The Business of
Natural Language
Computing
A Primer on Chatbots and Voicebots
JULY 2018
Julian Harris and Mick Endsor
with James Kingston, Milan Sanchania,
Archie Muirhead and Matthew Miller
2
| The Business of Natural Language Computing
Keep up-to-date on chatbots: make sure you're a Pro subscriber bit.ly/cx-chatbots
20180724-1532
Table of Contents
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
The CognitionX Research Subscription: up-to-date, effortless insight from the experts . . . . . . . . . . . . . . . . . . . . 3
About CognitionX . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Why you should use this primer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
Introduction: Conversational Computing: The Old New Frontier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
The Star Trek Reality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
Voice computing: the old new frontier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
It's still early days . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
The future: omnipresent voice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
CHATBOTS TODAY: ABOUT TO TAKE OFF. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7
What is a chatbot? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
The instant messaging tsunami . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
Chatbots are pretty hyped up still . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .10
Taskbots vs socialbots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
Mitsuku . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
Microsoft China XiaoIce ("Shiao-ICE") . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .12
What about the Amazon Alexa Challenge? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .12
CHATBOT DEPLOYMENT: FROM MINUTES TO YEARS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
How do chatbots interact with users? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
There are 25+ chatbot user platforms. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .15
Deploying chatbots: buy or build? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .16
End to end: turn-key business solutions for specific use cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
Conversations: dialog flows and memory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
Understanding: intent, topics, sentiment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .18
Parsing: nouns, verbs, parts of speech, word roots, grammar, and entity naming . . . . . . . . . . . . . . . . . . . .18
Patterns: machine learning, Finding patterns from example data, in new data . . . . . . . . . . . . . . . . . . . . . . .18
IMPACT SO FAR: 10 INDUSTRIES, 4 CORPORATE FUNCTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
Chatbots in business survey: big growth plans . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
Most chatbot deployment plans are in next year or so . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
Most chatbot plans expect substantial investment, and impact . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
Use of natural language is expected to grow substantially . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
Chatbots in China . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
WeChat: the dominant platform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
Xiaoi: the dominant chatbot provider . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
Microsoft XiaoIce: the dominant socialbot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
Chatbots are positively affecting most industries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
Education . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
Finance and insurance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
Government and public sector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Health, fitness and wellbeing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
Hospitality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
Real estate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
Retail and e-commerce . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
Sport, media and entertainment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
Telecommunications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
Travel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .31
There are 4 main corporate functions mostly affected by chatbots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
Customer servicing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
Customer servicing example: Autodesk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
Human resources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
Marketing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
Sales . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
THE FUTURE OF CHATBOTS: RICHER AND MORE PERVASIVE. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
The Future of Chatbots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
Speech is ubiquitous in humans . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
The autonomous assistive agent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
Higher ACE factors: better, richer, more emotional conversations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
NLU everywhere . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
Mixed mode conversations will have its day . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
Some businesses will be 100% chatbots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
Socialbots will become our concierges: for better or worse . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3
| The Business of Natural Language Computing
Keep up-to-date on chatbots: make sure you're a Pro subscriber bit.ly/cx-chatbots
20180724-1532
Introduction
The CognitionX Research Subscription:
up-to-date, effortless insight from the experts
The CognitionX Research Subscription is a new way to be informed and keep up to
date on developments in AI . For each of our growing portfolio of AI research areas,
we provide a uniquely valuable combination of resources .
Free for everyone:

News & article highlights: our experts cut through the noise to ensure you
have just a handful of the top articles every day .

Product directory: instant access to AI vendor profiles, product information
with related news . Comes with alerts too .

Expert network: connect to AI experts in real time to answer your questions .
For deeper insights, we offer a Pro subscription:
1. Primer: get your bearings with our regularly-updated overview of the land-
scape .
2. Case studies: CognitionX-exclusive case studies the real-world business im-
pact of client deployments for a range of industries, corporate functions and
use cases .
3. Insight reports: identifies themes, patterns and important topics that are hard
to see without the big picture, bringing together case studies and other re-
search for easily digestible insights to help inform your strategy and decision-
making .
4. Events: special deals on face-to-face events for conversations with experi-
enced practitioners digging into highly relevant issues, including CogX, our
annual festival of AI . Where events are recorded, they will be made available
to pro members .
About CognitionX
Founded in 2015 by Charlie Muirhead and Tabitha Goldstaub, CognitionX's
mission is to accelerate the adoption of AI across all organisations, and help
ensure a safe and responsible transition to an AI-driven society . We plan to do
this by providing a range of ways to gain access to AI insights and expertise,
including topic-based research programmes and an expert network that pro-
vides a new way for experts to earn a living .
4
| The Business of Natural Language Computing
Keep up-to-date on chatbots: make sure you're a Pro subscriber bit.ly/cx-chatbots
20180724-1532
Why you should use this primer
I've been fortunate enough to start my career in
the early 1990s, and participate directly in multiple
waves of technology disruption. This has included:


The internet itself, bringing the world of in-
formation together,

Social computing: bringing people together,


Mobile & touch computing: enabling the al-
ways-on society, and

Cloud computing, bringing unlimited cheap
computing resources at the click of a button .
Cloud computing has then enabled what is an outright silly number of further disrup-
tions, including of course artificial intelligence. I've spent many years at Google, and
in emerging technology in general, and this disruption is not going to stop . But how
as a business do you navigate this disruption? Where do you even start?
Firstly, choose your time-frame: 6 months, 12-18 months, or beyond? Chatbots are
affecting every aspect of how businesses communicate with people . In the coming
years we are expecting impact to increase as natural language technology improves .
Secondly, find a way to keep up to date. I learned a lot about what to keep in-house
and what to outsource . I broke down emerging tech evaluation into 3 stages:
1. Stay informed: knowing the landscape, and keeping up to date
2. Experiment: choosing and running experiments
3. Integrate: Integrating insights from outcomes into normal business
Knowing the landscape and keep-
ing up to date is time-consuming,
stressful and costly . I would have
personally loved to have used a
trusted research resource that dug
into the topics in sufficient depth to
move quickly to experimentation . Alas the options I found were either too high-level,
too slow, or too expensive .
This primer is a way to understand the chatbot business solution landscape:

