ARTIFICIAL INTELLIGENCE IN

BANKING
WHERE TO START?

Dan Latimore
September 2018

This is an excerpt from a previously published

Celent report. It was prepared for Finn.ai and
has been slightly modified due to progress in
the market, but the analysis has not been
changed, and the report was not sponsored by
Finn.ai in any way. For more information about
the full report, please contact Celent at
[email protected].
CONTENTS

Executive Summary ............................................................................................................ 1

Key Research Questions ................................................................................................. 1
Basics of Artificial Intelligence ............................................................................................. 4
Enabling Technologies .................................................................................................... 5
Applications in Banking ....................................................................................................... 8
Analytics .......................................................................................................................... 8
Chatbots ........................................................................................................................ 10
Robotic Process Automation ......................................................................................... 11
Report Generation ......................................................................................................... 12
Business Case for AI......................................................................................................... 13
Reduce Costs ................................................................................................................ 13
Mitigate Risk .................................................................................................................. 14
Increase Revenue / Improve CX ................................................................................... 15
Areas of Impact ................................................................................................................. 17
Front Office .................................................................................................................... 17
Middle Office .................................................................................................................. 17
Back Office .................................................................................................................... 17
Next Steps ......................................................................................................................... 18
Become Educated and Develop Principles ................................................................... 19
Lay the Groundwork ...................................................................................................... 19
Take Action .................................................................................................................... 20
Leveraging Celent’s Expertise .......................................................................................... 21
Support for Financial Institutions ................................................................................... 21
Support for Vendors ...................................................................................................... 21
Related Celent Research .................................................................................................. 22
EXECUTIVE SUMMARY

KEY RESEARCH QUESTIONS

What is artificial How can banks use

1 intelligence in a
banking context? 2 Why does AI matter
to banks? 3 AI to improve
business results?

Artificial Intelligence (AI) is extraordinarily popular when judged by today’s banking

headlines, but those headlines have outpaced today’s practical banking reality. Relatively
few banks have begun production or even full-blown research at this stage. For those
who think they’re lagging, the good news is that they’re not — there’s still some time. But
make no mistake: lack of progress today doesn’t mean that banks can afford to ignore AI;
they must formulate a strategy to deal with the opportunities that it promises to them and
to their competitors.

We should also note that AI is a big and complicated topic. This report provides an
introduction to AI by focusing on six areas.

1. Foundational technologies underpinning AI.

2. Key banking applications.
3. Business cases.
4. Relative project value.
5. Areas of impact (front, middle, and back office).
6. Next step for bankers to pursue.

Bankers looking for an introduction to the uses of AI in banking will find this report a
useful overview; over time additional reports will dive more deeply into specific areas.

Celent takes a pragmatic approach that defines AI in banking as technology that makes
inferences and decisions that used to require direct human involvement.

A series of fundamental and interrelated technologies around machine learning and

natural language underpin all of AI. Crucially, AI is not just better technology. It is not
faster processing, bigger data sets, or even thousands of rules rigidly applied. These
advances have yielded powerful results, but they’re performing old tasks better. That AI
can respond to ambiguous real-world inputs probabilistically is one of its critical features.
Building off the fundamental technologies to apply them in a banking context yields four
main AI applications today: analytics; bots; robotic process automation (RPA); and
report generation. Figure 1 lays out the basic AI relationships between the foundational
technologies and banking applications; all depend on massive amounts of data, the
lifeblood of AI.
 Figure 1: Artificial Intelligence in Banking

Source: Celent analysis. Arrows represent influence of source on target.

Today’s principal business justification for AI lies in cost reduction, but it can also mitigate
risks and increase revenues (primarily through the indirect route of improving customer
experience). AI is an umbrella term for a host of different technologies and applications;
when discussing it, being specific is helpful in narrowing the focus and clarifying whether
a solution truly uses AI or merely performs standard tasks better with new technology. AI,
for example, has applications in the front, middle, and back office, and can be either
customer-facing or employee-facing. Being clear about how, where, and why AI is to be
used is critical for a reasoned discussion about its pros and cons.

That being said, banks must develop an AI strategy so that they’re not left behind. As
they do this, three critical steps are necessary.

• Become educated and develop principles.

• Lay the groundwork.
• Take action.

More detail is provided in Figure 2.

Chapter: Executive Summary

2
 Figure 2: Next Steps on AI

Source: Celent analysis

The line between impressive technology and true AI is sometimes blurry, and we’ll try to
draw the distinction while recognizing that, in some cases (like RPA), being overly
pedantic causes more problems than it solves. And we’re not going to cover robo
advisors; our colleagues in wealth management have done a great job of that already.

Today’s AI is not yet ready to replace humans; instead it will augment them, letting them
move into more value-adding activities, freeing them from rote actions, making them
more efficient, and performing calculations that would be physically impossible for a
single person. We take an optimistic view of AI in banking — for those who embrace it, AI
will over time provide a better experience for customers and employees while delivering
real business value on every dimension.

