used to train the AI may
have those biases already baked in. If you use biased data to train an AI, then
that AI will reflect back those same biases. And, because the model can be
opaque to interrogation (as we saw in the previous section), it can be difficult
to spot them.
Here’s a simplified example to demonstrate the problem. In the world of
recruitment, AI can be used to short-list candidates based on their CV (curriculum vitae, or résumé). The AI system would be trained by feeding in many
CVs, each one labelled as either ‘successful’ or ‘unsuccessful’ based on whether
they got the job they were applying for at the time. The AI can then build up
a generalised picture of what a successful CV looks like, which it can then use
to short-list any new candidate CV that is submitted—a close enough match
between the candidate CV and the ‘successful’ model would mean the CV
gets through to the next round.
But how did those candidates in the training set get their jobs? By being
assessed by humans, with all our conscious and subconscious biases. If the
human recruiters had a propensity to reject older candidates (even if they
didn’t realise they were doing it), then that bias would flow through into the
AI model.
So, how to ensure there is no bias in the training set? The first answer is to
try and use as wide a sample as possible. In my example above, you would try
to take training CVs from many different recruiters, ideally those with as

The Challenge of Unintended Bias

135

neutral a mix of gender, race, age and so on as possible (just as polling firms
try to use a representative sample of the population).
This may not always be possible, especially if the data is from your own
company and is all you have. But even publicly available data sets can be
prone to bias. Some of the data sets used for facial recognition training do not
have enough representative samples from people of colour. This means that
any AI trained using that data could struggle to identify non-white faces. As
these biases in public data sets are unearthed, they are being slowly fixed (there
is even an Algorithmic Justice League which highlights bad practices like
these), but AI developers should be particularly cognisant of the issues with
any public data that they use.
In any of the above cases where there could be bias in the data, it will be
necessary to try and test for it. Just as in the previous section, where we looked
at testing for different influences in decision making, we can similarly test for
the influence on any particular feature for bias, as long as we know what the
‘right’, that is, unbiased, answer should be.
In the recruitment example, we would expect an unbiased response to show
that age does not influence the short-listing decision. If we are able to isolate
the age feature in our data set then we can test, by varying only the age in our
model case, whether it is influencing the outcome at all. As with the transparency test, this would need to be done across a range of samples (e.g. just men,
just women) to ensure that age was not influential for any particular group.
Assuming some level of bias has been detected, the next question is what to
do about it. Finding the type of bias (e.g. age, gender) will immediately help
identify any groups of data in the training set that may be causing it (there
may have been a sub-set of the data that was from a particularly young set of
recruiters, for example). This input data can then be excluded or altered to
eliminate the source of that bias.
The model can also be tweaked by changing the relative weightings between
different terms. If the training CVs in my example showed a strong correlation between Manager (role) and Male (gender), the mathematical relationship between gender-neutral words like ‘Manager’ and gendered words such
as ‘Male’ can be adjusted. This process, called de-biasing, is not insubstantial,
and usually involves humans to identify the appropriate and inappropriate
words, but it can be done.
One important thing to bear in mind is that not all bias is necessarily bad.
There may be instances where the inherent bias is important and where the
consequences justify the means, such as in identifying fraudsters. There is
clearly a fine balance between representing the truth, biases and all, and altering data to represent the accepted social position. Also, on a much simpler

136

8 What Could Possibly Go Wrong?

level, the AI will need to understand specific definitions in areas that could be
biased, such as in the difference between a king and a queen, if it is to work
properly.
Eliminating unintended bias from AI is, largely, still a work in progress and
must be considered carefully where there are specific consequences. Using
publicly available data sets does not, currently, ensure neutrality. Using data
from your own organisation will make de-biasing even more important, and
therefore must be built into the AI development roadmap.

Understanding AI’s Naivety
It was the French philosopher Voltaire who famously said that you should
judge a person by the questions they ask rather than the answers they give.
These may be very wise words when it refers to humans, but for AI the situation is even simpler: the machine doesn’t need to know what the question is
in the first place in order to give a respectable answer.
This is the whole idea behind Clustering, which I described in Chap. 3. You
can present a large volume of data to a suitable algorithm and it will find clusters of similar data points. These clusters may depend on many different features, not just a couple of things like salary and propensity to buy quinoa, but,
in some cases, many hundreds. The AI is providing mathematical muscle
beyond the capability of a human brain to find these clusters.
But these clusters are not (or, more accurately, don’t have to be) based on
any pre-determined ideas or questions. The algorithm will just treat the information as lots of numbers to crunch, without a care whether it represents data
about cars, houses, animal or people. But, whilst this naivety of the data is one
of AI’s strengths, it can also be considered a flaw.
For big data clustering solutions, the algorithm may find patterns in data
that correlate but are not causal. In the section on Clustering in Chap. 3 I
gave the rather whimsical example of an AI system finding a correlation
between eye colour and propensity to buy yoghurt. It takes a human to work
out that this is very unlikely to be a meaningful correlation, but the machine
would be naive to that level of insight.
The AI may also find patterns that do not align with social norms or expectations—these usually centre around issues such as race and gender. I’ve
already written in the previous section on the challenges of unintended bias,
but in this case an awkward correlation of purely factual data may naively be
exposed by the algorithm. The challenge for those responsible for that algorithm is whether this is a coincidence or there is actually a causality that has

Understanding AI’s Naivety

137

to be faced up to. How that is handled will have to be judged on a case-by-case
basis, and with plenty of sensitivity.
There is also the infamous example of the Microsoft tweetbot (automated
Twitter account) that turned into a porn-loving racist. It was originally
intended that Tay, as they called the bot, would act through tweets as a ‘carefree teenager’ learning how to behave through interactions with other Twitter
users. But it quickly turned nasty as the human users fed it racist and pornographic lines which it then learned from, and duly repeated back to other
users. Tay, as a naive AI, simply assumed that this was ‘normal’ behaviour. It
only took a few hours of interaction before Microsoft were forced to take the
embarrassing tweetbot offline.
One useful way of thinking about the naivety of AI is to consider how dogs
learn. Like all other dogs, my own, Benji, loves going for a walk. I know this
because he gets excited at the first signs that a walk might be imminent. These
include things like me locking the back door and putting my shoes on. Now,
Benji has no idea what the concepts of ‘locking the back door’ or ‘putting my
shoes on’ are, but he does know that when these two events happen in close
succession then there is a high probability of me taking him for a walk. In
other words, he is completely naive to what the preceding events mean—they
are just data points to him—but he can correlate them into a probable
outcome.
(This dog/AI analogy is quite useful and can be extended further: my dog
is quite lazy, so if he sees me lock the back door but then put my running shoes
on, he goes and hides to make sure I don’t take him with me. In this scenario,
he is using increased granularity to calculate the outcome this time—it’s not
just ‘shoes’ but ‘type of shoes’. Of course, he doesn’t know that my running
shoes are specially designed for running, just that they are different enough
from my walking shoes. It may be the different colour/shade, a different smell,
the different place where they are kept and so on. This demonstrates the
opaqueness issue I discuss in a previous section: I have no real idea (unless I
do some pretty thorough controlled testing) what aspect of the shoes switches
the outcome from ‘Excellent, I’m going for a walk’ to ‘Hide, he’s going for a
run’, but it clearly does have a binary impact. I should also point out that the
dog/AI analogy also has its limitations: Benji has lots of other basic cognitive
skills, such as knowing when it is time for his dinner without being able to tell
the time, but because AIs are currently very specialised in their capabilities, an
AI that predicted walks would not be able to predict dinner time.)
So, the naivety of AI systems can be a real headache for its users. Suffice it
to say that the outcomes from the clustering must be used carefully and wisely
if they are to yield their full value. Data scientists and AI developers must be

138

8 What Could Possibly Go Wrong?

aware of the consequences of their creations, and must apply heaps of common sense to the outputs to make sure that they make sense in the context for
which they were intended.

Becoming Over-Dependent on AI
As you will have seen from the many examples provided throughout this
book, AI can achieve some remarkable things, particularly when it is carrying
out tasks that would be impossible for humans to do, such as detecting clusters of like-minded customers (or fraudsters) in databases containing many
millions of data points.
The problem may come when companies become overly dependent on
these systems to run their business. If the only way to identify your best customers or people trying to defraud you is through very complex AI algorithms,
then there will always be a risk that these systems actually stop working effectively or, probably worse, stop working effectively without you realising it.
This all comes down to the complexity of the problems that are being
solved and therefore the minimal understanding of how they are really working. I have already covered the lack of transparency of how an AI may come
to a decision, but for very complex algorithms (and one AI solution usually
contains a number of different types of algorithms all strung together) there
will only be a few people—the AI developers and data scientists—that understand how these have been designed and built in the first place. Being dependent on a few (very in-demand) people is clearly a risk for any business. It’s
rather like the world of traditional IT where a core legacy system has been
hand-coded in an obscure language by one developer who could leave the
company, or get run over by the proverbial bus, at any time. The risk can be
compounded further if the AI system is being built by a third party who will
not necessarily have that long-term commitment to their client that an
employee might.
There are mitigating actions though. Most of them are similar to the
approach one would take with the bespoke legacy system coder—make sure
that they have documented everything that they do, and are incentivised to
stick around to help support the system. If they won’t be around for long,
make sure there is a robust succession plan in place.
For a third party, there can be more pressure applied, through contractual
obligations, to get the solution fully documented, but the most important
thing will be to select the right provider in the first place, which I discuss in
more detail in Chap. 9. In that chapter I also write about setting up a Centre

Choosing the Wrong Technology

139

of Excellence (CoE) that will be the guardian of the necessary skills and documentation to maintain and improve the AI solutions in the future.
Another factor to bear in mind when trying to solve complex problems is
that it is difficult to know whether the system is providing a correct, or reasonable, answer. For some AI solutions, such as measuring sentiment analysis, we
can test the output of the machine with what a human would reasonably say
was the correct answer; for example, the machine said the sentence ‘I was very
dissatisfied with the service I received’ would be predominantly ‘negative’ and
a human would agree with this. If there was a high volume of sentences to
assess (as would usually be the case) then a sample can be tested. But for complex problems, such as trading financial instruments or designing drugs, it is
nearly impossible to tell if the machine is making a correct decision. We can
look at the final result (the end-of-day trading position, or the efficacy of the
drug) but we will not be able to determine if this was the best outcome of
all—perhaps more profit could have been made or a better drug could have
been designed.
There is also a more philosophical risk from the over-dependency on AI: as
AI becomes more and more commonplace in our lives, we will eventually lose
the ability to do even the simplest cognitive tasks because we are no longer
practising those skills. Our ability to remember people’s names and ‘phone
numbers, and our ability to read maps are already being eroded by our dependency on smartphones and satnavs.
As AI capabilities become greater, they are bound to impact more of our
cognitive skills. Some people may argue that this is not necessarily a bad thing,
but we have developed all those skills over many millennia for particular reasons and, absent of the technology to help us, we quickly become exposed and
vulnerable.
But, coming back to a more practical basis, AI skills will, over time, become
more commoditised (just as HTML development skills have in the last
20 years) and the problem of dependency on highly skilled AI developers and
data scientists will gradually recede. But for now, the risk of over-dependency
should be built into the AI strategy as soon as it is clear that the solutions will
go beyond an off-the-shelf or simple platform approach.