How chatbots work today

The impact chatbots are making, and
What we expect the future to hold .
Our research subscription then picks up where the primer starts: a vehicle to keep
you up to date on chatbots, including case studies and insight reports . We hope you
find this primer and our companion research subscription uniquely valuable .
Julian Harris
Head of Technology Research, CognitionX
Chatbots are affecting every aspect of how
businesses communicate with people .
5
| The Business of Natural Language Computing
Keep up-to-date on chatbots: make sure you're a Pro subscriber bit.ly/cx-chatbots
20180724-1532
Introduction: Conversational Computing:
The Old New Frontier
Since we were children we've learned to use our
voices to talk with other people: to ask questions
or to ask for things, to understand, to connect, and
to learn and grow . Our voices are the most natu-
ral, versatile and universal way of communicating .
People have also known for decades that this is
the most natural way to interact with computers as
we've seen in science fiction . The Star Trek comput-
er for example, listens, understands and converses
with natural human speech . What has prevented
this being reality until now is the state-of-the-art of
computer technology . Keyboards, mice, windows and menus have been invented
in the meantime to get useful work done, but keyboards alone require hundreds
of hours of training to even approach half the speed of human speech1 and while
standards help, each application, web site and tool needs to be learned separately .
The Star Trek Reality
I'm particularly proud to have led the team that is helping make the Star Trek com-
puter a reality today . My company, Evi, developed deep tech for understanding lan-
guage and automatically answering questions and built a successful voice assistant .
We were then acquired by Amazon where we then continued to work hard on what
became Alexa and the Echo device . For me this was a ten year journey . Now, a nor-
mal day for tens of millions of families includes talking to items in the home that un-
derstand normal speech, and which speak back just like another person . For young
children, this is happening years before they start using a keyboard or mouse .
Voice computing: the old new frontier
So this voice frontier is very familiar, like an old companion that's been with us for all
of our lives, and yet we are just crossing it again with technology, so it's also new . It's
taken decades to overcome the technical problems to today's magical point where
we can build a product that you can speak to, and that understands you and that can
do useful things in our daily lives . This voice frontier is fundamentally changing how
we think of computers, including bringing the opportunity to retire old ways where
using voice is an easier and simpler alternative .
It's still early days
Of course, it's also early days and there are many gaps in the current products which
are being filled by thousands of engineers working in the big technology companies .
I thus expect to see constant improvements with the products we see today .
1
100 wpm is considered expert typing speed while normal human speech is 185 wpm
6
| The Business of Natural Language Computing
Keep up-to-date on chatbots: make sure you're a Pro subscriber bit.ly/cx-chatbots
20180724-1532
The future: omnipresent voice
Further into the future, I see a world where everything that you can currently do with
technology, every bit of information that you can currently retrieve from the internet,
every action that you can do with an app, every action that you can do online, is ac-
cessible through an omnipresent voice interface everyone knows how to use with
no training required . Everything will be controllable by voice and while you'll often
have the option of an alternative interface, the voice interface will be always acces-
sible and often the easiest, natural choice .
Our eyes are a natural way for consuming vast amounts of complex information quick-
ly, so screens will still have their place as a supporting role where large amounts of
information need to be presented . Text conversations ("instant messaging") are very
popular today and usage will continue to have a prominent role as a private, per-
sistent way of holding conversations that also sometimes benefit from a mix of other
elements . However in many areas I expect to see text replaced by voice as social
contracts evolve and speaking to technology becomes more normal and acceptable
and as the technology improves further . But critically, just as very complicated inter-
actions can be done on the telephone today, in the future, a voice-only conversation
will still exist as an option, particularly where no screen is available .
William Tunstall-Pedoe
Founder of Evi, Creator of Alexa
Chatbots today:
about to take off .
8
| The Business of Natural Language Computing
Keep up-to-date on chatbots: make sure you're a Pro subscriber bit.ly/cx-chatbots
20180724-1532
What is a chatbot?
The term "chatbot" has its origins as "chatterbots", coined by Michael Maudlin in
19942, using the common contraction of "bot" for "software robots" . In this guide
we use "chatbot" to describe messaging with computers, be it text, voice, or other .
Chatbots are sometimes called virtual assistants or virtual agents, but particularly in
recent years, this more colloquial term has pulled ahead3 ("voicebot" is still insignif-
icant):
Google trends: chatbot in blue, "virtual assistant" in red, "virtual agent" in yellow4
The most interesting chatbots use natural language processing (NLP) to have con-
versations with people in plain language, which we describe this as "conversational
computing" . ELIZA (1966), was the first program we'd today recognise as a chatbot:
its apparent ability to understand and reason was an illusion that helped inspire ex-
tensive investment in building thinking and reasoning machines .
Indeed for the rest of the 60s and
70s, intelligent machines were be-
lieved possible "within the next few
years" . Frustratingly, no meaningful
progress over several decades end-
ed in widespread defunding of AI
research in the early 80s, entering a
period dubbed the "AI winter" .5 While the designs were promising, the main problem
was woefully inadequate computing resources . Jeff Dean, Senior Fellow at Google,
underscored the enormity of the problem with his experience at the end of the 80s
as a computer science undergraduate:
"I thought maybe if we could get like a 60x speed up on a 64 processor machine
we could do much bigger problems. And so I worked on some algorithms for
that but as it turned out we needed like a million times more compute not 60x."
Jeff Dean on TWiMLAI6
2
http://bit .ly/2L0LCbi
3
http://bit .ly/2NW74f5
4
ibid .
5
http://bit .ly/2NjOEUw
6
http://bit .ly/2LlY1CN
In this guide we use "chatbot" to describe
instant messaging with computers (voice
or text) .
9
| The Business of Natural Language Computing
Keep up-to-date on chatbots: make sure you're a Pro subscriber bit.ly/cx-chatbots
20180724-1532
Today we are firmly out of this AI winter . The processing power available on a sin-
gle mobile phone today is more than the total global computing power available in
1971 .7 There are tremendous opportunities using natural language communication
with computers using text and voice, but the chatbot landscape today is mostly text
and button-based, and therefore is the primary subject of this version of the primer .
The instant messaging tsunami
How do we know if chatbots are just
hype or something worth investigat-
ing? What are the forces at play that
lead us to be convinced one way or
the other? One signal is what is hap-
pening with text communication,
often called "Instant Messaging"
(or "IM") .
Instant messaging is a huge phenomenon and it's not going away . In 2016, for the
first time, more people used messaging applications than social media .8 It's estimat-
ed that over 5 billion people send more than 100bn messages a day including SMS9,
WhatsApp, Weibo, WeChat Messenger and others, with both numbers not showing
any signs of slowing .
All of these these channels either currently or are expected shortly to offer ways of
supporting chatbots, which represents a huge opportunity for chatbot solution pro-
viders . Indeed there are already over 500,000 chatbots, and we can expect most
IT professional services providers add chatbots to their portfolio of offerings soon, if
they are not already doing so .
7
Source: CognitionX internal research
8
http://bit .ly/2L22hLi
9
http://bit .ly/2zGQNIb
Instant messaging is huge, and not going
away . 100 bn+ instant messages are sent
every day by most people on the planet,
and this trend is growing .
Messaging users
5bn+
a day
Chatbot
builder
products
1000+
Chatbots and instant messaging in context: July 2018
Messages
100bn+
a day
Chatbot solution
provider orgs
10k+
Number of chatbots
500k+
10
| The Business of Natural Language Computing
Keep up-to-date on chatbots: make sure you're a Pro subscriber bit.ly/cx-chatbots
20180724-1532
These are all growing as interest
in and deployment of chatbots
continues .
The overwhelming majority of these
chatbots that exist today are text-
based but voice is growing too:
Alexa claims to have surpassed 30,000 skills, and almost 25% of online shoppers,
for example, use voice assistants on mobile to shop .10
Chatbots are pretty hyped up still .
Currently, there are hundreds of thousands of chatbots and thousands of chatbot
solution providers . We expect to see 1,000+ chatbot builder products servicing spe-
cific geographies, industries and use cases in the coming years .
There's a popular communication tool called the Gartner Hype Cycle that attempts
to convey the impact of a hype bubble: a new technology emerges, thousands of
possible applications are postulated, startups are created, and most of them fail,
which itself then makes news . Investment dries up while the proven business cases
emerge . There is no timeframe on the hype cycle, though one comprehensive anal-
ysis showed that it can span years or decades11 .
Chatbot hype is't over, but close.
Chatbots
Hype cycle by Gartner, chatbot placement by CognitionX.
Image from Salesforce Ben12
Our analysis suggests that as of July 2018, chatbots are close to the peak of expec-
tations . Increasing numbers of case studies of successful chatbot deployments are
a great sign, and matched by still considerable hype articles talking about how chat-
bots "could" impact an area and how they "will" be a disruptive technology . This is
reflected in forecast growth in the chatbot landscape: a range of forecasting publica-
tions suggest between 25-35% growth over the next 5 years . Grand View Research
10 http://bit .ly/2urEjiy
11
http://bit .ly/2mnsVQo
12 http://bit .ly/2LpAdOa
The processing power available on a mo-
bile today is more than the total global
computing power available in 1971 .
11
| The Business of Natural Language Computing
Keep up-to-date on chatbots: make sure you're a Pro subscriber bit.ly/cx-chatbots
20180724-1532
forecasts growth of 24 .3%13, Orbis forecasts 37 .1%14 and MarketsandMarkets estimate
growth of 35 .2% to $3,2bn by 2021 .15 This is driving hype with Gartner predicting that
25% of companies will deploy a chatbot by 2020, up from 3% in 2017 .16
Taskbots vs socialbots
Taskbots vs social companion bots
Taskbots
Narrow focus
Efficient as possible
Outcome-driven
500,000+
Socialbots
Broad focus.
Build long-term empathy, trust and
companionship
Small numbers of bots, hundreds of
millions of users in Asia
Future growth
Who better to ask about what to
do than a trusted companion?
Ruuh
XiaoIce
Mitsuku
Rinna
Jessie
Most chatbots you'll be familiar with are what CognitionX like to call "taskbots" .
They're designed to be highly functional: complete a task efficiently and effectively .
The designers optimise for this, and smart designers identify when to apply natural
language and when to apply traditional user experience elements to achieve this
goal . For instance, a date picker requires more effort than typing a date key-by-key,
but if your journey involves substantial uncertainty around dates, a short conversa-
tion exchange might make the most sense .
Socialbots however, are a completely different kettle of fish . They are the purest
form of chatbots, focusing first and foremost on long-term empathy, trust and com-
panionship . A few examples are provided below .
Mitsuku
Owned and available as a module in the Pandorabots chatbot builder, Mitsuku is an
English-language text chatbot that over many years has built a comprehensive con-
versational vocabulary along with increasingly sophisticated building blocks, such
as its common sense module17 . Mitsuku's owners remain skeptical over the value of
deep learning for intelligence and use conversation log review as substantial inspira-
tion for how to develop Mitsuku's conversation prowess . Mitsuku has won a number
of awards18 including the Loebner award in 2017 for "most lifelike chatbot" .
13
http://bit .ly/2LfQeJL
14 http://bit .ly/2LfGW0f
15 http://bit .ly/2mizTWC
16 https://gtnr .it/2NVeSOo
17
http://bit .ly/2LpwTmi
18 http://bit .ly/2mn6ASX
12
| The Business of Natural Language Computing
Keep up-to-date on chatbots: make sure you're a Pro subscriber bit.ly/cx-chatbots
20180724-1532
Microsoft China XiaoIce ("Shiao-ICE")
On the other end of the ring is a deep learning solution by Microsoft drawing from
billions of previous conversations . For more on XiaoIce see the Chatbots in China
section .
What about the Amazon Alexa Challenge?
This was a deep learning initiative by Amazon to accelerate the development of so-
cialbots with Alexa . The goal was for the judges to hold a 20 minute free-form voice
chat with Alexa . There were around 20 US university contestants and the 2017 win-
ner held their own for 17 minutes . Amazon shared their insights from the challenge,19
including some data (40,000 hours of audio, and millions of conversations) . The only
criticism for the competition is that it's been restricted to US universities .
19 http://bit .ly/2LoQ2Vn
Chatbot deployment:
from minutes to years
14
| The Business of Natural Language Computing
Keep up-to-date on chatbots: make sure you're a Pro subscriber bit.ly/cx-chatbots
20180724-1532
How do chatbots interact with users?
We have identified five different messaging models involving chatbots including sim-
ple user-to-user instant messaging .
Customer to Staff: live chat . The staff member may be part of a team; the staff mem-
ber may have some predefined text to return . Examples here include LivePerson and
Intercom .
Customer to Bot: the only interaction a customer has is with a chatbot . Those who
can deliver this for their use cases will find this scales the best .
Human First vs Bot first: human first systems are similar to customer-to-staff but
may hand off to a bot for routine inquiries (information retrieval) . Bot-first systems
attempt to hold conversations and hand off to staff when appropriate .
Customer to Staff with Bot oversight: conversations are monitored by a bot and
feedback is provided either real-time or in bulk later to provide feedback on how to
communicate more effectively . Examples of this include Ixy20, Cogito21 and Daisee22 .
Customer to Bot to customer: brokered chat . Bots field questions and hand off to
the right person at the right time . Examples include Koko23 and our very own Cogni-
tionX Expert Network chatbot24 .
Customer to Bot
Customer to Staff
Human first: Customer to Staff then Bot
Bot first: Customer to Bot, then Staff (24/7)
Customer to Bot to Customer (anonymising P2P broker)
Customer to Staff with Bot oversight
(real time coaching / offline reporting)
Chatbot
interaction
models
20 http://bit .ly/2utw4Ck
21 http://bit .ly/2mmaHOX
22 http://bit .ly/2utwxEA
23 http://bit .ly/2NVV2T9
24 http://bit .ly/2L43AJG
15
| The Business of Natural Language Computing
Keep up-to-date on chatbots: make sure you're a Pro subscriber bit.ly/cx-chatbots
20180724-1532
There are 25+ chatbot user platforms .
The chatbot end-user platforms
July 2018
Skype
Twitter
LINE
Weibo
WeChat
Viber
Telegram
Facebook
Messenger
iMessage
Slack
Workplace by Facebook
Microsoft Teams
Google Assistant
Kik Messenger
Amazon Alexa
Cisco Spark
Cortana
Discord
Mycroft
Samsung
Bixby
Business Chat
by
Chatbots are deployed on one or more of over 25 platforms, some of which are large
and well-known, including Facebook Messenger, WeChat, Alexa, Google Assistant,
Skype and Slack . As of publication WhatsApp does not yet provide an official chat-
bot API but it is expected that one will be launched this year .
Some companies offer both consumer and business-focused platforms . For exam-
ple, alongside Messenger, its consumer platform, Facebook also offers Facebook
Workplace as a business platform .
Websites and Facebook Messenger are the primary deployment plans for respon-
dents to the CognitionX chatbot survey . However, in the next 12 months there is
strong interest in expanding deployment across native apps and voice channels .
16
| The Business of Natural Language Computing
Keep up-to-date on chatbots: make sure you're a Pro subscriber bit.ly/cx-chatbots
20180724-1532
Today
38%
18%
44%
0
0
18%
Next 12 months
33%
45%
39%
39%
32%
24%
Interested but no firm timeframes 29%
37%
17%
61%
68%
58%
Source: CognitionX internal research
Deploying chatbots: buy or build?
Another extraordinary area of growth and richness is a bustling ecosystem of tools
and technologies used to build chatbots that can hard to make sense of .