Chapter: Executive Summary

3
BASICS OF ARTIFICIAL INTELLIGENCE

Discussion of AI can rapidly turn philosophical; as much as we like a good debate, we’re
going to take a very pragmatic approach to a definition for financial services.

Banking artificial intelligence is technology that makes inferences and decisions

that used to require direct human involvement.

Cross-industry AI has been hot twice before (the term was coined in 1956). Today’s third
wave has hit banking. So it’s useful to highlight the textbook attributes of AI as taught in
computer science courses.

• Reasoning
• Knowledge
• Planning
• Learning
• Natural language processing
• Perception
• Manipulating the physical world

The ultimate goal is generalized intelligence: think of the computer in Star Trek or the
replicants in Blade Runner. We’re clearly a long way from that, but there’s no doubt that
we’ve made significant progress as the forces powering AI — computing power, data,
and algorithms — have all advanced to the point where the technology can play a
meaningful role in banks.

It’s also worth noting that the bar that defines which technologies qualify as AI is
continually moving; optical character recognition (OCR) used to be thought of as AI, but
now it’s simply a technology that consumers can access via scanners plugged into their
home computers.

It’s also necessary to draw the distinction between AI, which infers, and brute force
computing, which applies a set of rules, sometimes a very big set of rules, to a set of
inputs and returns an answer. If presented with inputs that its rules had never
encountered, the brute force method would not be able to respond; any AI worth its salt
would at least have a probabilistic answer. Some technologies, RPA in particular, are not
technically AI (at least not most of them), but advances are continually moving them
closer to using AI methods. While we don’t want to be pedantic about defining AI, the Chapter: Basics of Artificial Intelligence
term is often used somewhat sloppily. Simply applying a lot of computing power to a lot of
data is not AI, no matter how good the initiative might be for marketing.

Key What is artificial intelligence in a banking context?

Research
Question

1 Technology that makes inferences and

decisions that used to require direct human
involvement.

4
ENABLING TECHNOLOGIES
Today’s evolving AI is made possible by foundational technologies (computing power,
data, and algorithms) and the AI syntheses of those foundations, most notably machine
learning and language manipulation. Machine learning has captured a lot of press, but
natural language shouldn’t be downplayed, particularly because its implementation is
often much easier. We should note that vision is a fascinating area of AI research (e.g., a
machine being able to recognize an image of a cat), but today’s banking applications are
limited.

Machine Learning
Machine learning is a centerpiece technology of true AI.

Machine learning occurs when computers change their parameters/algorithms on

exposure to new data without humans having to reprogram them.

The biggest benefit of machine learning is that, by processing more data than a human
ever could, and then using that data to teach itself by spotting patterns that, again, would
be hard for humans to identify, it will be able to draw insights that previously would have
remained undiscovered. Although machine learning opens up possibilities, banks should
keep in mind three key ideas.

First, we may not know how a machine learning algorithm arrives at its conclusion.
Rather than testing hypotheses, as a human might in order to avoid “boiling the ocean,”
machine learning goes ahead and boils the ocean. It may present answers (generally
couched in terms of probabilities) based on the clustering of several obscure and
potentially unrelated variables whose causal roots are difficult to describe. In other words,
the algorithm may not be able to describe how it got to its answer. This poses obvious
auditability issues and makes it difficult to justify decisions to regulators. For example, if
AI decides to reject a loan application, what would the adverse action letter say?
Researchers are addressing this issue, but it’s not solved yet.

Second is that the amount of data that a machine learns from is critical. The more data,
the better the results will be. Gathering huge amounts of input will give the algorithms
more opportunities to test and learn, thereby refining their rules and, over time, producing
better results. Firms with access to more data will have a leg up over those firms whose
data is sparser. This is one area where there’s a definite size advantage that accrues to
larger banks.

Third and finally, learning can be unsupervised or supervised; having a human check in
on the results and give the algorithms nudges or feedback, based on perceptions that are
intuitive to humans but difficult for models to pick up, can generally speed up the learning
process significantly while improving accuracy. Chapter: Basics of Artificial Intelligence

5
 Figure 3: Machine Learning Ideas

Source: Celent analysis

Natural Language
One of the most notable areas where artificial intelligence has progressed is in the way it
deals with human language. For many years the industry joke was that good voice
recognition was just two years away — and that was the refrain for a decade! Today,
though, dictating via a phone or giving instructions to an automated voice when calling an
800 number generally gives good results (frustrations with Siri notwithstanding). In an
illustration of how AI technologies are mutually useful and reinforcing, many natural
language platforms use machine learning to improve their accuracy in recognizing or
generating human speech. Speech analytics offer an illustration of how AI technologies
are interrelated; speech analytics use machine learning to analyze patterns in speech,
learning, for example, how to identify an angry customer.

While natural language processing (NLP) has received the lion’s share of attention,
Celent believes that natural language generation (NLG) holds immense potential as a
means of turning raw data into useful and easily consumed insights.