Choosing the Wrong Technology
The field of AI is fast moving—what was not possible in one year can be
solved in the next. All the drivers for AI that I described in Chap. 2 are all
continuing to get better and more influential—big data is getting bigger,

140

8 What Could Possibly Go Wrong?

storage is getting cheaper, processors are getting faster and connectivity
between devices is now almost infinite. So, how do you choose which
approach and technology to take today, if tomorrow something better could
come along? (Users of Apple products, by the way, face this dilemma every
time they think they want to upgrade anything.)
One of AI’s constraints, that each application can only do one thing well,
is actually an advantage here. If an AI solution has been built up from a number of different capabilities (as per the AI Framework) then each of those
individual capabilities could be swapped out with a newer, better approach if
one comes along. Replacing one chatbot solution with another will not necessarily impact any Optimisation capability you have already built. You will still
have to train the new piece of AI but you won’t necessarily have to throw away
the whole solution.
Say, for instance, that you are using one of the AI platforms exclusively to
provide all your required AI capabilities. If one of the other platform providers brings out a better AI capability for, say, text-to-speech, then it is not too
big a challenge to switch out the current API connection for the new one.
(Some providers will, of course, try to lock you into their platform through
contractual means, or even just canny sales techniques, so you will need to
watch this as you select your approach and provider.)
Bigger investments will, of course, require greater levels of due diligence.
Putting in a multi-million-pound vendor-based AI solution will certainly
commit you to a roadmap that may be difficult to step off. But that is true for
any large IT investment.
It does become more challenging if there is a fundamentally new approach
available. Established users of Expert Systems will have looked at Machine
Learning when it first came along with some envy. But if we assume, as is
likely, that machine learning (and all its associated approaches like DNNs)
will be the fundamental core technology of AI for a long while, then it should
be a relatively safe bet to base a strategy around. (The only technology that
may have a material impact on machine learning is Quantum Computing,
but this is still very much still in the labs and will take decades to deliver practical day-to-day uses.)
Probably the bigger issue with committing to one particular AI technology
is the skills that you will need to retain to develop and support it. Generally,
individual developers will be specialists in a particular tool or platform; therefore, changing the tools may mean changing the developers. If these resources
are being bought in through a consultancy or implementation partner, then
you will just need to make sure that that firm has capabilities across all the
relevant tools and platforms that you may require (I cover this in more detail

Preparing for Malicious Acts

141

in the Vendor Selection section of Chap. 9). If you have started to build up a
Centre of Excellence based around specific technologies though, you may
have to think hard about whether the change is worth it.

Preparing for Malicious Acts
With great power (as Spiderman’s Uncle Ben famously said) comes great
responsibility. All the capabilities that I have discussed throughout this book
have shown the huge benefits that AI can bring. But that power could also be
used for nefarious means as well.
The ability for AI Clustering to, for example, identify customers that may
buy a certain product, could also be used to identify people who are ideal fraud
targets, especially if it is triangulated with other data sources. Criminal AI systems would usually have access only to publicly available data (which is always
more detailed than you think) but, as has been seen with the numerous largescale hacking attacks, other, more private data may also be available to them.
In early 2017, some banks introduced voice passwords for their online services—all the customer had to do was train the system on their voice and they
could then login by simply saying ‘My voice is my password’. Within months
it could be proved that this approach can be fooled using facsimiles of the
user’s voice—the BBC showed in one example how a presenter’s non-identical
twin could log in to his account—and the worry now is that AI-powered
voice cloning technology will be able to do the same job. The AI will need
voice samples to be trained on, but, again, these are surprisingly common
across social media, and if you are a famous person, then widely available.
There is already a commercial mobile app, CandyVoice, which claims to be
able to clone voices.
It won’t just be banking systems that could be fooled by voice cloning technologies; you and I might think we are listening to our mother on the
­telephone saying she has forgotten the password on her bank account, or our
boss asking us to send specific emails to suppliers or to make certain
payments.
Similarly, Image Recognition can be used to subvert Captchas—these are
the small images that appear when you try and make a payment on certain
websites. You can proceed only if you can type the numbers that are in the
photos, or identify only those pictures with, say, cars in. These are meant to
avoid software robots making unauthorised purchases, but are now being
bypassed by armies of people, or clever AI systems, entering the correct
answers.

142

8

What Could Possibly Go Wrong?

The difference between the type of online criminal behaviour we have been
used to in the past and what we see now is the scalability of the AI. Voice and
image recognition can be done at high volume and at low cost—if just a very
small fraction get through then the criminals will be successful. It’s the same
numbers game that spammers exploit.
AI can also be used to socially engineer people’s behaviours. During recent
political campaigns (Brexit in the United Kingdom is one that comes to mind)
software ‘bots’ were used to influence people’s opinions through targeting their
social media accounts with pertinent messages. Social media companies already
use AI to change the order of the articles in a person’s social media feed so that
the most ‘relevant’ are closest to the top. External companies can use a person’s
online behaviour to predict which way they might vote and then try to reinforce or change that using targeted posts and tweets. Extending this to include
more nefarious acts than voting is within the realms of possibility, especially if
the accounts that are being posted from are pretending to be someone else.
Artificial Intelligence systems can also aid malicious behaviour by being
trained to do the wrong thing. ‘Dirty data’ could secretly be fed into a training
set to change the learnt behaviours of the AI. The large search engines are
constantly having to defend themselves against malicious companies who create fake sites to increase their ranking in the search results.
In the earlier section on the issue of naivety in AI systems, I wrote about
Microsoft’s tweetbot, Tay. Tay didn’t understand the nastiness of the ‘data’ that
it was being fed, but Microsoft clearly didn’t expect the deliberate sabotage of
its creation by mischievous users. Although Tay is a slightly humorous example (unless you work for Microsoft, of course), it does show how AI can very
quickly be influenced by dirty data. Extrapolate this to a situation where the
AI is controlling millions of interactions with customers, or even financial
transactions, and you can start to see the potential risks from this.
The fixes for the types of behaviours I have described above are diverse, and
usually very technical (apart from insuring against loss, which should be done
as a matter of course). Ensuring that your voice recognition application cannot be a victim of voice cloning will be an inherent part of its development.
With Captchas, they need to be updated on a regular basis to ensure that the
latest technology is being deployed. To stop dirty data getting in requires the
usual defensive approaches to stop anyone from infiltrating your systems. The
level of defences you put up will depend on the sensitivity of the data and the
risk of anything going wrong.

 

Conclusion

143

Social engineering is more difficult to ameliorate. It is necessary to swim
against an incoming tide of more and more people who want to use, and generally trust, social media channels. The mitigating actions are awareness and
education. Just as young people (mostly) know that someone at the other end
of a messaging conversation may not be who they claim to be, people also
need to be aware that the chatbot at the other end of their chat conversation
may not be who, or what, it claims to be.
As with all online activities, being one step ahead of the criminals is the best
that can be hoped for. AI has the ability to do very good and useful things but,
as we have seen in this section, also has the potential to do bad things at scale.
Simply being aware of what can go wrong is the first step in combating it.

Conclusion
For readers hoping for a book all about the benefits and value of AI, this chapter may have come as a bit of a shock. But, as I said at the start, it would have
been remiss of me not to include details of how AI can also present risks to
businesses, by being either misappropriated or sabotaged.
I deliberately included some of the worst cases of those risks, as awareness
of what can possibly go wrong will be the most important part of being able
to mitigate them. Some of these defensive steps are part of the normal IT
development cycle, some are unique to AI development and some are more
sociological. For big, complex AI projects, you will be dependent on strategists, developers, security experts, data scientists and sociologists to ensure
that your application is not open to abuse or unnecessary risk.
In Chap. 10, the final one, I will cover some of the more philosophical
debate around the impact of AI, including how it affects jobs, and the big
question of what happens when (if ever) AI gets smarter than us.
But, before that, we need to look at how your AI efforts and projects can be
industrialised and embedded into your business so that they provide sustainable benefits for the long term.

144

8 What Could Possibly Go Wrong?

The AI Ethicist’s View
This is an extract from an interview with Daniel Hulme, CEO and Founder
of Satalia, a London-based AI development firm with a strong ethical
mission.
AB: Daniel, first tell me about Satalia, and how you approach AI.
DH:	In this capitalistic world right now, businesses are naturally looking at AI
to either make more money or to reduce costs. Satalia helps companies do
this by providing full-stack AI consultancy and solutions, but our higher
purpose is to ‘enable everyone to do the work they love’. We think we can
achieve that by building a completely new ‘operating system for society’
that harmonises technology with philosophy and psychology. We are taking the core learnings from the solutions we have developed for our clients and then making them available as blueprints and tools for people to
use to make their businesses and lives better. Satalia is also a global voice
for purposeful AI start-ups, of which there are not enough of in this world
right now.
Within Satalia itself, our staff are free to do what they want—they set
their own salary, working hours and vacation days, and they have no
KPIs. We combine AI and organisational psychology to enable our
employees to work on the projects they want to, unburdening them
from bureaucracy, management and administration. This gives them the
freedom to rapidly innovate in ways that they would rarely find anywhere else.
AB:	Can you talk to me more about how you see the ethics of artificial intelligence playing out?
DH:	There are three perspectives of AI that need to be considered differently if
we are talking about ethics.
The first is where decisions are made through a trained but static
model. The ethical question is who is responsible if that model doesn’t
behave according to our norms—if it is racist or sexist for example. This
is the Unintended Bias problem. This is linked to the big challenge of
building explainable a­lgorithms—how do we have transparency from
what is essentially a black box model? Now, if governments legislate for
these things, will that stifle innovation? Or to stay competitive should
companies be allowed to ‘move fast and break things’ as the current trend
goes? AI developers have a real responsibility in the middle of all of this,
and breaking things without careful consideration of the impact is not
the best way forward.

The AI Ethicist’s View

145

Secondly, there are the more general AI solutions where multiple types
of AI are combined into an ever-adapting system that changes its model
of the world when in a production environment. The most famous example of this is the Trolley Problem that designers of driverless cars are facing. Should the crashing car hit the child or the three adults, if there is a
choice? What if the car adapted its model in an unpredictable way ‘deciding’ to hit as many people as possible? You might actually end up with a
Preferences setting in your car that allows you to set these choices ahead
of time.
Here’s another example: in a burning building, there is a chance to
save either a baby or a suitcase full of £1 million. Most people’s instinctive
reaction is to save the baby, but if you think about it, perhaps the best
action for society would be the suitcase of money, as you could probably
save many more babies’ lives using that money. What is the right decision
for society? We’re at an intriguing time in humanity whereby we are having to start to codify our basic ethics.
Liability is also a big challenge with these adaptive systems. If a CEO
decides that his company should make an AI-powered machine that
administers medication to people, who is liable if that machines makes a
bad medication decision because it has learnt the wrong thing since it left
the factory? It’s almost impossible to predict how these algorithms will
behave. Just look at the Flash Crash that hit the financial market in 2010.
The third type of AI to consider is where the machines become more
intelligent than humans—this is the Singularity. I don’t believe that we
can build ethics into a super-intelligent machine (this is the Control
Problem)—they will have to be learnt. You could, perhaps, exploit game
theory to help it learn some ethics. Most people know of the Prisoner’s
Dilemma—the winning strategy is always one of ‘tit for tat’, that is, follow
what your opponent did previously. This is the sort of ethical approach
that a super-intelligent machine could learn, that has been embedded
into humanity over thousands of years: “do unto others as you would
have them do unto you.”
Obviously, the big problem with super-intelligent machines is that
they may not be concerned with the impact on humans for whatever
plans they have, and this is perhaps one of our greatest existential risks.
One thought I’ve had is that we should perhaps help them decide to
depart our planet and leave us in peace, a bit like how we might have
made an offering to the gods in ancient times in the hope they’d have
mercy on us.