Does a tool create an actual chatbot or is it just a design tool?

Can you customise the conversation flow?

Does it even do conversations?

What if I'm not happy with the conversation tools: what are my options? For
testing too?

How do I track and improve performance when live?

What do I do about data? What do I do when I move from the impressive
proof-of-concept that wowed the board to a real-world system that needs
to respect the specific manner of conversing of our customers, and the lan-
guage of our products? This is actually a new concept for software develop-
ment: that the data is so core, and so sensitive to the experience even for
traditionally non-data-driven applications .
To help answer these questions we've carved the chatbot solution space into 3 main
categories of solutions that attempt to answer these questions:

Design: tools to help rapidly (in minutes or hours) test out ideas .

Development: tools, technologies and frameworks to aid chatbot develop-
ment .

Optimisation: tools for testing, enriching and analysing chatbot performance .
How much do you build and how much do you buy? Particularly in this Age of Cloud
there has never been a richer and easier palette of options to choose from and
chatbots are no different . We've split out the development stack specifically into five
17
| The Business of Natural Language Computing
Keep up-to-date on chatbots: make sure you're a Pro subscriber bit.ly/cx-chatbots
20180724-1532
levels, guiding the balance between implementation speed over flexibility (i .e . more
bespoke work) .
Over the past year we've amassed the largest catalogue of chatbot builder technol-
ogies that we're aware of . With over 500 identified, we've attempted to show with
diagram below a sample of the products in each category .
Products are placed approximately in the area that we think the technology brings
the greatest value, but it is an incomplete picture . For example Rasa .ai offers a stand-
alone "Understanding" component (Rasa NLU), and Hashblu offers design features .
DEVELOPMENT
Patterns
Understanding
Conversations
End-to-end
Parsing
DESIGN
OPTIMISATION
The chatbot technology solution landscape
IBM Watson
Assistant
Amazon
SageMaker
Amazon
Lex
Humanise
Finstreet
App2Check
eXvisory
NLTK
qbox
End-to-e
nd
ns
ElseNet
SyntaxNet
Intelligent Bots
Qbox
by
Botframe
BotPreview
Apache
Apache
Stanford NLP
AI Platform
Turn-key business
solutions
Bespoke conversation
flows & memory
Intent, topics, sentiment
Nouns, verbs, parts of speech, word
roots, grammar, entity naming
Machine learning. Finding patterns
from example data in new data
We will be publishing the chatbot builder technology database
on directory.cognitionx.com in 2018.
End to end: turn-key business solutions for specific use cases
What is your specific use case? Is there in fact a solution out there already being
addressed? This is where we expect to see the most growth in the coming years .
Considering geography, industry and use case we anticipate literally thousands of
end-to-end products to appear, to then be followed by consolidation (see the trajec-
tories section) .
Conversations: dialog flows and memory
There are a few reasons why you may not want an end-to-end solution: the most
common scenarios are where:

The tools don't quite meet your need (right use case, insufficient language
support is common)

You want your chatbot's conversation mechanism to be differentiator and
want more control over the roadmap and retain core intellectual property, so
want to build it yourself
Most of the tech giants have a conversation development offering, except notably
Microsoft, that has left this to third party developers such as DF2020 Chatbot Au-
18
| The Business of Natural Language Computing
Keep up-to-date on chatbots: make sure you're a Pro subscriber bit.ly/cx-chatbots
20180724-1532
thor25, and Apple, that has opted instead to
offer a human-first live chat offering, in part-
nership with the major live chat platforms .
Most chatbot conversation development
tools today make it easy to design specific
conversation flows: if this, then that . Some
argue that this becomes complex and costly
to maintain, and that new frontiers in more
flexible conversation management (such as
the likes of Rasa .ai, Action .ai and Poly-AI)
presenting alternatives to this mechanism
and promise to offer more flexible conver-
sation flows .
Understanding: intent, topics,
sentiment
The next level of flexibility is the set of tools that we call "Natural Language Under-
standing" (or NLU) . These tools do not have memory and are designed specifically
to take natural language text and give back meaningful information: what was the
intent of the sentence or phrase (e .g . "find a product")? Was it good or bad sentiment
and why? What is the topic being covered?
NLU services are useful if you can provide value with no memory of past conversa-
tions . It's a pretty popular additional offering, for instance, to offer a natural language
experience for a company's frequently asked questions (FAQs) .
The companion to NLU is natural language generation (NLG): natural-sounding re-
sponses from the chatbot .
Parsing: nouns, verbs, parts of speech, word roots,
grammar, and entity naming
Intents are part of NLU, tied to a specific domain or use case . However practitioners
are finding that they don't always work well some systems are exploring methods
that discard the concept of intents altogether (e .g Rasa NLU 0 .12 and Poly-AI) . This is
looking to be more promising for cross-language support but out-of-the-box training
sets are less effective resulting in a lot more training required .
Patterns: machine learning, Finding patterns from
example data, in new data
As mentioned in the introduction, a key enabler in the last 5 years for natural lan-
guage computing has been the extraordinary resources now available for machine
learning: in the cloud, which has essentially unlimited memory and storage, along
with 1 million times the processing power available than the late 80s . This has meant
an opportunity to revisit a number of algorithms that despite being decades old,
have only recently been practical to use in daily life . Long-short-term memory (LSTM,
25 http://bit .ly/2mn7vTe
Conversation (level 4)
Flow.ai
Considerations:
When is it appropriate to use
free form text vs buttons and
date pickers?
When is it ok to insist on input
from a user (vs do the best they
can with what you have)
How do you handle multiple
intents?
Example dialog design from flow.ai
19
| The Business of Natural Language Computing
Keep up-to-date on chatbots: make sure you're a Pro subscriber bit.ly/cx-chatbots
20180724-1532
invented by Jrgen Schmidhuber) is one of the most widespread and a key compo-
nent of many natural language processing systems today (voice and text) .
Example: patterns & machine learning
"waistcoat": predicting it's a
noun due to its proximity to
other parts of speech
LSTMvis is an open source tool for diagnosing natural language processing performance
The landscape trajectories
When looking at a landscape we also need to consider trajectory over time: where
will things be in 12 months time? We have observed a number of movements:
Expansion of value up, out, and down: products, once established in one level
of value tend to expand to others . E .g . Microsoft Language Understanding service
(LUIS) is an Understanding service that recently (Q2 2018) added rudiments of mem-
ory, gesturing towards a desire to expand into the Conversations development tier .
Equally a lot of products offer analytics, enrichment and testing . Finally, Conversa-
tion development tools are expanding into use case-specific solutions .
Expansion of use cases: solving one problem well is a great start to a business and
once solved, not unlike most business strategies it's then a common play to expand
to service companion use cases, for example to cover a whole corporate function .
For example, Mya aims to solve a specific use case around CV/resume screening,
while competitor Eva .ai has the ambition to embrace the whole recruitment process
end to end, arguing its data perspective is richer and can provide more value .
Consolidation: there are two main types of companies absorbing chatbot builder
tech: customer relationship management solutions (e .g . Hubspot acquiring Motion .
ai) and tech giants (e .g . Google acquiring API .ai and renaming as DialogFlow) .
Impact so far:
10 industries,
4 corporate functions
21
| The Business of Natural Language Computing
Keep up-to-date on chatbots: make sure you're a Pro subscriber bit.ly/cx-chatbots
20180724-1532
Chatbots in business survey: big growth plans
As part of this primer we did an informal investigation to get a general sense of the
temperature of chatbot plans . For this sample we received 80 responses, around
half from the UK, around one quarter from the USA (SFO / NYC equal split), and small
numbers from Europe and Asia . That said, as responses came in, the distribution of
responses remained relatively consistent . We plan to update this with more compre-
hensive and widespread insights over time as part of the subscription .
As to how indicative the UK + USA is for other regions of the world, we believe the
the regions the surveys mostly came from (San Francisco, New York and London) are
innovators or fast followers, so for other regions the timeframes may be a little more
extended .
Most chatbot deployment plans are in next year or so .
The survey of companies' own chatbot progress and plans reflect a breadth of chat-
bot adoption and the varying stages companies are at in deploying them . At the
furthest stage of chatbot implementation, the survey found that around one fifth of
respondents have fully-launched a chatbot with customers and around 2/3rds expect
to do so within the next 12 months . Only 8% of respondents have no chatbot deploy-
ment plans in place .
Unsurprisingly, more companies are at the proof of concept stage with around 40%
of respondents at the stage reporting that they are currently piloting chatbots with
customers and closer to half planning to do so in the next year or so . One step
behind this, companies are exploring internal testing of chatbots . Around half are
currently doing so and one third plan to do so over the next year or so . Finally, at
the earliest stage of implementation, half have plans in place and around 40% are
looking to put chatbot plans in place, in similar timeframes .
22
| The Business of Natural Language Computing
Keep up-to-date on chatbots: make sure you're a Pro subscriber bit.ly/cx-chatbots
20180724-1532
Most chatbot plans expect substantial investment, and impact .
Over two thirds of survey respon-
dents expect to see a significant
business impact from chatbots
with around one third expecting
to see chatbots transform at least
one major business function . For
approximately 15% of respon-
dents, the potential business im-
pact of chatbots is currently un-
clear but this may change as the
number of successful use cases
continues to increase . Companies
will increasingly move along the path from planned deployment to internal testing,
customer piloting and full deployment with the expectation that they will see a signif-
icant impact on at least one major business function .
Use of natural language is expected to grow substantially .
A lot of value of a chatbot is being delivered successfully through buttons, menus,
and simple keyword matching: not a lot of machine learning . Today's use of natural
language in chatbots is typically less than 20% of conversation flow . However, there
is a strong appetite for flipping this on its head with the vast majority expecting this
percentage to be greater than 80% in future .
We think this is the right strategy: natural language communication is very easy to
get wrong; starting with carefully-scoped use cases to gain institutional knowledge
and offer a smooth transition for customers can help avoid being one of the fairly
common high profile embarrassments (such as a transport company in Australia re-
cently not understanding common place names) .
23
| The Business of Natural Language Computing
Keep up-to-date on chatbots: make sure you're a Pro subscriber bit.ly/cx-chatbots
20180724-1532
Chatbots in China
When talking about chatbots, it's important to take a global perspective, particularly
as many parts of Asia have, through necessity, been much more advanced in their
smart phone and chatbot development . In particular we've chosen to spotlight China
as the volume and development are well worth watching . We'll be including a China
perspective regularly in our research . With the Chinese smart home industry expect-
ed to reach $23bn by 201826, voice capability will be of increasing importance to the
Chinese chatbot landscape .
WeChat: the dominant platform .
China's conversational computing ecosystem has been marked by early chatbot
adoption, the structural relationships of China's internet giants, cultural characteris-
tics of Chinese consumers, and practical necessities of having over 200 base char-
acters in their writing system . The goliath of Chinese social media, WeChat, has
been using chatbots since 2013 . Attracting 1 billion monthly users, WeChat is of-
ten compared to Facebook Messenger, and indeed Facebook is looking to emulate
WeChat's "super app" model which integrates a huge range of third party services
where essentially the whole internet is redeployed insided WeChat27 . This is every
company's golden scenario -- Western providers in the 90s such as America Online,
MSN and Prodigy had such plans before the web took off . WeChat first began as a
social messaging platform, and now offers a range of sticky features such as seam-