Processing
Natural language processing is the ability for technology to use human
communication, naturally spoken or written, as an input that prompts computer
activity.
Chapter: Basics of Artificial Intelligence
NLP is critical to AI because it gives people an intuitive way to communicate their intent
to a program. The human’s unstructured words are parsed and converted into machine-
understandable instructions. 1 Once the program has understood and processed the
instructions, it responds. The response can be in a computer-friendly format (think of
simply showing a list of transactions when the query is “Show me my last three credit
card purchases”), a useful graph, or spoken or written in plain language.

1
A quick note on Natural Language Understanding (NLU), a subset of NLP, is in order. NLU is described as an
“AI-hard problem,” meaning that it can’t be solved by a specific algorithm, nor can it be solved by today’s
technology alone — it would need human intervention. For our practical purposes, however, we consider NLU
to be a more sophisticated type of NLP that takes into account a wide variety of input variations, tonality, irony,
implications, and assumptions about pre-existing knowledge. Some service providers say that they provide
NLU, but it’s generally an aspiration, since true NLU doesn’t yet exist.

6
Generation
Generating plain language responses is a critical component of being able to eliminate
the need for human intervention and can automate many types of routine tasks (think
report generation) that today are typically very time-consuming.

Natural language generation (NLG) is the ability of technology to produce human-

quality prose based on a wide variety of inputs.

NLG can be either written or spoken, although if spoken, the voice quality doesn’t
necessarily have to be indistinguishable from humans. NLG does not rely on a limited
number of predefined responses or fetching a single piece of data for a customer, but
instead sorts through large amounts of available data to produce a human-sounding
response. It can take the form of speech, or of a multipage report summarizing financial
results, as just two examples. See Figure 4 for an example of NLG writing a story on
Denny’s earnings in 2015 and compare the result to the effort by a talented NPR reporter.

Figure 4: Natural Language Generation in Action

Chapter: Basics of Artificial Intelligence

Source: NPR. http://www.npr.org/sections/money/2015/05/20/406484294/an-npr-reporter-raced-a-machine-

to-write-a-news-story-who-won

With two basic technologies explained, let us now turn to some banking-specific use
cases that exploit them.

7
APPLICATIONS IN BANKING

AI is not the brute application of computing power to huge databases using a set of static
rules, no matter how complicated. A huge set of if-then statements, for example, does not
constitute AI (however, if the algorithm constructed new if-then statements based on its
prior analyses, that would be a case of machine learning). Remember, AI must make
inferences, not just decisions. We’ll examine four specific cases of AI in banking,
touching on the highlights of each and being mindful that each of these four is worthy of a
report in its own right. Another area adjacent to banking that we’ll mention but not
examine is robo advisors, a topic covered in a host of other Celent reports from our
Wealth Management colleagues.

Key Why does AI matter to banks?

Research
Question
AI can help banks become more efficient and
2 effective by reducing costs, mitigating risk, and
increasing revenues by applying newly capable
technology in analytics, bots, RPA, and report
generation.

ANALYTICS
An area of intense banking focus for the last several years, analytics has benefited from
technological improvements even without AI. Fundamental advances in processing
power and data manipulation have given human analysts increasingly powerful tools.
Couple this with machine learning progress, and firms have the potential to uncover
insights previously undiscoverable.

AI-driven analytics test vast quantities of data to search for patterns, groupings,
and correlations by employing a wide variety of techniques, including data mining
and hypothesis testing. Moving from merely descriptive to predictive, near real-
time analysis is a key component of some analytics.

Chapter: Applications in Banking

8
 Figure 5: Three Intertwined Technological Advances Driving Analytics

Source: Celent analysis

The amount of processing power required is simply beyond the ability of humans to
provide. Thousands of transactions are being submitted every minute from all over the
world. For one employee to give a timely response to a single transaction is impossible: a
human simply couldn’t retrieve and analyze the data necessary to make an informed
assessment quickly enough. Scaling makes the problem even more intractable. Similarly,
analyses that used to take days can now be performed in minutes thanks to technologies
like in-memory processing.

This processing power is being brought to bear on ever greater amounts of data. Banks
have a head start on many industries in that they already possess an immense amount of
information about their customers, an asset that they still consistently underutilize, in part
because of the siloed nature of their organizations. This data, though, when combined
with behavioral information (for example, How many times does the customer interact
with the bank? When does the customer make payments?) and external data (Which
websites has she been visiting?) can yield valuable insights that can shape how the bank
interacts with its customers and make the relationship vastly more personal. And with the
vast amounts of data being ingested, there are more chances to learn from mistakes and
make corrections. Data is the lifeblood of machine learning, and analytics that rely on
huge data sets are ideal use cases.

Finally, improvement in the algorithms that power machine learning happen more
quickly with more data. Every time a mistake slips through (an erroneous decline, or a
missed fraudulent transaction), a program using machine learning notes that and makes
Chapter: Applications in Banking

an appropriate adjustment. Machine learning speeds the process of improving many

types of analytics algorithms, and many different sorts of analytics challenges lend
themselves to AI: risk modeling, biometric identification, fraud detection, or credit
underwriting to name a few. Each of these areas is fascinating, complex, and deep.
We’ll touch briefly on credit underwriting for the sake of focus.