146

8 What Could Possibly Go Wrong?

AB:	That’s a lot of different ethical issues to consider, especially if the future of
the planet is at stake.
DH:	We may not see the Singularity come along anytime soon but we absolutely have to think about the impact of an AI-driven capitalistic society,
and to work out—both individually and collectively—how we can create
a radically better society today.

9
Industrialising AI

Introduction
This book is predominately about how to introduce AI into your business, how
to take those first steps in understanding and then implementing prototype AI
capabilities. But at some point, you are going to want to move beyond that,
especially if you think that AI has the potential to transform your business.
So far you should have a good grasp of what the different AI capabilities
are, where they are being used effectively in other businesses, the best way to
create an AI strategy and start building AI prototypes, and being aware of
some of the risks involved.
This chapter then is about how to industrialise those AI learnings and capabilities within your business. It focuses on how you move from a few projects
to a mature capability that can implement most, if not all, of the AI requirements that your business will need in the future.
The AI strategy that was discussed in some detail in Chap. 6 will have given
you an excellent starting point. It should include your AI Ambitions which
will determine how far down the industrialisation road you want and need to
go. I have assumed in this chapter that you will want to at least create a permanent AI capability within your organisation, one that is able to be a focal
point for AI activity, and act as a catalyst for ideas and opportunities. If your
ambitions are more or less than this, then just dial up or down my recommendations appropriately.
You will also recall from the Change Management section in Chap. 6 that
there is often a ‘slump’ after the initial success of a prototype. Now is the time
to ‘go large’ and ensure that the momentum from those early wins is t­ ranslated
© The Author(s) 2018
A. Burgess, The Executive Guide to Artificial Intelligence,
https://doi.org/10.1007/978-3-319-63820-1_9

147

148

9

Industrialising AI

into continued success. Having a robust plan for how that will happen will go
a long way to achieving that. This chapter will give you the foundations for
that plan.

Building an AI Eco-System
You may already have been working with a number of third parties to help
you build a strategy and the initial AI project builds. Some of them you may
want to continue working with on a longer-term basis; others may have just
been useful for a specific task and are no longer required. If you are going to
build a long-term AI capability though, you will need to start thinking about
how much of it will be supported by third parties, and which pieces you
intend to build up internally.
A useful way to think of this is as an ‘Automation Eco-System’. This may
include software vendors, strategists, implementation partners, change management experts and support teams, some of whom will be internal and some
of whom will be external. Some may have a bigger part to play in the early
days, whilst others will become more important once some of the projects
have been firmly established.
If you are planning to use a third party to provide any of the Eco-System
capabilities, be aware that some may be able to provide more than one; an
implementation partner, for example, may also have change management
capabilities. Whether you use both those capabilities from that provider will,
of course, be up to you. I cover some of the subtleties of this in the section on
Vendor Selection later in this chapter.
An Automation Eco-System may end up looking something like this
(Fig. 9.1):
Strategy
Prototyping

Implementation
AI Vendor
AI Vendor

RPA Vendor
Support

AI Vendor

Assoc. Tech. Vendor

AI Vendor

Assoc. Tech. Vendor

Change Management
Fig. 9.1 AI Eco-System

Building an AI Eco-System

149

Each of the capabilities can be provided internally or externally. Apart from
the software vendors, the roles can cover AI specifically, or Automation
(including, say, RPA) generally. I will cover each of these roles in turn.
Strategy This area concerns the sorts of things you are reading about in this
book, particularly those in Chap. 6. These are the activities that I, as an automation advisor, carry out for my clients. It involves building the Automation
Strategy, which will include the Maturity Matrix, the Heat Map, the Business
Case and the Roadmap. It will likely involve the technical approach (vendor
versus platform versus bespoke) and, connected to that, the approach to this
Eco-System. It can also include supporting the selection of the software vendors and the other service providers. Most importantly, it should provide
thought leadership to the organisation.
As the AI market is very complex but relatively young there are very few
advisors that are able to cover the business strategy aspects as well as the technical approaches. Some of these (including myself ) are able to provide prototyping services as well. The connection between these two capabilities is pretty
tight; therefore, it may make sense to cover both of these with one provider.
Although the bulk of the strategy work is carried out up-front, it can be
beneficial to retain the strategy advisor longer term to ensure that the benefits
are realised and the internal capability can be realised. Most clients who wish
to pursue a predominantly internal capability (with little dependency on
external parties) keep a light-touch external strategic involvement to provide
appropriate checks and balances.
Prototyping The building of the prototypes, pilots and PoCs requires a flexible, agile approach, and follows on closely from the output of the automation
strategy (and, in some cases, can form part of the automation strategy). There
are not (at the time of writing) many consultancies specifically focused on
prototyping—this capability is generally provided by either a few of the strategy advisors, or by the implementation providers.
The activities are those that are largely covered in Chap. 7 of this book,
including the build of PoCs, RATs, MVPs and Pilots. If your company is just
starting its AI journey then it makes sense to buy in this capability until you
have enough experienced and skilled resources in-house.
AI Vendor The AI Vendor could include a packaged software provider, a platform provider or a vendor of some of the various development tools that are used
to create bespoke AI builds, depending on the technical approach being taken.

150

9 Industrialising AI

For some of the platform providers, and maybe some of the development tools,
you may have existing relationships in place. No matter what arrangement
though, the software obviously sits at the core of your Automation Eco-System,
and many of the other members of that eco-system will be dependent on those
choices.
It should also be noted that some of the packaged software vendors can
provide their own implementation resources. This is usually the case for those
that have yet developed their software to be ‘implementation partner ready’.
One or two of the platform providers, such as IBM, also have large implementation service capabilities, although other third-party providers can also be
used with their platforms.
It will be likely that you will end up with a number of different software
vendor relationships, irrespective of which technical approach you take. Even
with a platform-based strategy, it is probable that you will need to bring in
other vendors to supplement any gaps in the platform capabilities.
Implementation The Implementation Partners are usually brought in following successful prototyping (unless they are carrying out those activities as
well). They are generally concerned with the industrialisation of the AI capability and can provide a full range of development resources. The Big Four
global consultancies can provide these sorts of services, as well as a number of
their smaller competitors.
Whether you need an Implementation Partner at all will depend on how
keen you are to build an internal capability, and how quickly you want to get
there. A Strategy Advisor and a Prototyping Partner can get you a long way
towards self-sufficiency, but if you want to move fast (or if you are not interested in building an internal capability at all) and have deep pockets then an
Implementation Partner makes sense.
An Implementation Partner will be able to continue the work started during the Strategy and Prototyping stages, and help build and ‘bed in’ those
applications. Depending on their capabilities they may be able to build further applications as well as deliver some or all the Change Management
requirements.
RPA and Associated Technologies Vendor As I detailed in Chap. 4, AI is
rarely the only technology you will need to deploy to realise the full benefits
available. You may also need to bring in an RPA vendor, a Cloud vendor, a
Workflow vendor, and a Crowd Sourcing vendor, depending on the complexity

Building an AI Eco-System

151

of your solution and what existing relationships you already have (many
enterprises looking at AI will already have a Cloud capability, for example).
Some vendors will work better together, either technically or culturally, so
do consider this if you have a choice. Build these eco-system members around
your AI vendors, looking at who they have worked well with before and
whether there may be any particular integration challenges or not.
Your RPA vendor, should you require one, will be an important choice.
Some AI vendors have strategic relationships with RPA vendors, and some
Implementation Partners may also have ‘preferred’ RPA vendors that they
work with. In a very few examples (to date) there are software vendors that
can provide both RPA and AI capabilities, with some of these also offering
Crowd Sourcing as well. Generally though, most enterprises bring in separate
vendors for these crucial capabilities.
Change Management This is an important capability for your eco-system
and should be considered early on in the lifecycle. Many large organisations
will be able to provide a general change management capability internally, but
the specific characteristics of an AI project should be borne in mind to determine whether this will be the most appropriate approach.
The capability can be provided by a generalist third-party provider but
those that can claim to have experience in the unique characteristics of an AI
implementation will usually be preferable (the exceptions being when you
have an existing, trusted provider or you need one with experience in your
specific industry sector).
Support With this I mean the ongoing management of the developed AI
solutions. That could mean ensuring that the applications are available
through management of the underlying infrastructure and networks, or
ensuring that the application is providing accurate and meaningful results. It
could also mean the debugging of any bespoke or procured software. Each of
these activities will generally be done by different parties and will depend
heavily on the technical approach that you take, any existing IT policies and
the complexity of the developed solutions.
For infrastructure support, you may have passed the responsibility to an
outsourcing provider or a Cloud provider, but if you have decided to keep it
in-house your IT team may have to support some unfamiliar technologies,
including the management of high-powered GPUs and large storage arrays.

152

9 Industrialising AI

Management of the AI applications (apart from actual code errors) could
be carried out in-house if you have built up the necessary capability, probably
in an Automation Centre of Excellence, which I describe later in this chapter.
It may well involve recruiting AI developers and data scientists, neither of
which are readily available or cheap at the moment. There are very few providers, though, that specialise in supporting AI applications—they tend to be
either the Prototyping firms or the Implementation providers. The skills will
need to include the ability to understand the business requirements, the data
and the technology, hence the challenge in finding the perfect provider, if that
is the route you are taking.
If the procured or bespoke-build software is faulty, then it will usually be
down to the software vendor to fix it—this is the classic ‘third line’ support and
should be included in any contracts between your organisation and the vendor.
I have not included Sourcing Advisors in this eco-system. They are likely to
be a short-term requirement at the start of the journey, and therefore would
not form part of an on-going capability. I do talk more about the role of
Sourcing Advisors in the next section though.
Whether you end up creating an eco-system with these elements, or you
break it into smaller chunks, it is important to consider all the aspects I have
mentioned above. For those elements that you plan to buy-in, you will need
to have a robust selection process, which I cover in the next section.