less business and payment integrations allowing users order food, taxis, buy tickets,
open bank accounts, transfer money, share files, and generally interact with brands
through the app . Further fuelled by the Chinese cultural trait of assessing product/
vendor quality through multiple question, chatbots have proliferated .
Xiaoi: the dominant chatbot provider .
It's not typical to talk about solution providers when talking about the solution land-
scape as there are typically tens of thousands . However in China, Xiaoi (pr . "Shia-
oi", i) is notable in being by far the country's dominant chatbot developer .
Launching in 2004 with an MSN based chatbot, Xiaoi is estimated by one 2017 re-
port as controlling 90% of the chatbot market in China and counts 40 of the top 50
Chinese banks as clients,28 processing about 200 million requests a day .29
26 http://bit .ly/2miJjRY
27 https://tcrn .ch/2JsuFke
28 http://bit .ly/2Js84V9
29 http://bit .ly/2uoOXXm
24
| The Business of Natural Language Computing
Keep up-to-date on chatbots: make sure you're a Pro subscriber bit.ly/cx-chatbots
20180724-1532
Microsoft XiaoIce: the dominant socialbot
Once again it's not typ-
ical to include specific
chatbots when paint-
ing the landscape but
Microsoft XiaoIce (pr .
"Shiaow-ice",
) is striking exam-
ple . With 100m users
in China alone30, 25%
of whom have at one
point said "I love you",
XiaoIce's
conversa-
tion skills have become extremely sophisticated, integrating skills like weather into
the conversation, empathy, and adjusting for tone of voice based on recipient, mem-
ory, supporting text and "full duplex" voice conversations (where speakers slightly
overlap each other) . XiaoIce is more than just a chatbot but a personality, noted for for
the playful, creative nature of its interactions; indeed, it has even generated its own
poetry31 and appeared to read newscasts . The success of XiaoIce and its pan-Asian
siblings (e .g Microsoft Japan's Rinna) inspired Microsoft's launch of English-speak-
ing Tay, a Twitterbot . All the more
flummoxing then was that within 24
hours Tay became a holocaust-de-
nying racist32 . Indeed it appears that
trolling is a distinctly Western phe-
nomenon . Zo (US/English) and Ruuh
(India/English) were launched some
time later, with markedly more mut-
ed imitation inclinations .
30 http://bit .ly/2Lgd4RG possibly the most interesting paper we've read this year
31
http://bit .ly/2NlKJXv
32 https://tcrn .ch/2utt09w
Transcript of conversation with Microsoft Japan's Rinna. It's not
a taskbot so it takes its time chatting before adding a cookie
coupon. This is an extremely powerful trust platform.
With 100m users in China alone, 25% of
whom have at one point said "I love you",
XiaoIce's conversation skills have become
extremely sophisticated .
25
| The Business of Natural Language Computing
Keep up-to-date on chatbots: make sure you're a Pro subscriber bit.ly/cx-chatbots
20180724-1532
Chatbots are positively affecting most industries
Chatbots are affecting any touchpoint where businesses communicate with people .
Ten sectors reveal the widespread potential applications of chatbots:
1. Education
2. Finance and insurance
3. Government and public sector
4. Health, fitness and wellbeing
5. Hospitality
6. Real Estate
7. Retail and e-commerce
8. Sports, media and entertainment
9. Telecommunications
10. Travel
Education
Traditional approaches to teaching and education are being reshaped by chatbots .
The opportunities are considerable with 264m children out of school33 and 758m illit-
erate adults .34 Companies are responding to the opportunity . Already there are over
10,000 education bots deployed on Facebook Messenger .35 Use cases range from
new forms of teaching through to administrative support and student engagement .
Professor Ashok Goel, a professor at Georgia Tech, created Jill Watson, a teaching
assistant, using IBM Watson tools . Using 40,000 questions and answers from four
semesters of teaching, Goel trained Jill Watson to answer student questions . It had
mixed success at first, providing incorrect and "strange" answers but over time Goel
was able to improve Jill, using new questions and memory of previous questions
and answers . By the time Goel deployed it with his students, Jill's answers had an
accuracy of 97% .36
Another chatbot use case is improved student engagement . The University of Ade-
laide has piloted a chatbot deployed on Facebook Messenger to respond to admis-
sions enquiries from prospective students . On its first day, the chatbot responded to
2,100 unique conversations leading to a 40% decrease in calls to the university with
the added benefit of contributing to a fall in telephone hold times from 40 minutes
to 90 seconds .37
Finance and insurance
In finance and insurance, chatbots are enhancing the customer experience in both
online and physical settings . The most prominent use case is customer service where
banks and insurance companies are looking to enhance customer experiences . On-
line, insurance companies are turning to chatbots . Singapore Life developed SingLi-
fe, a machine learning assistant on Facebook Messenger to provide customers with
33 http://bit .ly/2uDHfI6
34 http://bit .ly/2Ni3bQK
35 http://bit .ly/2NluY2D
36 https://read .bi/2KY5mfv
37 http://bit .ly/2LqFPYo
26
| The Business of Natural Language Computing
Keep up-to-date on chatbots: make sure you're a Pro subscriber bit.ly/cx-chatbots
20180724-1532
a simplified experience for life insurance coverage . Allstate, an American insurer,
developed the Allstate Business Insurance Expert (ABIE), an AI chatbot designed
to answer insurance queries from small business owners,38 and Finstreet .co39 is a
Bangkok-based chatbot that is substantially better at generating bank leads than the
banks themselves due to being seen as more trustworthy .
Banks are also deploying chatbots across their operations .40 Customer service is the
most prominent use case with banks forecast to automate up to 80% of customer
queries .41 A survey by CGI found that 58% of customers demand a personalised ser-
vice and 52% would like their banks to monitor their spending and provide savings
advice .42 HiCharlie is a chatbot deployed on Facebook Messenger and over text
messaging which has helped users save on average $80 per week .43
Bank of America has Erica, a voice
and text-enabled chatbot designed
to provide customer service and help
customers make banking decisions .
Capital One has Eno, a text-enabled
chatbot to help customers manage
their money . And the State Bank of
India (SBI) has SBI Intelligent Assistant, that functions like a bank representative to
handle customer queries .44 And chatbots deployed by financial services firms are
not just confined to the online world . HSBC has deployed Pepper, a robot created by
Softbank Robotics, to greet customers in its main New York City bank .45
The sector is still working to understand where chatbots can be deployed success-
fully . Nordnet, a Swedish online bank, has discontinued its use of the Amelia chatbot
developed by IPSoft . Originally intended to improve onboarding processes and cus-
tomer satisfaction, Nordnet has concluded that the customer response is 'ok but not
overwhelming' .46
Government and public sector
Government and the public sector are turning to chatbots to deliver services to cit-
izens . In 2017, the city of Los Angeles launched the City Hall Internet Personality
(CHIP) chatbot in collaboration with Microsoft for the Los Angeles Business Assis-
tance Virtual Network (BAVN) . In its first 24 hours, CHIP answered over 1,400 ques-
tions from 180 users . Since deployment, CHIP has cut the number of emails to BAVN
from 90 to between 30 and 40 per week and has increased its answer knowledge-
base from 200 to 700 questions .47 At a national level, the US Department of Home-
land Security uses Emma, chatbot supporting English and Spanish interaction, which
38 http://bit .ly/2NVhjAy
39 http://bit .ly/2uujOSb
40 http://bit .ly/2Lg2ud6
41
http://bit .ly/2KY5TOx
42 http://bit .ly/2L3m2Ck
43 http://bit .ly/2L254Eg
44 http://bit .ly/2miCjVc
45 https://tcrn .ch/2NhsySC
46 https://read .bi/2Jts21E
47 http://bit .ly/2LfUc57
Charlie is a chatbot deployed on Face-
book Messenger and SMS, helping users
save on average $80 per week .
27
| The Business of Natural Language Computing
Keep up-to-date on chatbots: make sure you're a Pro subscriber bit.ly/cx-chatbots
20180724-1532
handles over 1m monthly interac-
tions .48
In January 2017, the Dubai Electricity
and Water Authority (DEWA) created
RAMMAS, the first ever government
chabot developed on the Google AI
platform . DEWA is available on the
authority's website, mobile (IOS and Android), Alexa, Facebook and as a physical
robot . RAMMAS has answered over 698,000 requests and can also process bill pay-
ments .49
Health, fitness and wellbeing
Chatbots are being tested and deployed for a range of healthcare applications, in-
cluding diagnosis, personalised healthcare management and therapeutic treatment .
In medical diagnosis, Babylon Health tested its chatbot against a set of questions
from the Membership of the Royal College of General Practitioners (MRCGP) test .
Babylon Health claim that it achieved a score of 81% compared with an average
score of 72% for human doctors based on results from 2012 to 2017 .50
Furthermore, chatbots offer the
prospect of personalised health-
care and lifestyle management . For
example, Haptik offers a medicine
reminders chatbot .51 Chatbots are
also being developed for therapeu-
tic treatment . Woebot, a chatbot
grounded in cognitive behavioural
therapy, provides mental health support for individuals .52 In a study of its effective-
ness, researchers at the Stanford School of Medicine in collaboration with Woebot
found that the app significantly reduced the symptoms of depression compared with
a control group .53 Similarly, Tess54, another mental health chatbot, found that interac-
tions with it reduced symptoms of depression by 13% and anxiety by 18% .55
Hospitality
The hospitality industry is being transformed by chatbots . Like retail and e-commerce,
the hospitality industry will look to chatbots to increase customer loyalty through
on-demand support and a more personalised service . In particular, multilingual chat-
bots could converse with guests in their preferred language and reduce the burden
on hotels to hire staff fluent in the languages of all possible guests . Chatbots have
other uses in the hospitality industry . They can be used in the reservation channel,