The situation is familiar to any lender. Will a credit applicant receive a loan, and if so, on
what terms (amount, rate, repayment period)? And what criteria will be used to determine
the answers to these questions? A properly constructed machine learning algorithm can
analyze not only all of the underwriting data used today, but also other factors, assuming
that they can be cleanly brought into the data set. The algorithm can be taught by feeding
it data from prior underwriting decisions and analyzing which loans performed and which

9
didn’t. Rather than testing hypotheses, the algorithms can simply data-mine the entire
portfolio to determine which variables (or clusters of variables) are correlated with
performance. It can then reweight the importance assigned to various variables and
develop revised underwriting criteria. Each time that it learns more (based on updated
performance data), it can tweak its criteria. Of course, in jurisdictions like the US where
denying a consumer credit requires an explanation, the bank must be sure of its rationale
and willing to defend it to consumers, regulators, and the press.

CHATBOTS
Perhaps the most visible instance of AI today is the chatbot.

A chatbot is a technology service powered by algorithms that interacts with a

customer in a natural (human-like) manner, either by voice or text.

Some chatbots are given personas in the form of a virtual agent, which may be named
and given some sort of avatar and personality. Typically accessible via a bank’s own
digital properties, virtual agents are marketed explicitly to customers as non-human
technology tools that will let them interact with the bank better than IVR-powered phone
trees or call centers.

Chatbots have various levels of sophistication. The more basic accept a limited range of
inputs and produce a limited set of answers (think of a host of if-then answers, which
mean that they’re technically not AI), and at their most basic may be essentially menu-
driven. The more sophisticated ones parse customer queries and can detect the
questioner’s intent even when it’s expressed in a wide variety of different ways. Whether
an unassuming text interaction or the more explicit virtual agent, the more sophisticated
rely on NLP and NLG.

For example, consider the query, “What is my routing number?” A simple rules-based
engine may simply pick up on the term “routing number” and direct the customer to the
page on the bank’s site that tell them how to find it on their checks. A more sophisticated
AI would understand the difference between “What is my routing number?” and “What is
a routing number for?” In the first case it would simply give the customer her routing
number, while in the second, would provide an explanation of its function. In neither case
would a customer be directed to a site, but would instead be given the direct answer in
the context of that particular chatbot conversation.

It’s important to note that chatbots should not be left on their own. When a customer
presents a situation that the chatbot is not equipped to deal with, it must quickly hand the
interaction off to a human quickly and seamlessly so that the customer isn’t dissatisfied.
In these relatively early days in particular, it’s critical that banks make a good first
impression. No one wants a repeat of the disgust that most consumers feel toward IVR
systems and their impenetrable menus.
Chapter: Applications in Banking

One other form of bot is a physical manifestation of AI, typically in humanoid form (Figure
6). While some banks have begun to experiment with these physical robots to solve for
particular use cases (e.g., the need for multilingual interactions for the upcoming
Japanese Olympics), Celent doesn’t view this as the current sweet spot of banking AI
technology.

10
 Figure 6: Pepper the Robot

Source: http://spectrum.ieee.org/automaton/robotics/humanoids/softbank-pepper-humanoid-robot-us-debut

ROBOTIC PROCESS AUTOMATION

Based on a prior caveat, there’s an argument to be made that RPA doesn’t qualify as AI.
Here’s why:

RPA uses a number of techniques to mimic routine human activities automatically,

repeatedly, faster, and more accurately.

In the most basic case, think of any set of routine bank activities performed by junior-level
employees sitting in a cube. They receive inputs (whether on paper or digitally), examine
those inputs, apply a rule to them (with no discretion in this extreme scenario), and then
send the output forward to the next step in the process. In this edge-case, there is no
inference required, no latitude allowed. RPA steps into a gray area when interpretation is
required. Does the supporting documentation from utility companies suffice to prove
continued residency for purposes of a mortgage application, for example? For us, RPA
that uses machine learning to continually improve counts as AI; that which simply has a
large and ever-expanding library of scenarios, generally accreted manually by human
supervisors, is simply plain automation. That’s not to say that RPA isn’t a valuable
technique that many banks should be examining, but simply to emphasize that much
RPA isn’t in fact AI.

RPA powered by AI is nevertheless the next stage in the evolution of making back offices
Chapter: Applications in Banking

more efficient. The first wave was labor arbitrage — of moving basic functions to
jurisdictions to be performed by lower-paid workers Replacing those workers with
programs aims to lower costs, increase throughput, and improve accuracy. If the first
stage of RPA is simply “paving the cowpath,” and the next is adopting new processes
based on capabilities of new technology, then AI-powered RPA will make judgments that
today require humans and, eventually, begin to suggest and implement process
improvements on its own.