Selecting the Right AI Vendor
To a large extent, selecting AI software vendors and service providers will follow the same rules as with any IT vendor or provider, but there are some key
differences with AI that can present some additional challenges as well as
opportunities.
One of the first questions that will need to be answered is what exactly is
being procured. For software, the answer to this will depend on the technical
approach that has been decided upon—off-the-shelf, platform or bespoke
build (or any combination of these). There is likely to be one piece of software
that is at the centre of most things, particularly in the early stages, and this
should be the initial focus. This could be the off-the-shelf software capability
that is required for the pilot, or the platform that the bulk of the applications
will be built out from.
Your organisation may be constrained in what software it can choose. This
may be due to the IT strategy stipulating particular criteria (e.g. it must operate as-a-service), specific standards it wants to adhere to (e.g. no JavaScript) or

Selecting the Right AI Vendor

153

general procurement rules (e.g. third-party company must be at least two
years old). Beyond these constraints though, it is important to consider the
following when selecting AI software vendors:
• Proof of capability—because there is so much hype in the market many
vendors will exaggerate their AI credentials in order to attract attention to
themselves. Some will have only a minimal AI capability as part of their
overall solution (maybe a bit of simple NLU embedded in it somewhere)
but claim it all to be ‘powered by AI’.
Now, of course, the solution you actually require may not necessarily
need full-on AI, but if you have a specific application in mind then you will
need to make sure you can cut through the hype. The understanding of AI
capabilities that you will have gained from this book will hopefully go a
long way to helping achieve that, but you can also call on external advice as
well if need be.
An important step in being able to understand and prove the capability
of the vendor will be to have a comprehensive demonstration of the system
and, if appropriate, a small test. Tests, as with pilots and PoCs, can be tricky
for AI solutions as they generally need a lot of data and training to get them
to work effectively. If the vendor can do something meaningful here
though, do take up the offer.
• Proof of value—you should already have at least an outline business case
by this stage, but it is useful to understand how the vendor measures the
value of their solution. Some will have an approach that allows you to
quickly assess their value whilst others may expect you to do all the hard
work on this (this may, of course, mean that they will struggle to create any
value from their solution, and it should be seen as a potential warning). If
the vendor does have a model that can validate your own business case
assumptions, then do take them up on that approach. The key thing is for
the value that they can demonstrate to align as much as possible with your
own business case.
• References—taking up client references should be an obvious thing to do,
but many enterprises don’t bother (and live to regret it later). The challenge with an immature market like AI means that client references and
case studies will be thinner on the ground. How far you want to pursue
this will depend on the perceived fit of the vendor’s solution to your
requirements, their uniqueness and your appetite for risk. If you want to
bring in a vendor who has no existing clients, then you may want to see
what opportunities there are for you to be a ‘foundation client’ for them—
this is covered below.

154

9

Industrialising AI

• Pricing model—the commercial aspects are obviously very important when
selecting a vendor, but it’s not all about the headline price. In Chap. 6
I described some of the different pricing models that are being used, and you
should explore which ones might be the most relevant to you, and whether
the vendor can provide them. For variable pricing, model different scenarios, especially if it materially impacts the business case—a price per API call
may sound attractive at first, but if your application is very popular, those
escalating costs may start to outweigh the benefits. Also, seriously consider
gain-share or risk-reward models, but make sure that they are based on real
numbers that you can measure easily.
• Cultural fit—although you are just buying some software, it is always
worth finding a vendor company that has a good cultural fit with yours.
The AI project may be a long, and sometimes stressful, journey for both
parties, so you need to know that there is some common ground and purpose between both of you. Cultural fit can be assessed simply from having
a good feeling about it, but it can also be worth trying to do this more
objectively, especially if you want it to form part of your selection criteria.
• Future proofing—ensuring that the software you buy will still be relevant
and working in five years’ time can be difficult to determine, but it should
definitely be considered carefully. In the previous chapter, I talked about
the risks of selecting the wrong technology and those considerations can be
applied here. But it is probably the more traditional elements of the technology that are the ones that are most at risk of going out of date. Ask what
technologies the software is built on, and what technology standards does
it rely on. External expert advice may also be useful here.
• Foundation client—procuring software that has not been used in anger
with any other clients may seem like an unnecessary risk, but there can be
benefits from taking this approach. All other things being equal (you have
tested their capability, the technology is good and future-proofed and there
is a good cultural fit) then, versus a more established vendor, a new vendor
may be willing to offer you some significant benefits in return for taking
the risk of being their first client. Usually the benefits amount to heavy
discounts, especially if you also agree to promote your use of the software.
But it can also mean that you get the opportunity to shape the development of the solution, and make it more closely match your own requirements. This will also ensure that you get the undivided attention of the
vendor’s best people, and that can count for an awful lot in delivering a
successful implementation.

Selecting the Right AI Service Provider

155

• Technical requirements—it should go without saying that the technology
needs to fit into your current environment and align with your IT standards and strategy. But with AI you will also need to consider the data. Is
your data the right sort of data for the vendor’s solution, is it of good
enough quality and quantity? These are key questions that must be satisfied
early on in the selection process.
• Professional services requirements—the final consideration is how the
vendor’s software will be implemented. What professional services are
required, and who can provide them (the vendor, third parties or your own
organisation)? If it requires a third party, then understand how many providers out there can do it, and which ones are most familiar with the specific vendor’s software. Some vendors may have a partnership program
where they will have certified or approved service providers. If you are
already developing your eco-system, then you will need to determine how
much commonality there is between these. You will also obviously need to
evaluate the cost of the services at this stage. The selection of service providers is covered in the next section.

Selecting the Right AI Service Provider
In the earlier section concerning the Automation Eco-System I also talked
about the type of service providers that may be required, including strategy,
change management and implementation consultancies—these will also
have their unique AI-related challenges and opportunities when procuring
them in.
Each of the services capabilities you require should be considered as separate ‘work packages’, which can be procured individually or in groups. From
the Eco-System considerations, you should have a starting point for which
capabilities need to be bought in, and how they might be grouped together.
For example, you may think that you need to buy-in Strategic Advisory,
Implementation Services and Change Management, but you think that you
will be able to handle the Support Services in-house, and you also think that
a single provider could do Implementation and Change Management. In this
example then you would be looking for two providers: one for the Strategic
Advisory work package and one to handle the Implementation and Change
Management work packages.
During the selection process (or processes) these initial assumptions should
be tested, but they do provide a useful starting point. You should therefore
provide the flexibility in your Request for Proposal (or whichever procurement

156

9 Industrialising AI

method you are using) for providers to bid for each work package separately
and combined. You may find as you go through the procurement process that
one provider is really good at Implementation, while another is better at
Change Management—in this case you can always adjust your approach to
instead procure two separate providers for these work packages.
You should also test your in-house capability (if you have one) against the
external providers if you are not sure whether to ‘buy or build’. You can treat
them as a bidder in the procurement process and even price up the cost of
their services.
Another important consideration when looking to identify potential providers is your incumbent or strategic providers. Many large organisations have
preferred suppliers and you may have to align with these. AI is such a specialised technology, though, that (unless your incumbents are actually AI specialists) you should argue strongly to bring in providers with the necessary skills
and experience in AI.
The strategic advisory role should be your first procurement consideration,
as many of the other decisions will be based on the Automation Strategy that
is developed with your advisor. Robust, demonstrable methodologies, independence and (maybe most crucially) cultural fit will likely be strong factors
in your selection here. For the procurement of the remaining work packages,
I think the following are the key AI-specific considerations, whether the
resources come from an established provider, software vendors or contractors
(I have used the words ‘service provider’ below to cover these):
• Experience—it may be difficult for a lot of providers to demonstrate a real
depth of experience in implementing AI, simply because the market is so
young, so you will need to bear this in mind. You will need to look for
capability in the particular approaches and tools that you have decided
upon and any relevant partnerships with software vendors (see below).
There will be some experience that is technically specific (skills in a certain
tool, for example) and others that is more industry sector specific (e.g. ability to understand customer data for telecoms firms), and there will realistically be a compromise that has to be made between these two (unless you
are lucky enough to find the perfect match). It may be that you have the
data knowledge already in your organisation.
• Partnerships and Independence—for the tools and approaches that you
have selected, you will want to look for a service provider that has a strong
relationship with these. Many vendors have ‘partnerships’ with providers,
which can range from a loose connection to an almost symbiotic one—it is
important to understand the level and history of those partnerships.

Selecting the Right AI Service Provider

•

•

•

•

157

In most cases the service provider will be rewarded by the vendor for
selling their software licences. But you may also want your service provider
to be independent of the software vendors, and a balance has to be struck
here. Probably the ideal arrangement is where the service provider has
strong, independent relationships with a number of vendors and will be
happy to recommend and implement any of them. If you can’t find true
independence in the service providers then you will need to rely on your
strategic advisor to provide this.
Cultural Fit—this is often overlooked in a service provider selection process but, in my experience, is the one factor that causes most relationships
to break down. Because your AI projects could be quite fraught at times,
you will want to work with a provider that shares your values and objectives, and has people that get on with your people. Cultural fit can be formally measured during the selection process or it can be assessed informally,
but it should not be ignored.
Approach—most service providers will have methodologies, and these
should, of course, be evaluated and compared carefully. But you should
also try and look ‘under the bonnet’ (or ‘under the hood’) to see how
they may approach different challenges, such as testing, poor data or
biased data. Their answers should tell you a lot about their underlying AI
capabilities.
Pricing Model—as with the software vendors, pricing is going to be
important but so is the pricing model. With service providers, there is the
opportunity to introduce ‘gain share’ or ‘risk/reward’-type approaches
where the provider will share some of the risks but also some of the gains
of the project. This helps align their objectives with yours. The challenge
with these pricing models is getting practical metrics—they need to be
general enough that they relate to business outcomes but specific enough
that they can be measured. The metrics also need to be timely—service
providers will not agree to a target that can be measured only in two years’
time, for example. Your strategic advisor, as an independent resource,
should be able to help you identify and implement practical KPIs.
Knowledge Transfer—the final consideration is how the service provider
will transfer all the relevant knowledge to your internal team. This assumes
that, at some point in the future, you will not want to be dependent on a
third party for your AI capability—some enterprises want to get to this
point as quickly as possible, whilst others are happy to remain dependent
for as long as it takes. The knowledge that will be transferred from the service provider to your teams will include a range of things, and will depend
on the relationship and contract you have with them. It may involve some

158

9 Industrialising AI

licensing of Intellectual Property (IP) from the provider (such as methodologies and tools), but will generally be in the form of work shadowing and
the hands-on experience gained by your team. As your resources get more
skilled, and you bring new employee resources in, the service provider can
start to back away; as you ramp up, they will ramp down.
There is another potential third party that you may wish to engage in.
Sourcing Advisors are experienced procurement specialists that can help you
with both the vendor and service provider selection processes. They will be
able to help create a sourcing strategy, find appropriate companies to evaluate,
build and manage the selection process, and support negotiations.
The over-riding requirement for any sourcing advisor is their independence. Then you should look for capability in AI and in your specific industry
sector. If they do not understand the AI market well enough (it is, after all,
sprawling and dynamic) then you may want to team them up with your strategic advisor (or find someone that can do both).
With both the vendor and the service provider selection there will need to
be a fine balance between a robust selection process and the flexibility that
dealing in a young, dynamic technology requires. Be prepared to change your
plans if things don’t initially work out, or better options come along. As the
cliché goes, be ready to ‘fail fast’. And this means ensuring that the arrangements you have in place with your vendors and providers are as flexible as
possible, whilst still demonstrating your commitment to them. Work closely
with your procurement department to make sure that they understand the
special nature of AI projects and the demands that it may place on their usual
methodologies.
Now that we have an approach for bringing in the necessary third parties,
we need to look at the internal organisation and the skills and people that
might be needed.