48 http://bit .ly/2upb3J6
49 http://bit .ly/2KXwZoU
50 https://bbc .in/2Le7Q94
51 http://bit .ly/2ml9SWE
52 http://bit .ly/2zIJ4sT
53 http://bit .ly/2NgV5I3
54 http://bit .ly/2JtyeXh
55 http://bit .ly/2JtyeXh
Tess, another mental health chatbot,
found that interactions with it reduced
symptoms of depression by 13% and anxi-
ety by 18% .
The US Department of Homeland Securi-
ty uses Emma, chatbot supporting English
and Spanish interaction, which handles
over 1m monthly interactions .
28
| The Business of Natural Language Computing
Keep up-to-date on chatbots: make sure you're a Pro subscriber bit.ly/cx-chatbots
20180724-1532
help the industry offer personalised services, support customer engagement and
retention and provide assistance throughout the hospitality process .
Adoption of chatbots in the hospitality sector is global . The Cosmopolitan in Las
Vegas has deployed Rose, ConcierGo is used in The Andaz Singapore and BeBot
from Bespoke Inc is used in the Otani hotel in Tokyo . These are all variants of a chat-
bot concierge that provide on-demand customer assistance including: recommen-
dations for nearby attractions, making restaurant bookings and answering queries
from guests . In 2017, 7% of guests at the Cosmopolitan engaged with Rose and those
who did spent 30% more than those who did not and were 33% happier when they
checked out .56 Similarly, Hashtag Hotels have deployed Humanise .AI's chatbot, ob-
serving early success:57 50% of all guests have been engaging with the new system,
with >10% guests advancing to purchasing items .
Real estate
In real estate, chatbots can be used
to assist in the sales process, pro-
viding instant responses to custom-
er enquiries, pre-qualifying leads,
collect customer information and
schedule property viewings . Chat-
create developed a Facebook Mes-
senger chatbot for a client to gen-
erate and pre-qualify sales leads . In
its first ten days and with an ad spend of 94 EUR, the company achieved 1,064 link
clicks, 243 messaging conversations, 60 pre-qualified leads and 3 apartment reser-
vations .58 Structurally created Aisa Holmes, a personal lead assistant for real estate
brokers . For some users, Aisa Holmes was able to increase lead volume by as much
as 400% and sales by up to 10% while also supporting job growth to convert the ad-
ditional leads generated into sales .59
Retail and e-commerce
Retail and e-commerce stands to benefit greatly from the deployment of chabots to
enhance customer engagement and drive sales . A survey by Mastercard found that
20% of EU customers have shopped using a voice assistant such as Amazon Alexa
or a chatbot . The survey also found that by 2022, the value of "conversational com-
merce" could reach $40 billion or 6% of US online spending .60
Chatbots can also deliver a personalised e-commerce experience for customers .
Companies are already realising these benefits in online sales . For example, Amtrak
deployed Ask Julie, a chatbot developed by Next IT, for booking tickets online and
handling customer queries . Julie answered over 5 million questions in a single year,
saving over $1m in customer service email costs and generating 30% more revenue
56 http://bit .ly/2JxHb28
57 http://bit .ly/2L2PloM
58 http://bit .ly/2uynxxk
59 http://bit .ly/2mkWrGf
60 https://bloom .bg/2NTrnKo
Aisa Holmes was able to increase lead vol-
ume by as much as 400% and sales by up
to 10% while also supporting job growth
to convert the additional leads generated
into sales .
29
| The Business of Natural Language Computing
Keep up-to-date on chatbots: make sure you're a Pro subscriber bit.ly/cx-chatbots
20180724-1532
per booking .61 Similarly, Adidas used a Facebook Messenger chatbot to create an
interactive registration process for its female-focused community space . 2,000 peo-
ple signed up through the chatbot in the first two weeks with repeat use of 80% and
retention at 60% .62
BabyCenter commissioned ubisend to build a chatbot on Facebook Messenger to
improve marketing and sales . Once deployed, BabyCenter saw an average 84%
read rate of automated messages and an average click through rate from the Face-
book messenger chatbot to their website . The chatbot increased engagement by
1,428% compared with industry benchmarks .63 Sephora, the global beauty retailer,
used the Sephora Assistant chatbot to reduce the number of steps required to book
a beauty makeover by five . This resulted in an increase of 11% in the booking rate .64
Sport, media and entertainment
Sports, media and entertainment are huge industries with a core focus on customer
experience and engagement . In September 2017, FC Barcelona launched a chatbot
on Viber, the messaging app, to engage its over 4m followers .65 Dream11, India's
largest sports game, used a chatbot developed by Haptik to respond to over 80% of
the 1m+ customer support queries during the 2018 Indian Premier League tourna-
ment .66 The chatbot was able to resolve customer queries in an average of 32 sec-
onds compared with 4-24 hours required by a human customer service member .
Beyond customer support, chatbots
are also used in the sports, media
and entertainment industries to pro-
vide information and book tickets to
events . New York City's Lincoln Cen-
ter worked with Pypestream to add
Wolfie, a chatbot, to their website .
Customers can ask questions and receive information on over 3,000 annual events
at the Lincoln Center . Wolfie also makes recommendations and books tickets .67 Ten-
nis Australia implemented a chatbot to directly sell tickets for the 2018 Australian
Open via social media . They achieved 170% higher sales conversions compared with
their traditional marketing efforts .68
Telecommunications
Telecommunications companies are using chatbots to improve customer servicing .
Charter Communications implemented Alme in November 2012 to enhance custom-
er support . Charter recognised that 38% of customer requests were for username
and password retrieval, these have been fully automated . Alme also helped reduce
live-chat volume by 83%, delivered 5x return on investment in the first six months,
61 http://bit .ly/2uEfOxN
62 http://bit .ly/2Nlyji2
63 http://bit .ly/2KY6TSN
64 http://bit .ly/2Jp3vLb
65 http://bit .ly/2Ld5sPP
66 http://bit .ly/2NUUSLW
67 http://bit .ly/2NOPI41
68 http://bit .ly/2uqpCfw
Tennis Australia achieved 170% higher
sales conversions compared with their tra-
ditional marketing efforts .
30
| The Business of Natural Language Computing
Keep up-to-date on chatbots: make sure you're a Pro subscriber bit.ly/cx-chatbots
20180724-1532
delivered a 44% cost reduction in the first year and reduced the time taken to re-
set passwords by 50% .69 Beyond customer servicing, telecommunications compa-
nies are experimenting with chatbot companions for users . AT&T launched Atticus, a
chatbot "built to talk entertainment ."70 Intelligent social companion bots represent a
new frontier in chatbot technology .
Sales and marketing is another key function where telecommunications companies
are testing chatbots . In the second half of 2016, CenturyLink implemented Angie, an
automated sales assistant developed by Conversica . By June 2017, Angie was send-
ing 30,000 emails a month, supporting customer representatives who may have up
to 300 accounts . In a small pilot study, Angie was able to understand 99% of emails,
referring the other 1% to a manager . The company has earned $20 in new contracts
for each dollar invested in Angie .71
Travel
The travel sector is adopting chatbots to assist customer bookings, provider custom-
er support and complaints resolution and act as a personal assistant to customise a
traveller's itinerary and experiences . Industry research by SITA in 2017 found that
14% of airlines and 9% of airports currently use chatbots with 68% and 42% respec-
tively forecast to adopt chatbot services by 2020 .72 The Dutch airline KLM has Blue-
Bot (BB) a chatbot on Facebook Messenger that is able to interact with customers in
9 languages .73 In BB's first month, KLM saw the volume of Facebook messages in-
crease by 40% and since it was deployed, over 1 .7m messages have been sent by
more than 500,000 people .74
Changing customer preferences
and behaviour are driving chatbot
adoption in the travel industry . Re-
search of 19,000 travellers in 26
countries by Booking .com found
that 50% do not have a preference
for a human or computer interac-
tion as long as their question is an-
swered and 80% prefer self-service
options . In response, Booking .com developed the Booking Assistant chatbot as an
English-language pilot version in 2017 . By the end of 2017, the Booking Assistant was
available globally to assist in English-language bookings and was able to resolve
30% of customer questions in under five minutes .75
69 http://bit .ly/2Nlifgq
70 https://soc .att .com/2Lfk1SQ
71
http://bit .ly/2NmEGBJ
72 http://bit .ly/2JqMtfK
73 https://klmf .ly/2LfV1uJ
74 http://bit .ly/2zIYTA1
75 https://booki .ng/2mpcVx8
50% of 19,000 travellers surveyed do not
have a preference for a human or comput-
er interaction as long as their question is
answered and 80% prefer self-service op-
tions .
31
| The Business of Natural Language Computing
Keep up-to-date on chatbots: make sure you're a Pro subscriber bit.ly/cx-chatbots
20180724-1532
There are 4 main corporate functions mostly
affected by chatbots
Inside a company, there is the opportunity to beneficially use chatbots whenever a
person has to interact with a computer in any function or industry . This is reflected in
the diverse range of use cases outlined below . Businesses are looking for chatbot
use cases where they can improve engagement and increase satisfaction while re-
ducing costs and response times .
Customer servicing
Companies are finding success in deploying chatbots to build, manage and improve
customer relationships . Globe Telecom used Gie, a hybrid bot developed by Ser-
vicefriend for Facebook Messenger . Gie resulted in a 22% increase in customer sat-
isfaction rate compared with the call centre, a 3 .5X increase in employee productiv-
ity and 50% reduction in calls to the call centre for customers interacting with Gie .76
Similarly,
Swedbank
deployed
Nina, a chatbot developed by Nu-
ance Communications, across its
entire customer base . Within three
months, Nina averaged over 30,000
customer conversations per month,
achieving a first-contact resolution
of 78% .77 This subsequently grew to
over 40,000 customer conversations a month with a resolution rate of 81% .78 In eval-