11
REPORT GENERATION
NLG can turn data into prose. As one example, the first drafts of Little League baseball
stories in local papers are being written by NLG technology based on the box scores of
games.

NLG vendors tout the use cases of writing reports or synopses that synthesize large
amounts of structured data and put them into a prose narrative that highlights the key
points. One example might be a quarterly sales report that tracks results from across
many different regions and product categories. The NLG platform would ingest the raw
data and then write a report distilling the key points. Rather than a human business
analyst poring through the raw data, searching for high and low points, calculating
CAGRs and rates of change, the NLG engine would do all of this.

Another example might be constructing a spending report for a retail banking client,
someone who to date has never had any kind of prose analysis from their bank. Many
clients will prefer prose summaries, rather than a set of numbers, tables, and graphs that
takes time, and some degree of expertise, to decipher.

It’s important to note that NLG should be used to write the first draft, at least for reports
that aren’t being mass-produced for distribution to thousands of clients. Analysts will want
to ensure that what’s written reflects their emphases and voice, and also want to confirm
that they understand the conclusions!

With four key areas of banking AI laid out, let us now turn to the business justifications of
AI.

Chapter: Applications in Banking

12
BUSINESS CASE FOR AI

As daunting as AI may seem for many banks, Celent believes that most institutions
should be exploring at least the basic forms of the many technologies that fall under its
umbrella. The main and most basic benefit is reducing cost, but firms shouldn’t ignore
risk mitigation (via fraud reduction or better underwriting outcomes, for example) and
revenue increases (generally via an improved customer experience or better marketing).
We offer some of the more promising use cases that are practical today, with the caveat
that this field is rapidly evolving; more implementations may follow in the coming months.

Key How can banks use AI to improve business results?

Research
Question

3 Reducing costs and mitigating risks are the initial

benefits of AI; increasing revenue through improved
customer experience is more ambitious.

REDUCE COSTS
There are many ways that well-implemented AI can reduce costs. Several ideas follow in
Figure 7; the list is by no means exhaustive. AI may handle many basic inquiries that
nevertheless require some degree of judgment. It may help humans do their jobs better
or faster. And it can prepare accurate reports that won’t miss crucial details.

Figure 7: Reducing Costs with AI

Chapter: Business Case for AI

Source: Celent analysis

13
Of course, AI technologies never need a day off and can work around the clock. Note that
while these actions tend to augment the role of the human for rote or basic tasks, they
still need varying degrees of human supervision.

MITIGATE RISK
Risk is a huge and complicated issue in banking. Figure 8 describes the risks that banks
must manage on an ongoing basis.

Figure 8: Celent Risk Management Taxonomy

Source: Celent

Technology has, of course, been both a cause of increased risk and a way to mitigate it.
While most risk mitigation methods do not have an AI component and are instead
applications of “dumb” (but nevertheless sophisticated) algorithms and processes, AI will,
over time, be able to help mitigate many of the individual risks described above. Today,
credit risk and fraud risk hold the greatest promise for the application of AI.
Chapter: Business Case for AI

For credit, origination/underwriting is the specific area that can most benefit from today’s
machine learning technology. Banks can use machine learning to evaluate historical
underwriting performance by looking for approved applications that later defaulted and
analyzing them to see if there are factors that would have had predictive power. Some
examples that appear to be predictive on an experimental basis included choice of
specific words and whether applicants apply in capital letters. Of course, rejected
applications won’t be as useful in this scenario, but even slight improvements in default
rates can have a significant impact on the bottom line.

In financial crime, much technological benefit again comes from plain old automation. AI
can nevertheless improve outcomes, particularly as the integration of data and algorithms

14
progresses. To reduce fraud, machine learning can assess a candidate transaction and
compare it to stored transactions, both proper and fraudulent, and make a probabilistic
assessment as to whether to approve the transaction or flag it for further investigation. As
in all cases of machine learning, more data improves outcomes, and the more
transactions the algorithm ingests, the better it will perform. Improved authentication is
another area where AI can help, often through biometrics.

Figure 9: Mitigating Risk with AI

Source: Celent analysis

To implement effective anti-money laundering (AML) procedures, machine learning can

again compare candidate transactions to prior versions and make a quicker and more
thorough initial assessment of the legitimacy of the transaction. Additionally, the AI
should, assuming that it has been programmed properly by humans, always be up to
date with the latest regulations. Automation (though not necessarily AI) can provide a
suitable audit trail.

For cybercrime, AI’s thoroughness and ability to adapt to new techniques can again be
very useful. And AI can improve underwriting and credit assessment, assuming that it has
access to the right data, properly configured. When assessing credit risk, the ability of the
AI to data-mine and identify potentially unrelated clusters of factors that, taken together,
correlate with an outcome is another tool in the underwriter’s kit.

Error reduction (typos, miscalculations, and misreadings) is another way that AI can help
mitigate risk. Because the AI never gets bored or tired, and can calculate flawlessly, it
won’t be subject to the kinds of errors that humans might.