Building an AI Organisation
There are a number of things to consider when building an organisational
capability to manage your AI (and other automation) efforts, all of which will
be based on how bold your AI ambitions are. You will remember from Chap. 6,
as part of building an Automation Strategy, I stressed how important it was to
understand your ultimate AI ambitions: do you want to simply ‘tick the AI
box’, to improve some processes, to transform your function or business, or
even create new businesses and service lines? Assuming that you are some-

Building an AI Organisation

159

where in the middle of this range, that is, you want to improve processes and
transform some areas of your business through AI, then you will need to create a team or even a CoE to ensure that you are successful and can extract the
maximum value from your efforts.
In this section I will describe what an AI CoE could look like. Based on the
skills, capabilities and ambitions of your own organisation you should be able
to assess which elements are relevant for you, and therefore what your own
CoE might look like. I will also describe how a CoE can be integrated into an
overall company organisation.
It is important for an Automation CoE to have a mission. This will describe,
and help others understand, its purpose. You will be able to find the words
that are most relevant to your own requirements and objectives of your company, but generally the CoE’s mission should be focused on driving the introduction and adoption of AI technologies. It should act as a central control
point for assessing AI technologies and monitoring progress of ongoing projects. Perhaps most importantly, it should provide leadership, best practices
and support for projects and teams implementing AI solutions within the
business.
Two of the key inputs to determine the scale and structure of your CoE,
apart from your AI ambition, are your AI Heat Map and Roadmap. These will
list the functions, services and processes that are being targeted for automation, and will give the priority of how they are likely to be rolled out. Particular
aspects to consider will be the different technologies that will need to be
deployed, the complexity of the solutions and the current state of the data
that will be exploited.
The roles that are generally included within an Automation (and particularly an AI) CoE fall into four main functions: the Management of the CoE,
an overall Architecture Management, the Implementation Teams and
Operations. Some of these roles, including, particularly, parts of the
Implementation Teams, can be provided by third-party providers or vendors,
especially in the early days of the CoE. The following is a guide for how some
organisations are structuring their automation CoEs, but can obviously be
adjusted for your own needs and aligned with your company culture.
The Management Team should ideally start out relatively small, including
at least a manager and some project management capability. The manager
should have responsibility for the whole CoE and the people assigned to it, as
well as the communications across the organisation and upwards (communications could be taken by an internal communications specialist later on). For
management and control of projects, the team (which could just be one person at first) should be responsible for planning, project management, resource

160

9

Industrialising AI

tracking and reporting across all the various AI projects in the CoE. Other key
areas that usually come under this team would include the co-ordination of
training and education for both members of the CoE and users of the
systems.
Taking on the formulation of the business case and solutions are the
Architect Team. This team is best headed up by someone who understands
the business functions and processes but also the technical aspects as well. The
team will look at the opportunities for automation in the wider context (so
would have control of the AI Heat Map) and would create business cases for
each of them, carrying out the initial scoping of effort and technologies, all of
which would be done in close collaboration with the Subject Matter Experts
(or Process Owners or Business Analysts) from the Implementation Team.
The Architect Team would be responsible for managing the pipeline of
opportunities, ensuring that these are being proactively sought, and reactively
received, from across the business. They would also include technical architects (unless this role is retained within the IT Department), data scientists (if
required) as well as managing customer experience.
The Implementation Teams are where the bulk of the CoE resources
would reside. It would consist of a number of project teams, each focused on
a specific solution. Depending on the size and complexity of each project they
could include: a project manager, a project architect, developers (who would
have responsibility for importing and training data, creating and configuring
the models), SMEs/Business Analysts/Process Owners (carrying out solution
design and validation) and Quality Assurance. If appropriate, specialist
resources such as customer experience specialists, linguists, web developers
and integration developers could also be assigned to projects.
Agile development approaches are the most suited to building AI solutions.
The specific type of Agile (e.g. Kanban, Scrum) is not that important, just
that the developers and SMEs work closely together and on fast iterations.
Because of this, and the transformational nature of AI, the members of the
Implementation Team (and the CoE generally) should have a strong mix of
both business and technical knowledge. For regulated processes, which throw
up additional control requirements, there are a number of Agile approaches
that try to cater for these, including R-Scrum and SafeScrum.
Support and maintenance of the released applications (fixing bugs and carrying out any enhancements) can be handled in a couple of different ways.
For small implementations, the original project team can retain responsibility
for support, whilst for larger or more complex projects a separate support
team should be created. Therefore, young or small CoEs have the support
responsibility within the Implementation Team, whereas larger, more mature

Building an AI Organisation

161

CoEs will generally create a new team for this purpose, or have it embedded
as part of the Operations Team.
The final group is the Operations Team. This team will have responsibility
for the deployment, testing, updating and upgrading of any of the live systems. They will also have responsibility for the technical integration of the AI
solutions with any other systems. For organisations used to taking a DevOps
approach (where the Operations and Development resources work as integrated teams), then much of this Operations Team would be subsumed into
the Implementation Team.
As I mentioned earlier, the above groups are only meant as a guide and can
be flexed to match your own company’s requirements and practices. Another
consideration regarding the setting up of a CoE is how it fits into the overall
organisation structure of the company.
Some companies, such as Aviva and Clydesdale and Yorkshire Bank, have
created ‘lab’ or ‘innovation hub’ environments (the Aviva one is actually called
a Digital Garage). These are useful vehicles to generate additional buzz around
the initiatives and to foster a culture of innovation. They usually focus on the
early stage development of ideas. CoEs though tend to be a little more practical and will include most, if not all, of the capabilities to identify, build and
run new technology initiatives.
A key consideration, especially for larger organisations, is whether to have
a single, centralised CoE or to spread the capability across the various divisions or business units whilst retaining a central control function. The
approach that usually works best for AI, at least initially, is to keep everything
as centralised as possible. This is for a couple of reasons.
Firstly, the impetus and momentum for the early AI initiatives usually
come from one area of the business (because of greatest need, biggest opportunities, most enthusiastic management, etc.), and this can form the genesis
of the CoE. When other business units see the benefits that are being generated, then they can tap into the existing capability rather than creating it
themselves from scratch.
Secondly, an ‘independent’ team not involved in the day-to-day running of
the function is much more likely to identify, and be able to implement, transformational change. Left to their own devices, business units rarely consider
the full range of opportunities available and will instead focus on the simpler
process improvements. Both are valid types of opportunities, but AI has much
to offer with regard to transformational change and therefore it should be
promoted at every opportunity.
A final consideration regarding the organisational structure of your industrialised AI capability is whether to appoint a senior executive to oversee it all.

162

9 Industrialising AI

Many large companies, especially in the financial services sector, already have
Chief Data Officers (CDO). These roles take responsibility for the enterprise-­
wide governance and utilisation of information as an asset (including in some
cases a large element of revenue generation). For some companies, such as
Chartis, Allstate and Fidelity, the CDO has a big influence on the overall
strategy of the company.
A relatively new position that is starting to be seen in some businesses is the
Chief Automation Officer (CAO). These can sometimes be called Chief
Robotics Officers or Chief AI Officers (CAIO). This role should try to embed
automation, and AI, at the centre of the business strategy. A CAO, more than
a CDO or a CTO, can look outward and forward. They will tend to have a
greater focus on the business opportunities of automation than, say, a CIO.
(Some analysts believe that a CAO is more suited to a move up to CEO than
a CIO might be.)
CAOs are still relatively rare, and it may be that it is a role that is needed
for a certain period of time, to get automation up and running and firmly
embedded in an organisation, and then have those responsibilities eventually
absorbed back into business-as-usual.
I am not aware, at time of writing, of any organisations with CAIOs yet,
although there is plenty of talk about the role. Again, it might be a short-lived
necessity, and many people are sceptical that this level of focus is required in
the boardroom. It may be that the CDO or CAO (if they exist in a company)
are able to cover AI specifically.
Whether an organisation chooses to bring in a CDO, CAO or CAIO, the
people carrying them out need to have some pretty special capabilities: they
will need at least a reasonable understanding of the technologies and the data
infrastructure; they will need to be able to work cross-functionally, since automation opportunities can exist across the whole business and will require the
cooperation of a number of departments; they will need to behave like
­entrepreneurs within the business and they will need to have the industry
standing and inter-personal skills to attract and retain the best talent. That is
not an insubstantial shopping list.
We have now come to end of the journey to understand, create and industrialise AI in businesses. Throughout the book, I have focused specifically on
what AI can do for organisations today, and the practical approaches that can
be taken to exploit the inherent value in this technology. But now, in the final
chapter, it is time to look to the future: how AI may develop, what new
opportunities and risks it will create, and how we can best plan to get the
most from it.

The Data Scientist’s View

163

The Data Scientist’s View
This is an extract from a dialogue with Richard Benjamins, who, at the time
of the interview, was working as Director, External Positioning & Big Data
for Social Good at LUCA, part of Telefonica, a European Telecoms Company.
He is now Group Chief Data Officer & Head of the Data Innovation Lab,
AXA, a global insurance company.
AB:	In your role at Telefonica’s data-led business, LUCA, you understand the
inherent value, as well as the challenges, in using Big Data for AI applications. Can you tell me more about this?
RB:	Firstly, you have to look at the numbers. Six years ago, McKinsey estimated that Big Data would bring $300bn in value for healthcare and
€250bn for the European Public Sector, but in their most recent update
at the end of 2016, they reckon that the numbers were actually 30%
higher than this.
But none of that tells you how you should get the value from your own
big data and how to measure it, and that’s the biggest challenge for organisations who want to be able to realise that value. I think there are four ways
that you can get at that inherent value.
Firstly, and probably most simply, you can reduce the cost of the IT infrastructure that is being used to manage your big data, using open-source tools
such as Hadoop. This can save significant amounts of money, and is easy to
measure. Secondly, big data can be used to improve the efficiencies of your business, that is, enable you to do more with less. Thirdly, it can be used to generate
new revenue streams from your existing business—this can be a challenge
though as it is difficult to know where to start, and it can be hard to measure the
value. The final source of value is where completely new revenue streams are created from big data, that is, new products where data is at its centre, or by creating
insights from data to help other organisations optimise their own businesses.
I believe that in the next few years, the dominant source of big data value
will be from business optimisation, that is, by transforming companies into
data-driven organisations that take data-driven decisions.
AB:	So how do you start to put a number against business optimisation
through big data?
RB:	This is the big challenge. Of course, you should have a base case and compare the difference before and after, but big data initiatives are rarely the
only reason for doing something, so it’s difficult to isolate the value. There
are two ways to help with this though. The first is to do a segmented
experiment; for example, have one group of customers that are being analysed through big data, and another that are not (and maybe even another

164

9 Industrialising AI

where a different approach is being used). The second is to do something
with big data that has never been done before – that way you can compare
the results directly with your base case.
From my experience, though, mature organisations know that there is
value in big data and will not be overly obsessed with trying to measure
every bit of value. When companies reach the stage where big data has
become business-as-usual, then it changes the way that departments interact
and the value naturally flows from this.
AB:	In order to manage all of this data and value, the role of the Chief Data
Officer (CDO) must be crucial?
RB:	Definitely. We are seeing a big increase in the number of businesses that
have a CDO. According to one survey I have seen, 54% of companies
now have a CDO, up from just 12% in 2012.
Where the CDO sits in the organisation is an open question still. It’s getting closer to the CEO level every year—in Telefonica for example, five
years ago the CDO was originally five levels below CEO but now reports
directly into that position.
The best position for a CDO to be in is where they have responsibilities
across various functions or departments, so that their role, and objectives,
are not defined by a single department. Reporting into the COO is probably a good balance as they will then have good cross-functional visibility.
AB:	Big Data is obviously closely related to Artificial Intelligence. Do people
get confused between these terms?
RB:	The term Artificial Intelligence can be used for many things including Big
Data and Machine Learning. The hype that is around AI can be useful
because it brings an increase in interest and attention but it is important
to keep in mind what we are talking about. It is not just about building
wonderful applications, it is also about asking fundamental questions
related to how people think, how they solve problems and how they
approach new situations. When thinking about AI I think it is important
to consider these three things: AI can solve complex problems which used
to be performed by people only; what we consider today as AI may just
become commodity software; and, AI may help us shed light on how
humans think and solve problems.
And remember, in the big scheme of things, this is only the beginning...