uating Nina, Swedbank learned that rolling it out to 20% of their customer base as a
proof of concept would have provided a better environment to train Nina .
Customer servicing example: Autodesk
Autodesk is a multinational and cross-industry
software development company . In May 2018
Autodesk upgraded its text-based assistant, Ava,
to one that used Soul Machines virtual employee
technology to offer a 3D virtual character . Speak-
ing at CogX in June 2018,79 Autodesk's Director
of Machine Assistance, Rachel Rekart, highlight-
ed the efficiencies offered by chat technology .
Now used in 85 different use cases, Ava speaks
to 100,000 customers a month - more than Au-
todesk's entire 350 person Customer Support Agency combined . Human agents
can solve approximately 25 customer cases per day - Ava solves 2500 per day with-
out any human intervention required . Average response time for an enquiry with Ava
is 5 minutes; with a human agent, 1 days .
76 http://bit .ly/2L1vAhh
77 http://bit .ly/2ml5DdG
78 http://bit .ly/2urI3R4
79 http://bit .ly/2zO5cCo
Gie resulted in a 22% increase in custom-
er satisfaction rate compared with the call
centre, a 3 .5X increase in employee pro-
ductivity and 50% reduction in calls .
32
| The Business of Natural Language Computing
Keep up-to-date on chatbots: make sure you're a Pro subscriber bit.ly/cx-chatbots
20180724-1532
Notable for its 'humanlike' and ethnically ambiguous avatar, Ava now has voice and
video functionality . This functionality is used to detect emotion on the part of the
user, further increasing Ava's responsiveness .
Human resources
Human resources (HR) is one prominent area where chatbots can be successfully
deployed with a range of possible applications . These include candidate screening
and engagement, employee engagement, and helpdesk support . A 2017 survey by
Allegis found that 58% of respondents felt comfortable interacting with chatbots in
the recruitment and interview process . 66% said they would be glad to have chatbots
help schedule interviews and 61% said they were comfortable with AI tools being
used to assess their skills .80 In recruitment, the AI recruitment assistant Mya is able to
achieve up to 79% decrease in time to hire, up to 90% re-engagement by candidates
and up to 93% of candidates completing the recruitment screen from Mya .81
User comfort with chatbots in HR is demonstrated by successful use cases . Working
with a global financial services company, ubisend was able to build a chatbot to help
the HR team handle an average of 6,000 calls and 9,000 monthly emails from em-
ployees . The chatbot delivered a 37% reduction in human time needed .82
CognitionX has a companion research subscription and primer dedicated to the im-
pact of AI in HR83 .
Marketing
Chatbots are demonstrating much greater engagement rates than traditional mar-
keting techniques, especially email . For email, open rates are between 5% and 15%
while click through rates are 5-10% . In contrast, Facebook Messenger chatbots enjoy
open rates of between 75% and 95% with click through rates of 20% to 30% .84 Six
and Flow, a marketing agency, used a chatbot from Drift to increase leads by 23%,
reduce their sales cycle by 33% and grow their business by 15% .85
Voice interaction with chatbots also offers considerable opportunities for marketers .
A March 2018 study by Voicebot, PullString and RAIN Agency found that 19 .7% of US
adults own a smart speaker, up from less than 1% two years ago . In addition, 26%
of smart speaker owners have made a purchase by voice, 11 .5% make purchases
by voice monthly and 16 .7% of the general public is likely or very likely to order by
voice .86
Sales
Closely linked to marketing, the sales function in companies from across a wide
range of industries is also benefiting from chatbot deployment . Argomall, an online
consumer electronics company in the Philippines, deployed a chatbot on Facebook
80 http://bit .ly/2mkpsBT
81 http://bit .ly/2KWY9wl
82 http://bit .ly/2zFiSiS
83 http://bit .ly/2LfcAv0
84 http://bit .ly/2urd7R3
85 http://bit .ly/2NlxpCs
86 http://bit .ly/2LhxBoP
33
| The Business of Natural Language Computing
Keep up-to-date on chatbots: make sure you're a Pro subscriber bit.ly/cx-chatbots
20180724-1532
Messenger . Product sales from the chatbot increased by 10% delivering a 23x in-
crease in return on investment in the first months following its launch .87 Website
Hosting Insider implemented Winnie, a chatbot on Facebook Messenger designed
to help website owners find a hosting provider . Winnie achieved a 72% click through
rate of users clicking through to an affiliate hosting provider .88
RapidMiner deployed MarlaBot, a LeadBot developed by Drift, to provide an im-
proved bot-assisted sales experience . Using MarlaBot, RapidMiner was able to cap-
ture over 4,000 leads and the bot influenced 25% of the company's open sales
pipeline worth over $1m and is the source of 10% of all new sales pipeline created .
It also improved customer satisfaction with users 30% more likely to remain monthly
active users and a 20% higher NPS score .89
87 http://bit .ly/2Jpih4p
88 http://bit .ly/2uDREDI
89 http://bit .ly/2usj005
The Future of Chatbots:
Richer and More Pervasive .
35
| The Business of Natural Language Computing
Keep up-to-date on chatbots: make sure you're a Pro subscriber bit.ly/cx-chatbots
20180724-1532
The Future of Chatbots
We are at the maturation of what CognitionX likes to call chatbots v1: solid business
value is being delivered in a wide variety of situations, utilising low-friction, fea-
ture-rich customer platforms like Facebook Messenger and Telegram . The success-
ful use cases we have seen so far have relied mostly on traditional visual elements
and relatively inflexible conversation flows . Confidence is reflected across the board
including the amount of investment and volume of planned deployments across in-
dustries and corporate functions .
As for any emergent space, there
will be a period of consolidation, that
is already under way . As the mes-
saging chart showed, only a small
handful of chatbots are used, and
the rest will die . Nor can the mar-
ket sustain the dozens and dozens
of conversation development tools
and so there we can expect to see acquisition / acquihire, shuttering, or standalone
success, expanding language support, enterprise integrations, and much more ad-
vanced conversation flow support .
We will arrive at a much smaller number of solid proven test use cases . These will
highlight where chatbots can be successfully deployed and where their use is un-
proven or unsuccessful . This will push chatbots along the technology hype cycle to
the slope of enlightenment before reaching the plateau of productivity where the
technology is successfully integrated into everyday business operations .
The next frontier presents an opportunity for a step change in technology and
sophistication . But truly natural language is not currently very easy: the underlying
innovation engine, deep learning, has afforded revolutionary improvements in
images and recorded audio, but text is vastly more information-dense: users are
forgiving if pixels are not quite in the right place, orientation, or colour, but even
moving a word in a sentence can dramatically affect its meaning . We can look to
Google Duplex, and to some extent
Microsoft XiaoIce and London
startup Action .AI as demonstrations
what the next frontier of chatbots
looks like: very focused but much
more flexible conversation flows,
handling
interruptions,
subject
changes, corrections, errors, and
conflicting
and/or
ambiguous
information . Future platforms will make this more accessible to everyday businesses,
but connecting fully comprehensive natural language dialog to underlying system
capability comes with unavoidable data and functional complexity and cost: just
because it recognises the words you say doesn't mean it can do anything with them .
This may act as a brake on deployments across industries and corporate as cost and
complexity may not be justifiable for quite some time .
Winning chatbot builders will include lan-
guage support, enterprise integrations,
and much more advanced conversation
flow support .
Natural language recognition is fine, but
just because it recognises the words you
say doesn't mean it can do anything with
them .
36
| The Business of Natural Language Computing
Keep up-to-date on chatbots: make sure you're a Pro subscriber bit.ly/cx-chatbots
20180724-1532
Speech is ubiquitous in humans .
Given effective processing, it is also independent of graphical user interface . The
utility of speech is evident, however the developmental state of NLU is not the sole
limiting factor in voice interface adoption . Environments in which privacy, latent noise
and information density are an issue still favour the use of GUIs . Voice is making its
first forays in business in hotel rooms and conference rooms with Alexa for Business .
In general, voice and text chatbots will continue to improve due to improved domain
specific training data, improved conversation flow tools and better access to conver-
sation design best practice . Business uptake will continue to be strong as business
cases become easier to justify with a growing body of precedents . In the shorter
term (over the next 12 months) we can expect to see voice spreading across intimate
business spaces (hotel rooms, meeting rooms, etc) with screen support continuing
to grow . On the periphery we can also expect to see progress with other forms
of conversational engagement: private voice, subvocalisation, and mind control all
look promising . With voice-enabled beds already available, the ethical and utilitarian
boundaries of voice in the smart home will continue to be tested .
The autonomous assistive agent .
If the purpose of chatbot technology is to reduce the time and complexity of on-
line interactions, then the end-goal for consumers is a totally autonomous assistive
agent . The adoption of such assistant bots would have material implications on busi-
ness, namely in chatbot deployment and back office RPA tuned for bot-to-bot inter-
action . While current AI is not sophisticated enough to predict highly fluid consumer
preferences, the space should be observed carefully .
Beyond the future role of voice and fully autonomous assistive agents in the future
of chatbots, the success or failure of innovations in what we like to call "chatbots v2"
will be determined by five factors:

Higher ACE factors

NLU everywhere, not just inside chat

Growth in mixed-mode experiences

Full automation

Socialbots will become our concierges .
Higher ACE factors: better, richer, more emotional
conversations
Our ACE factor attempts to reflect on the essence of chatbot quality and is so far
proving to be useful, with Autodesk showing 10% increase in customer satisfaction
when they added emotional components (including nonverbal cues such as frown-
ing) . We therefore see future chatbots to improve their ability to handle ambiguity,
rich conversations, and emotional cues .
NLU everywhere
We believe that in future user expectations will evolve to the point where they ex-
pect to have a natural language conversation in any free-form input, be it a search
37
| The Business of Natural Language Computing
Keep up-to-date on chatbots: make sure you're a Pro subscriber bit.ly/cx-chatbots
20180724-1532
box, or an information field . Today, SalesForce Einstein and Google Analytics are
great examples of this .
Mixed mode conversations will have its day .
Mixed-mode experiences voice and screen, chat and web will become increas-
ingly common and important as customers become equally accustomed to interact-
ing with chatbots via voice applications as they are with text-based applications . For
example, a screen plus voice interface makes intuitive sense and companies will be
able to combine these to offer customers a better and more engaging experience .
Google's Assistant and Amazon's Alexa already highlight the possibilities for screen
and voice interaction . Rather than an experiment, this will be an inevitable part of a
conversational computing experience where information is presented either visually,
verbally or both according to what is most effective in the situation .
Some businesses will be 100% chatbots .
Increasingly sophisticated chatbots offer the prospect of a 100% chatbot-driven busi-
ness . Kakao Bank, South Korea's largest internet only bank, is launching a chatbot
that could answer up to 80% of customer enquiries .90 As companies gain a richer
understanding of customer requirements and reimagine fundamental business de-
sign with chatbots, they will get ever closer to customer support that is 100% online
using chatbots .
Socialbots will become our concierges: for better or worse
Multiple signals suggest to us that a major frontier to be breached is the socialbot
concierge: your first point of call to start your conversation, to draw in other ser-
vices as needed . Today, at best, Al-
exa and Google Assistant play this
role, but it's extremely basic; Micro-
soft's having more luck with XiaoIce
and siblings . Where it gets interest-
ing is when a single request spans
multiple services, such as "Order
my usual Chinese takeout after my
last meeting today"91 and it's with
this perspective it becomes so clear
why Amazon is investing so much in developing a socialbot . Socialbots build im-
mense levels of trust, and you take advice from those you trust: including purchase
decisions .
This last frontier is both exciting and worrying . Are socialbots the new Operating
System? Indeed, they are poised to become the new control-point through which
all access to information, products and services is given . Historically this was the
desktop operating system, then the mobile, and now our beloved social companion .
Fingers crossed governments play their role to ensure the right safeguards are in
place to keep the marketplace fair and equitable .
90 http://bit .ly/2uH6TMe
91 http://bit .ly/2LmU2Wh
Are socialbots the new Operating System?
Indeed, they are poised to become the
new control-point through which all ac-
cess to information, products and services
are given .