INCREASE REVENUE / IMPROVE CX

Despite a great deal of hype today, AI’s ability to increase revenue and improve the
customer experience has not been widely proven at scale – though some early results
are encouraging. In terms of prioritization, at this stage there’s more potential to improve
Chapter: Business Case for AI

the customer experience than there is in increasing sales directly.

15
 Figure 10: Increasing Revenue and Improving Customer Experience with AI

Source: Celent analysis

Many observers gravitate toward using AI to increase revenue, although the challenge
with many of these customer-facing activities is that they run the risk of annoying the
customer if executed poorly. Banks should apply rigorous testing and quality assurance
cycles to any customer-facing AI deployment prior to launch to ensure they don’t alienate
customers with beta technology.

In areas where processes are working effectively today, it’s best to leave well enough
alone; focus on where things are broken or could be done appreciably better. Improving
employee effectiveness is one instance of this, particularly by relieving them of rote and
non-value-added tasks. For example, some early adopter banks say they’ve seen a
reduction in calls to their customer care centers due to customer / chatbot interactions.

Speeding up the customer experience is one more area where improvements can be
made in a relatively low-risk way. A good example is the ‘always on’ nature of AI and how
this could allow banks to extend customer support from traditional business hours to a
24/7 model.

Chapter: Business Case for AI

16
AREAS OF IMPACT

There are many different ways of examining how AI will affect a bank. In addition to
business value drivers, another perspective is to examine the functions in the bank that
will be affected. AI has the potential to improve all parts of the bank, from front end,
customer-facing areas; to the middle office; and through to the back office. While some
argue that these terms are becoming obsolete, they’re still a useful construct to consider
how AI can create business value. In each case we’ll look at customer benefits, as well
as risks for both the consumer and the bank. The prior section covered the business
benefits for the bank

FRONT OFFICE
Direct customer interactions can either be addressed by AI directly, most prominently
with chatbots or virtual agents, or by enabling employees to do their jobs better (that is,
by enabling them to be faster or more accurate, or more productive — efficiency), or by
augmenting their capabilities (effectiveness). Another area of customer interaction is
through next best offers or financial advice or nudges. And finally, AI may simply improve
the employee experiences.

Customer-facing AI can be either one-way (a push of offers, for example), or two-way,

where an ongoing set of interactions takes place, initiated by either the customer or the
bank. The benefits to customers include improved advice, better offers, and saved time.

While the benefits are tangible, there are risks inherent in exposing AI directly to
customers. If the implementation is flawed, the mistakes may earn the bank or credit
union public ridicule, or it may poison the well of customer perception such that
customers who had a bad experience will not willingly interact with the AI again.

MIDDLE OFFICE
The functions of the middle office include employees supporting other employees,
indirectly supporting customers, or conducting compliance activities. Some salient
examples are in report generation, underwriting and credit decisioning, and risk and
compliance monitoring. While middle office activities are generally an effectiveness play,
efficiency has a role, too, as AI technology helps process more work at lower cost.
Identifying exceptions (and becoming better at doing so) is one example of AI helping
employees become both better and faster at their jobs.

The risks of middle office AI are relatively low from a consumer standpoint as long as
outcomes (like false positives) aren’t degraded. For all of these AI activities, human
supervision is critical. Internal reports generated via NLG, for example, should only be
considered first drafts; the responsible analyst should sanity-check them, revise them for
tone and voice, and ensure that she knows the substance of the report.

BACK OFFICE
Processing and reconciliation, typical back office functions, can use AI to detect
Chapter: Areas of Impact

anomalies and exceptions. Layered on top of non-AI RPA, they can serve as a backup
and a second set of eyes to make sure that processes are proceeding as they should be.
Direct benefits to customers are hard to envision in a back-office AI situation, but the
risks also aren’t as high. A great deal of threat detection and risk mitigation will also take
place in the back office. The biggest risks in back office implementations lie in over-
reliance on AI and in the cost associated with putting initiatives in place.

17
NEXT STEPS

Every bank should develop a strategy for incorporating AI into its technology stack. AI
can reduce costs, mitigate risks, and improve the customer experience.

Figure 11 provides a summary of the economic rationale and areas of impact of each of
the AI areas we’ve been discussing.

Figure 11: How and Where AI Can Help Banks

Source: Celent analysis

“Think about how to implement AI” is an unrealistic and unhelpful description. AI is an

umbrella term that covers a host of specific technologies and implementations. To break
this huge task into manageable pieces, Celent offers three specific high-level steps to
make AI less daunting and more manageable:

• Become educated and develop principles.

• Lay the groundwork.
• Take action.

A summary is provided in Figure 12.