10
Where Next for AI?

Introduction
This book has always been about what executives need to do today to start
their AI journey. Rightly so, but it is also important to understand what is
coming down the line. Organisations need to be prepared as much as possible to exploit any future advancements as well as mitigate any of the developing risks.
This final chapter tries to predict the near future for AI. This is, of course,
a bit of guessing game, but hopefully it is a well-informed one. Firstly, I look
at which of the AI capabilities I have discussed in the book are likely to flourish and thrive the most, in other words which should be watched most closely
for significant developments.
I then have a go at predicting when AI might become ‘business as usual’,
that is, when we are no longer talking about it as a new and exciting thing, in
a similar way that outsourcing has now become simply another tool for the
executive to deploy. This section covers some of the more general predictions
about AI and its use in business.
And finally, I close the book with some words of advice about how to
future-proof your business (hint: it includes AI).

© The Author(s) 2018
A. Burgess, The Executive Guide to Artificial Intelligence,
https://doi.org/10.1007/978-3-319-63820-1_10

165

166

10

Where Next for AI?

The Next Opportunities for AI
In Chap. 3 I described the core capabilities of AI and then in Chap. 5 I gave
numerous examples of how those capabilities are being used in businesses
today. But the technology will inevitably get better, and the speed, accuracy
and value of these capabilities will only increase, probably at an ever-­
accelerating pace.
This section provides my summary of how each of those capabilities might
develop over the next few years—some will accelerate whilst others may hit
some bumps in the road.
• Image recognition has already advanced dramatically over the last five
years and, particularly, over the last couple. Its development over the next
few years will continue at this heady pace as more image sets become available for training purposes, and better algorithms, faster processors and
ubiquitous storage will all contribute to the ability to improve accuracy but
also tag moving images as well. There has already been some progress in this
area but I think that we will see objects in films, clips and YouTube movies
identified and tagged automatically and efficiently. This would mean you
could, for example, search for an image of a DeLorean car across all of the
movies online, and it would return the frames in the films where the car
appeared.
The next development beyond that will be to identify faces in movies,
just as Google Images and Apple’s Photo applications can do today for still
photos. Being an optimist I am hoping that the medical applications of
image recognition, such as in radiology, will really start to take off—they
will have a huge impact on society, and therefore deserve to be promoted.
The technique of having AI systems play (or battle) with each other,
called GANs, will enable systems to create new images from scratch. Put
simply, one system, after an initial period of training, will create an image,
and another will try and work out what it is. The first will keep adjusting
the image until the second system gets the right answer. This technique
could also be used to create scenery for video games or de-blur video images
that are too pixelated.
Image recognition will have to face the challenges around bias, which
are certainly coming to the fore at the time of writing. Ensuring unbiased
image sets should be a mission for all those creating and controlling data,
and it shouldn’t become an issue only when things have clearly gone
wrong.

The Next Opportunities for AI

167

• Speech Recognition, as can be evidenced in domestic devices such as
Amazon’s Alexa, Google’s Home and Apple’s HomePod, is already very
capable, although there are plenty of examples of poor implementations as
well. I predict that this capability will become much more prevalent in a
business-to-business environment, powering, for example, an HR
Helpdesk, or a room booking system.
Real-time voice transcription already works reasonably well in a controlled environment (a nice, quiet office with a microphone on a headset)
but advances in microphone technology in parallel with better algorithms
will make speech recognition viable in most environments.
In the section on malicious acts in Chap. 8 I talked about voice cloning.
As long as this technology can stay out of trouble (or the people wanting to
use it for good can keep one step ahead of the hackers) then this has the
potential to help many people who have voice-crippling diseases.
• Search capability really took off in 2017, especially in specialised sectors
such as legal. There are a number of very capable software vendors in the
market which means that this will be a useful introduction for businesses as
they look to adopt some level of AI capability. The algorithms will inevitably get better and more accurate, but I think that the future for this capability is as part of a wider solution, either by extending its own capabilities
(e.g. to be able to auto-answer emails from customers) or by embedding it
in others (as part of an insurance claims processing system, for example).
• Natural language understanding is closely connected to both Speech
Recognition and Search, and will benefit from the advances there, particularly in areas such as real-time language translation. But I think the hype
and elevated expectations that have surrounded chatbots will have a detrimental effect on their adoption. I predict that there will be a backlash
against chatbots before they start to make a recovery and become widely
adopted and used every day. Those firms that implement chatbots properly
(i.e. for the right use cases, with appropriate training, and capable technology) will see successes, but those that try and take shortcuts will speed the
oncoming backlash.
The sub-set of NLP that focuses on generating natural language (NLG)
will see some big strides, and therefore wider adoption, over the next few
years. It will benefit from broader acceptance of AI-generated reports, particularly where they add real value to people’s lives, such as with personalised, hyper-local weather forecasts. As more data is made available
publicly, then more ideas will surface on how to extract value from it and
to communicate those insights to us using natural language.

168

10

Where Next for AI?

• Clustering and prediction capabilities can, for this purpose, be considered
together. Two things will happen regarding the use of big data: firstly, as
businesses become more ‘data aware’, they will start to make better use of
the data they have, and also exploit publicly available data as well; and,
secondly, less data will be needed to get decent results from AI, as a­ lgorithms
become more efficient and data will become ‘cleaner’ and (hopefully) less
biased.
Financial Services will be the main beneficiary from Prediction, being
used in more and more companies (not just the global banks) to combat
fraud and up-sell products and services. Cross- and up-selling will also
become the norm for retailers.
The challenge for these capabilities will be whether they will be used
productively and ethically. Increasing the number of click-throughs and
delivering targeted adverts may make sense for some businesses, but it is
never going to change the world, and will no doubt frustrate many customers. And, of course, the questions around bias and naivety will remain on
the table and will need to be addressed for every use case.
• Optimisation, out of all the capabilities, has, for me, the greatest promise.
The work that is being done with Reinforcement Learning (that was at the
heart of AlphaGo’s victory) has enormous potential in the worlds of risk
modelling and process modelling, and specific examples where energy bills
can be significantly reduced (as at Google’s data centre) demonstrate that it
should have broad and important applications for us all.
One of the biggest constraints on AI right now is the availability of
properly labelled, high-quality data sets. GANs, which I mentioned earlier,
have the potential to allow AI systems to learn with unlabelled data. They
would, in effect, make assumptions about unknown data, which can then
be checked (or, more accurately, challenged) by the adversarial system.
They could, for example, create ‘fake but realistic’ medical records that
could then be used to train other AI systems, which would also avoid the
tricky problem of patient confidentiality.
But reinforcement learning and GANs are the AI technologies that
could still be described as ‘experimental’, so we may not see these being
used on a day-to-day basis for a few years yet. In the meantime, we have
systems that can make important decisions on our behalf, including
whether we should be short-listed for a job, or approved for a loan. The
main advances here, I think, will be in how they are used, rather than any
major advances in the underlying technologies. Eliminating bias and providing some level of transparency will be the key areas to focus on.

When Will AI Be Business as Usual?

169

• As we are looking into the future, it is worth considering some of the questions around Understanding, which I stated in Chap. 3 was still, as an AI
capability, firmly based in the AI labs, and may be there for a very long
time, if not forever. Having said that, a really interesting area of development to keep an eye on is that of Continual Learning. This is an approach,
currently being developed by DeepMind, that hopes to avoid ‘Catastrophic
Forgetting’, the inherent flaw in neural networks that means that a system
designed to do one thing won’t be able to learn to do another unless it
completely forgets how to do the first.
Because the human brain learns incrementally it is able to apply learnings from one experience and apply it to another, but neural networks are
extremely specialised and can learn only one task. DeepMind’s approach,
using an algorithm they call Elastic Weight Consolidation (EWC), allows
the neural network to attach weight to certain connections that will be useful to learn new tasks. So far, they have demonstrated that the approach
works with Atari video games, where the system was able to play a game
better once it had been trained on a different game.
Although there are a number of labs looking at tackling Artificial General
Intelligence, EWC is probably the most exciting opportunity we have for
getting closer to an answer to this most intriguing of challenges.

When Will AI Be Business as Usual?
In the previous section I talked about the developments I expect to see in the
next few years for each of the AI capabilities. In this section I want to be able
to generalise those predictions and look at the wider context around
AI. Specifically, I will look at the question of when (if ever) AI will become
‘business as usual’.
The traditional way of understanding the maturing of any technology is to
see where it is on the Gartner Hype Cycle. This is a well-used graph which
plots the life of a technology from its initial ‘technology trigger’, through its
rise in popularity to a ‘peak of inflated expectations’ that is then followed by a
decline into the ‘trough of disillusionment’. After that it recovers through the
‘slope of enlightenment’ and finally reaches the ‘plateau of productivity’. It is
this final stage that can be defined as business-as-usual.
The challenge with plotting AI on this curve is that AI is many different
things and can be defined in many different ways. We could plot some of the
individual capabilities, such as NLU (heading into the trough of disillusionment) or some of the technologies, such as machine learning (right at the peak

170

10

Where Next for AI?

of inflated expectations) or GANs (as a technology trigger). To generalise
though, most of the machine-learning-based AI technologies are pretty much
peaking with regard to expectations and are therefore primed to enter the
steep slide into disillusionment.
Being in the trough doesn’t mean that everyone abandons the technology
because it doesn’t work anymore—in fact it’s usually quite the opposite.
Developments will continue apace and great work will be done to create new
applications and use cases. It’s just that the perception of the technology across
businesses and society will shift; there will be less hype, with people taking a
more sober, considered approach to AI. The best thing about being in the
trough of disillusionment is that everyone’s expectations are now much more
realistic. And that surely must be a good thing.
How long it stays in the trough will depend on a number of factors. The
four key drivers I described in Chap. 2 (big data, cheaper storage, faster processors and ubiquitous connectivity) will only continue on their current trajectory. Big data may be the one area that makes or breaks it through.
I think the greater availability of public data will be one of the main reasons
that AI is adopted by the mainstream and reaches the plateau of productivity
quicker. But there are challenges to this, as the tech giants hoard their own
proprietary data sets farmed from their billions of ‘customers’ (who, by the
way, should more accurately be called suppliers, since they provide all the data
that is used by the tech companies to sell advertising space to the advertisers,
their actual customers).
But even their data will become more difficult to farm—there is a complicit
deal between the data provider (you and me) and the user (tech giants) and
this will need to continue despite being constantly eroded by examples of
misuse or abuse or hacking. It’s clear what the tech giants need to do to keep
their supply of data coming in: there is a balance of utility, trust and transparency in the use data, and increased transparency can often compensate for
lower levels of trust.
Once we have more publicly available data and higher levels of transparency from the tech giants we can start to exploit all the connections between
these data. Rather than just having public transport data, for instance, we can
bring in our own personal data that might include our inherent preferences
for how we like to travel. (These could be set manually, of course, but better
to pull that information from actual, real-time behaviours rather than what
we think we liked a year ago.)
This idea of hyper-localised, hyper-personalised information may well
be the key to widespread acceptance of the benefits of AI. Some examples
can already be seen (e.g. hyper-localised next-hour weather forecasting,

When Will AI Be Business as Usual?