Chapter: Next Steps

18
 Figure 12: Next Steps on AI

Source: Celent analysis

BECOME EDUCATED AND DEVELOP PRINCIPLES

• Keep an eye on the market; learn the landscape.
Basic AI technologies are improving rapidly, and bank leaders are making great
progress in specific implementations. Monitor market developments as a cost-
effective and efficient way of conducting research.
• Assess portfolio of potential initiatives.
Based on your particular situation, evaluate the implementations that will make the
most sense for your institution given your strategic priorities. Building on your market
analysis, decide when your target technologies are mature enough to begin
implementation, or at least experimentation.

LAY THE GROUNDWORK

• Develop proper governance.
We hear questions from our banking clients on where AI should sit within the
organization. There’s not a one-size-fits-all answer, but our guidance is that it should
mirror a bank’s digital efforts. AI governance, though, must include the proper internal
stakeholders, determine the initiatives to pursue, and define and measure desired
outcomes.
• Seek ways to help humans, not necessarily replace them.
AI is not a panacea, and unrealistic expectations about its capabilities can doom AI
efforts. Early AI won’t replace humans, but will help them do their jobs better and
make those jobs more interesting. Don’t expect more than that, at least not in these
early days.
Chapter: Next Steps

• Consider using a partner.

Celent believes that there is a lack of AI skills within many banking IT departments.
This is by no means an indictment of banks, for this a new and emerging technology
with relatively few experts. Consider finding a trusted partner who can help you in
your AI journey.

19
• Evangelize and educate throughout the organization.
AI may cause trepidation in the bank. Develop a communications strategy that clearly
lays out objectives and keeps colleagues apprised of progress. There are fears that
technology and AI will put people out of work. You must address those concerns
head on and without platitudes; emphasize that AI can enhance productivity and
improve the employee experience.
• Put your data house in order.
AI success is directly correlated to data access; not just how much data you have,
but how well you’re able to use it. Make sure that you can leverage what you have,
and realize that if you’re a bigger institution, here’s one clear advantage of your
scale.

TAKE ACTION
• Start small to gain an early win.
Succeeding with a small early implementation is more important than failing
ambitiously. Pick a well-defined and manageable project whose failure won’t be
catastrophic and whose success will be inspirational. Consider back-testing on
existing data to train the AI and to experiment. Be realistic: AI is not a panacea, nor is
it even appropriate for most technology situations. But where it can be helpful, use it,
and be ready to shift course quickly based on your learnings.
• Be upfront and transparent with customers.
Transparency is a new banking imperative that customers have learned from other
industries. Google became famous for its beta releases — customers (at least
initially, before cynicism set in) appreciated that there might be bugs, and generally
understood the rationale. Take a leaf from this book and be clear about your goals
and methods with your customers.
• Track progress and adjust.
No AI implementation will work perfectly out of the box; design your tracking and
feedback mechanisms from the beginning so that you can gauge progress from the
outset and make the inevitable adjustments based on full information.

What’s considered AI today will likely become just technology tomorrow. Remember
OCR. Banks and credit unions should nevertheless think about which specific business
challenges AI can help solve and begin to prepare today.

Was this report useful to you? Please send any comments, questions, or suggestions for
upcoming research topics to [email protected]. Chapter: Next Steps

20
LEVERAGING CELENT’S EXPERTISE

If you found this report valuable, you might consider engaging with Celent for custom
analysis and research. Our collective experience and the knowledge we gained while
working on this report can help you streamline the creation, refinement, or execution of
your strategies.

SUPPORT FOR FINANCIAL INSTITUTIONS

Typical projects we support related to artificial intelligence include:

Vendor short listing and selection. We perform discovery specific to you and your
business to better understand your unique needs. We then create and administer a
custom RFI to selected vendors to assist you in making rapid and accurate vendor
choices.

Business practice evaluations. We spend time evaluating your business processes,

particularly around artificial intelligence. Based on our knowledge of the market, we
identify potential process or technology constraints and provide clear insights that will
help you implement industry best practices.

IT and business strategy creation. We collect perspectives from your executive team,
your front line business and IT staff, and your customers. We then analyze your current
position, institutional capabilities, and technology against your goals. If necessary, we
help you reformulate your technology and business plans to address short-term and long-
term needs.

SUPPORT FOR VENDORS

We provide services that help you refine your product and service offerings.
Examples include:

Product and service strategy evaluation. We help you assess your market position in
terms of functionality, technology, and services. Our strategy workshops will help you
target the right customers and map your offerings to their needs.

Market messaging and collateral review. Based on our extensive experience with your
potential clients, we assess your marketing and sales materials — including your website
and any collateral.