171

hyper-­personalised what-to-wear recommendations) but the bringing to
together of all the different sources of data to provide really useful information that is relevant to just you at that specific time in that specific location
will be, to use a cliché, a game changer.
If there is trust and transparency around the data that consumers find useful,
then they are more likely to allow businesses open access to that, therefore
increasing the utility even further. (This is a mind-set change from today where
consumers know their data is being exploited but don’t ask too many questions
because they believe the benefits are usually worth it.) And businesses will benefit from this as well, and can use the same ‘three-dimensional data’ approach
to help run their businesses better internally—salespeople can be much better
informed, as can couriers, caterers, HR managers, executives and so on.
So, there is a huge dependency on the levels of trust in data usage, which
will be driven by the level of transparency that there is in how the data is
farmed and used. Other areas that might delay the date at which AI reaches
the plateau of productivity could be a general backlash against the technology,
especially as the amount and types of jobs begin to be materially affected.
Right now, though, one of the biggest challenges to the widespread adoption of AI is actually a shortage of skilled and experienced people. Data scientists and AI developers will become superstars (with superstar salaries) as the
tech giants and AI organisations, such as OpenAI, fight over these scarce
resources. The pipeline for supplying skilled people, which includes the education systems of the world, will take time to adapt, with quick fixes such as
online training courses taking the immediate strain.
People will also make a difference on the demand side as well. The
Millennials, as we all know, have grown up with many of these technologies
and are both familiar and dependent on them. They will absorb new AI applications without thinking and will have fewer issues with some of the ethical
challenges I have described (although this is clearly a generalisation—you
only have to read the interview with Daniel Hulme in this book to realise that
there are people who care deeply about the dangers of AI and how it can benefit the world).
Customers for AI will therefore have a much bigger impact on the adoption
of the technology. It is fair to say that AI is currently a vendor-driven market—the innovations and ideas in the market are mainly borne out of the fact
that it can be done, rather than it needs to be done. This balance will slowly
shift over the coming years as people became more familiar with the capabilities of AI and there are more examples of its use in society and in the
workplace.

172

10

Where Next for AI?

Things like speech recognition, and particularly voice authentication, have
already become part of some people’s normal lives as they become familiar
with, and then maybe dependent on, their Google Home, Amazon Alexa or
Apple HomePod devices. Once this technology enters the business
­environment, in, say, an HR Helpdesk, we will start to see the beginnings of
AI becoming ‘business as usual’.
The ‘fully connected home’ dream that AI product vendors try and sell currently falls far short of expectations—there are lots of different products and
standards out there which rarely work together and can be difficult to set
up—but one can see signs that the seamless and effortless use of AI to help
run our homes is on the horizon. So when people go to work, they will expect
the same sort of experience, and this will drive further developments, and the
wider acceptance, of AI in business.
One of the most interesting developments I am starting to see is the creation of ‘vertical’ AI solutions, where a number of different AI (and associated
technology) capabilities have been brought together to solve a particular challenge in a particular domain, such as financial services, or HR, or healthcare.
Many of the AI ‘solutions’ being developed right now fall short of being complete solutions—they just look at a very specific capability or service—and
these will quickly become commoditised or embedded into existing enterprise
systems. ‘Full stack’ vertical solutions, however, will have AI at their core, and
will focus on high-level customer requirements that can be better met using
AI, or will be used to discover new requirements that can be satisfied only by
using AI. These solutions will take a much more holistic approach and draw
upon subject matter expertise just as much as they do technical know-how. In
my mind, once we start to see more and more of these types of solutions, the
quicker AI will achieve business-as-usual status.
Another bellwether for AI’s acceptance onto the plateau of productivity
will be how business’s organisational structure changes to accommodate
AI. There are already a number of enterprises that have created Automation
(and AI) Centres of Excellence (I discussed these in Chap. 9) and these will
become much more commonplace in the next few years. Full acceptance
though will come when enterprises no longer need these CoEs—when AI is
seen as the normal way to do business. By this point we won’t have AI specialists—everyone will be a generalist with AI skills. Just as in the 1970s there
were people who had jobs ‘working the computer’ and now everyone just
‘works with computers’, so the same will be true for AI.
There is a slightly tongue-in-cheek definition of AI which says it is ‘anything that is 20 years in the future’, but there is probably some truth in that.
The AI capability that we have now and use every day would have seemed like

Future-Proofing Your Business

173

black magic 20 years ago, yet we don’t really consider it to be AI anymore. We
are always looking forward for the new shiny thing that we can label AI. So it
is almost a self-defeating exercise to try and predict when AI will become business as usual. I have, though, tried in this section to identify those things that
will signal when AI (as we consider it today) has reached the plateau of productivity. But if your business is waiting until then before looking into or
adopting AI then you will be too late. The time to start your AI journey is
now: to future-proof your business so that you can defend against, and prepare for, the new AI-driven business models and opportunities that will inevitably come.

Future-Proofing Your Business
In the very first chapter of this book I implored you to ‘not believe the hype’.
Over the subsequent nine chapters I have hopefully provided you with a practical guide on how to approach AI so that you are as informed and prepared
as you can be. I also hope that you are now as excited as I am about the benefits that AI can bring to business and society, and that you can see how AI
could be beneficial to your company.
But the most important thing I want you to have taken away from this is
that every company, every senior executive, needs to be thinking about AI
now, today. The development of this technology is already moving apace but
it will only get faster. Those who wait will be left behind. Now is the time to
start future-proofing your business to capture the full value from AI.
There are three broad phases to this: understand the art of the possible,
develop an AI strategy and start building stuff.
Understanding the art of the possible is the first stage in future-proofing
your business. By reading this book you will have taken an important step
forward, but you should also be reading many of the very decent books and
resources that are also available, many of them for free on the internet. There
are, of course, significantly more poor resources on the internet than there are
decent ones, but sites such as Wired, Quartz, Aeon, Disruption, CognitionX,
Neurons, MIT Technology Review, The Economist, The Guardian and The
New York Times make very good starting points. You should also consider
attending conferences (there are now a plethora of automation and AI conferences available, some of which you may find me speaking at) and also signing
up for AI masterclasses and/or boot camps.
The ‘art of the possible’ should, of course, be grounded with some level of
realism. Therefore, it is important to understand what other enterprises are

174

10

Where Next for AI?

doing, particularly those in the same sector or with the same challenges.
Advisors can be useful for this as they (we) tend to have further insights than
those that are available publicly. Advisors will also be able to challenge you on
what can be done and to help you ‘open your mind’ to all the available
opportunities.
Some useful thought experiments that can be done at this stage include
imagining what a company like yours might look like in 10 or 20 years’ time,
or to try and imagine how you would build a company with the same business
model as yours from a blank sheet of paper—what are the core things you
would need, what could you discard and what could be delivered through
automation? Holding ‘Innovation Days’ can also work—these would bring
together stakeholders from the business to listen to the art of the possible
(from experienced third parties) and have demonstrations from AI vendors of
their solutions. This should then spark new ideas that can be fed into the AI
Strategy.
Following on from the initial thinking, the next phase will all be about
building the AI Strategy. A good chunk of this book describes this process in
detail, so I won’t repeat myself here. Suffice it say that the strategy should be
built around identifying and addressing existing challenges, as well as identifying new opportunities that can be addressed by AI.
Creating and delivering innovation within any business is never an easy
ride. It takes time and effort, with planned investments and a clear mandate.
For AI, you could create a small working group with that specific focus, or
you could sit on the back of any innovation capability you may already have
in-house. The risk with the latter of these is that you may end up with the
‘lowest common denominator’ and miss out on some of the bigger AI opportunities. The advantage is that you can call upon additional resources and
other technologies that may be instrumental to your solutions.
The important thing to remember about any program to implement innovation, but particularly with AI, is that it will require a cultural shift. People
will need to think very differently, and they will be challenged about many of
their core beliefs of how a business should be run. Be prepared for some heavy
discussions.
The final phase to start future-proofing your business is to build some AI
capability. Based on your AI Strategy, building some prototypes will be the
first real test for AI, and the first time that many people will have seen AI in
action in your business. This act of disruption will have a profound affect and
will be a great catalyst to start new conversations across the business. Ensure
that you make the most of this opportunity to showcase AI and bring the rest
of the business along with you.

Final Thoughts

175

You should always be ready for new developments in the technology. AI is
always getting better—advances that we weren’t expecting for decades are suddenly with us, and you need to be informed and ready to bring those into
your business if they are relevant. Having an AI, or automation, CoE, perhaps
backed up with external advice, will ensure that the market is being constantly
scanned and all potential opportunities are captured as soon as possible.
Some businesses, once they have some prototypes under their belt, take the
disruptive power of AI to the next level and build parallel digital businesses.
These in effect compete with the legacy business, and are therefore not encumbered by existing systems or culture. Even if you do not create a new business
unit, this healthy competitive mind-set can be useful to bring out the full
potential of the technology.

Final Thoughts
This has been a fascinating book to write, and I hope that you have, at least,
found it an interesting and thought-provoking read. I’d like to think that you
are now much better prepared and informed to start your AI journey with
your business.
That journey won’t be an easy one—disrupting any business with new technology has more than its fair share of challenges—but the rewards should
certainly be worth it. Be very aware that all your competitors will be considering how AI can help their businesses, but the advantage will be to those who
actually start doing something about it now. As I said in the previous sections,
if you are waiting for a mature and stable AI market then you will be too late.
Thank you for spending your precious time in reading my book. If you
need any further help or advice, then you know where to find me. But for
now, put down the book, and go start your AI journey.

Index

NUMBERS AND SYMBOLS

B

1-800-Flowers, 76

Baptist Health South Florida, 81
Barclays, 60, 77
Baxter, 64
benefits, 1, 6, 18, 22, 30, 37–9, 41, 50,
51, 59, 60, 63, 64, 70, 73, 77,
78, 80–2, 84, 87, 89, 92–5, 98,
100, 102, 103, 105–9, 112, 113,
129, 141, 143, 149, 150, 154,
161, 167, 170, 171, 173
Berg, 87
Berkeley, 18
bias, 27, 47, 48, 77, 86, 87, 125,
130, 131, 134, 136, 144,
157, 166, 168
big data, 3, 11, 13–15, 19, 25–7, 41,
56, 66, 87, 95, 114, 125, 129,
136, 139, 163, 164, 168, 170
Brown, 18
business case, 27, 73, 98, 103,
105–13, 115, 118, 149,
153, 154, 160

A

Advanced Research Projects Agency
(ARPA), 12
Ageas, 80
AIA, 78, 79
Alexa, 2, 17, 27, 34, 39, 76, 167, 172
Alexander Mann, 86
algorithms, 1, 2, 6, 9, 14, 18–20, 26,
27, 36, 43, 45, 47, 50, 54, 57,
58, 68, 83–5, 114, 119, 124,
125, 129, 132, 136, 138, 144,
145, 166–9
Amazon, 2, 16–18, 27, 34, 56, 57, 68,
72, 76, 119, 167, 172
ambitions, AI, 100, 106, 113, 147,
158, 159
Apple, 17, 27, 140, 166, 167, 172
Associated Press (AP), 82
automation, 2, 3, 5, 6, 24, 53, 58, 71,
85, 91–4, 96–8, 100, 107, 108,
110–13, 119, 123, 149, 158–60,
162, 173–5
Axa, 84, 163

C

Care Quality Commission (CQC), 87
Celaton, 70, 78, 79, 88, 89

© The Author(s) 2018
A. Burgess, The Executive Guide to Artificial Intelligence,
https://doi.org/10.1007/978-3-319-63820-1

177

178

Index

change management, 8, 89, 110,
112, 113, 120, 147, 148,
150, 151, 155, 156
chatbots, 5, 12, 39, 45, 57, 63,
74–8, 82, 83, 93–5, 97, 101,
102, 115, 132, 140, 143, 167
cloud, 16, 18, 19, 26, 52, 55–8, 76,
91, 114, 119, 150, 151
clustering, 3, 4, 37, 38, 41, 46,
51, 66, 93, 95, 108, 136,
137, 141, 168
Clydesdale and Yorkshire Banking
Group (CYBG), 77, 161
Cognitiv+, 53
cognitive reasoning, 6, 43, 62, 74,
80, 101, 107, 114–16, 124
consultancy, AI, 144
Cortana, 2, 39
crowd sourcing, 55, 67–9, 102,
125, 131, 150, 151
Customer Satisfaction (CSAT), 93,
103, 107
customer service, 4, 18, 24, 30, 36,
52, 61, 63, 73–8, 84, 89, 93,
99, 101, 102

E

D

H

DARPA, 12
data, 1, 13, 30, 55, 74, 95, 118,
129–32, 152, 166
data scientists, 23, 54, 57, 119–21,
124, 130–2, 137–9, 143,
152, 160, 163, 171
Davies Group, 79
DeepMind, 42, 45, 50, 66, 83, 169
Deep Neural Networks, 8
Deutsche Bank, 81
developer, AI, 3, 18, 23, 55–7,
135, 137–40, 144,
152, 171
Durham, Police Force, 86, 130

hacking, 141, 170
heat map, AI, 100–6, 112, 113, 159, 160
Hidden Markov Model (HMM), 34
HMM. See Hidden Markov Model
(HMM)
HSBC, 77, 86

eco-system, 148–52, 155
expert systems, 11–13, 41, 43, 124,
132, 133, 140
F

Facebook, 13, 15, 16, 32, 45, 83
Farfetch, 87
framework, AI, 3, 4, 7, 29–54, 57, 73,
101, 113, 120, 140
G

GANs. See Generative Adversarial
Networks (GANs)
general artificial intelligence, 4, 23, 48,
56, 65, 70, 73, 145, 165, 169
Generative Adversarial Networks
(GANs), 45, 131, 166, 168, 170
Goldman Sachs, 85
Google, 2, 13–15, 18, 32, 40, 42, 50,
57, 66, 68, 83, 84, 86, 94, 119,
166–8, 172
graphical processing units (GPUs), 17,
26, 57, 151

I

IBM, 16, 18, 19, 25, 26, 32, 75–7, 80,
119, 120, 150
image recognition, 3, 7, 20, 32, 33, 35,
40, 65, 79, 80, 96, 125, 141,
142, 166

Index
  

infrastructure, 26, 53, 57, 63, 76, 82,
83, 103, 119, 151, 162, 163
Internet of Things (IoT), 55, 65–7,
72, 87, 91, 97
IoT. See Internet of Things (IoT)

179

NLP. See Natural Language Processing
(NLP)
NLU. See Natural Language
Understanding (NLU)
North Face, 77
NVidia, 17

L

London School of Economics, 26
M

machine learning, 3, 7, 8, 13–15,
19–22, 25, 29, 32, 39, 41, 43,
54, 62, 85, 114, 124, 130–2,
140, 164, 169, 170
machine translation, 40
Maturity Matrix, AI, 96, 97, 99,
100, 105
McKesson, 81
Minimum Viable Product (MVP),
122, 123, 149
MIT, 5
MNIST, 15
MVP (see Minimum Viable Product
(MVP))
N

naivety, 47, 48, 136, 142, 168
narrow artificial intelligence, 4, 23, 70
Natural Language Processing (NLP),
35, 51, 54, 64, 107, 167
Natural Language Understanding
(NLU), 2–4, 25, 31, 33–5,
38–41, 51, 54, 56, 57, 64, 65,
74, 76, 78, 81, 82, 87, 114, 167
Netflix, 76, 96
neural networks, 8, 11, 13, 21,
22, 50, 169
neurons, 8, 20
Nexar, 80

O

opaqueness, 23, 47, 132, 137
optimisation, 3, 4, 9, 19, 25, 30,
31, 41–6, 51, 63–6, 74, 79,
86, 93, 95, 97, 102, 108,
132, 140, 163, 168
Otto, 85
P

pilots, 12, 92, 122, 123, 149, 152, 153
Pinsent Masons, 82
planning, 3, 5, 41, 46, 51, 79, 107,
132, 148, 159
platform, AI, 18, 117–21, 140
PoC. See proof of concept (PoC)
prediction, 3, 4, 7, 31, 37, 46–8, 51,
54, 62, 83, 84, 95, 108, 130,
132, 165, 168, 169
processors, 16, 17, 19, 20, 22, 25, 30,
57, 140, 166, 170
proof of concept (PoC), 86, 91, 92,
115, 122, 123, 149, 153
prototypes, 89, 92, 111, 112, 117,
122–4, 147, 149, 174, 175
R

Rainbird, 114, 115
RAT. See Riskiest assumption test
(RAT)
reasoning, 95
reinforcement learning, 8, 42, 44,
45, 168

180

Index

ReThink Robotics, 64
riskiest assumption test (RAT),
122–4, 149
risks, 1, 2, 26, 27, 33, 39, 48, 53, 54,
60–2, 65, 73, 75, 79, 81, 84, 86,
87, 93, 95, 98, 103, 106, 107,
118, 123–5, 129, 130, 134, 138,
139, 142, 143, 145, 147, 153,
154, 157, 162, 165, 168, 174
roadmap, AI, 9, 92, 111–13, 136, 159
robotic process automation (RPA), 8,
27, 35, 36, 52, 55, 58–63, 76,
79, 91, 93, 95, 97, 98, 100, 102,
107, 108, 119, 131, 149, 150
robotics, 60, 64, 65, 83, 97
robust, 152
ROSS, 25
Royal Bank of Scotland (RBS), 75, 77
RPA. See robotic process automation
(RPA)
Rue La La, 85
S

Satalia, 144
search, 3, 7, 13, 14, 25, 26, 32,
35, 36, 38, 46, 59, 62, 79,
81, 82, 93, 95, 97, 106,
107, 142, 166, 167
SEB, 75, 76, 83
selection, 19, 47, 77, 118, 141,
148, 149, 152, 154–8
self-service, 63, 93, 95, 102, 108
Sensai, 126
service providers, 149, 152, 155–8
Siri, 2, 17, 27, 39
Slaughter & May, 82
speech recognition, 2–4, 7, 15, 18,
31, 33–5, 38, 39, 51, 56, 64, 65,
81, 95, 102, 107, 108, 167, 172
Stanford, 19, 64

Stanford Artificial Intelligence Lab
(SAIL), 11
Staples, 76
Stitch Fix, 77
storage, 16, 19, 22, 25, 32, 56, 140,
151, 166, 170
strategy
AI, 9, 91, 92, 96, 100, 110, 112,
113, 117, 120, 121, 125, 126,
139, 147, 148, 174
automation, 91–4, 99, 105, 123,
149, 156, 158
business, 91–4, 100, 112, 113, 127,
149, 162
supervised learning, 7, 20, 32, 34, 35,
39, 44, 45, 51, 68
T

Telefonica, 163, 164
TeliaSonera, 83
training, data, 7, 14, 15, 18, 23, 30,
34, 47, 68, 75, 86, 123–5, 131
transparency, 8, 27, 87, 132, 135, 138,
144, 168, 170, 171
U

UC San Diego Health, 81
UK Meteorological Office, 82
Under Armour, 78
understanding, 3, 4, 9, 11, 23, 29–31,
40, 41, 47–50, 71, 74, 83, 87,
88, 94–6, 101, 102, 110, 111,
123–5, 132–4, 136–8, 147, 153,
162, 169, 173
unintended bias, 47, 48, 130, 134–6,
144
unsupervised learning, 7, 20, 29, 30,
32, 37
USAA, 84, 86

Index
  
V

W

vendors, 2, 9, 19, 36, 52, 70–2, 89, 96,
114, 117–20, 131, 148–50,
152–9, 167, 172, 174
Virgin Trains, 78, 88

Watson, 18, 19, 25, 26, 32,
33, 75–7, 80
Winter, AI, 11–13, 17, 19,
22, 23, 126

181



---

Source Exif Data:

File Type                       : PDF
File Type Extension             : pdf
MIME Type                       : application/pdf
PDF Version                     : 1.4
Linearized                      : Yes
Language                        : EN
XMP Toolkit                     : Adobe XMP Core 5.2-c001 63.143651, 2012/04/05-09:01:49
Create Date                     : 2017:11:06 15:36:25+05:30
Creator Tool                    : Adobe InDesign CC 2017 (Windows)
Modify Date                     : 2017:11:07 23:01:24+05:30
Metadata Date                   : 2017:11:07 23:01:24+05:30
Producer                        : Adobe PDF Library 15.0
Format                          : application/pdf
Document ID                     : uuid:66aea5e3-3217-43c9-adcc-dc9f4b804af7
Instance ID                     : uuid:bae45cd8-e7f7-4747-809b-1d36a6a9dbdd
Rendition Class                 : default
Version ID                      : 1
History Action                  : converted, converted
History Instance ID             : uuid:564aea4f-253c-4f25-8d4f-594152f31437, uuid:ecd8a48e-bbb9-4733-a0c9-37b482257c58
History Parameters              : converted to PDF/A-2b, converted to PDF/A-2b
History Software Agent          : pdfToolbox, pdfToolbox
History When                    : 2017:11:06 15:41:59+05:30, 2017:11:07 23:01:23+05:30
Part                            : 2
Conformance                     : B
Schemas Namespace URI           : http://ns.adobe.com/pdf/1.3/
Schemas Prefix                  : pdf
Schemas Schema                  : Adobe PDF Schema
Schemas Property Category       : internal
Schemas Property Description    : A name object indicating whether the document has been modified to include trapping information
Schemas Property Name           : Trapped
Schemas Property Value Type     : Text
Page Layout                     : SinglePage
Page Mode                       : UseOutlines
Page Count                      : 187
Creator                         : Adobe InDesign CC 2017 (Windows)

EXIF Metadata provided by EXIF.tools