Chapter: Leveraging Celent’s Expertise

21
RELATED CELENT RESEARCH

Under the Spotlight: Innovative Vendors in Financial Crime Case Management

Technology
July 2017

Understanding the Investment into AI in Banking: Celent Digital Panel Series 6

July 2017

Personal Financial Experiences (PFE): Putting the 'E' in Personal Finance

July 2017

PFM Is Dead, Long Life PFE: The Advent of Personal Financial Experiences
July 2017

Applying Conversational Commerce to Wealth Management: Aligning IT to the Machine

World
May 2017

Robo in a Box: The UK Adviser and the Future of B2B Robo Advice
April 2017

Alior Bank: Dronn, The AI Agent

April 2017

Mizuho Bank: AI and Social Media Banking

April 2017

Robo UMA: Automated Advice and the Battle for the Affluent Investor
March 2017

A Survey of Retail Banking Channel Systems in North America: Omnichannel Emerges

February 2017

Artificial Intelligence in Insurance: Use Cases from Early Adopters

December 2016

Treating Cyber-Risk as an Operational Risk: Governance, Framework, Processes, and

Technologies
October 2016

Chapter: Related Celent Research

22
 Chapter: Related Celent Research

23
Copyright Notice
Prepared by

Celent, a division of Oliver Wyman, Inc.

Copyright © 2017 Celent, a division of Oliver Wyman, Inc., which is a wholly owned
subsidiary of Marsh & McLennan Companies [NYSE: MMC]. All rights reserved. This report
may not be reproduced, copied or redistributed, in whole or in part, in any form or by any
means, without the written permission of Celent, a division of Oliver Wyman (“Celent”) and
Celent accepts no liability whatsoever for the actions of third parties in this respect. Celent
and any third party content providers whose content is included in this report are the sole
copyright owners of the content in this report. Any third party content in this report has been
included by Celent with the permission of the relevant content owner. Any use of this report
by any third party is strictly prohibited without a license expressly granted by Celent. Any
use of third party content included in this report is strictly prohibited without the express
permission of the relevant content owner This report is not intended for general circulation,
nor is it to be used, reproduced, copied, quoted or distributed by third parties for any
purpose other than those that may be set forth herein without the prior written permission
of Celent. Neither all nor any part of the contents of this report, or any opinions expressed
herein, shall be disseminated to the public through advertising media, public relations,
news media, sales media, mail, direct transmittal, or any other public means of
communications, without the prior written consent of Celent. Any violation of Celent’s rights
in this report will be enforced to the fullest extent of the law, including the pursuit of
monetary damages and injunctive relief in the event of any breach of the foregoing
restrictions.
This report is not a substitute for tailored professional advice on how a specific financial
institution should execute its strategy. This report is not investment advice and should not
be relied on for such advice or as a substitute for consultation with professional
accountants, tax, legal or financial advisers. Celent has made every effort to use reliable,
up-to-date and comprehensive information and analysis, but all information is provided
without warranty of any kind, express or implied. Information furnished by others, upon
which all or portions of this report are based, is believed to be reliable but has not been
verified, and no warranty is given as to the accuracy of such information. Public information
and industry and statistical data, are from sources we deem to be reliable; however, we
make no representation as to the accuracy or completeness of such information and have
accepted the information without further verification.
Celent disclaims any responsibility to update the information or conclusions in this report.
Celent accepts no liability for any loss arising from any action taken or refrained from as a
result of information contained in this report or any reports or sources of information
referred to herein, or for any consequential, special or similar damages even if advised of
the possibility of such damages.
There are no third party beneficiaries with respect to this report, and we accept no liability
to any third party. The opinions expressed herein are valid only for the purpose stated
herein and as of the date of this report.
No responsibility is taken for changes in market conditions or laws or regulations and no
obligation is assumed to revise this report to reflect changes, events or conditions, which
occur subsequent to the date hereof.
For more information please contact [email protected] or:

Dan Latimore [email protected]

AMERICAS EUROPE ASIA

USA France Japan

200 Clarendon Street, 12th Floor 28, avenue Victor Hugo The Imperial Hotel Tower, 13th Floor
Boston, MA 02116 Paris Cedex 16 1-1-1 Uchisaiwai-cho
75783 Chiyoda-ku, Tokyo 100-0011
Tel.: +1.617.262.3120
Fax: +1.617.262.3121 Tel.: +33.1.73.04.46.20 Tel: +81.3.3500.3023
Fax: +33.1.45.02.30.01 Fax: +81.3.3500.3059

USA United Kingdom China

1166 Avenue of the Americas 55 Baker Street Beijing Kerry Centre

New York, NY 10036 London W1U 8EW South Tower, 15th Floor
1 Guanghua Road
Tel.: +1.212.541.8100 Tel.: +44.20.7333.8333 Chaoyang, Beijing 100022
Fax: +1.212.541.8957 Fax: +44.20.7333.8334
Tel: +86.10.8520.0350
Fax: +86.10.8520.0349

USA Italy Singapore

Four Embarcadero Center, Suite 1100 Galleria San Babila 4B 8 Marina View #09-07
San Francisco, CA 94111 Milan 20122 Asia Square Tower 1
Singapore 018960
Tel.: +1.415.743.7900 Tel.: +39.02.305.771
Fax: +1.415.743.7950 Fax: +39.02.303.040.44 Tel.: +65.9168.3998
Fax: +65.6327.5406

Brazil Spain South Korea

Av. Doutor Chucri Zaidan, 920 – Paseo de la Castellana 216 Youngpoong Building, 22nd Floor
4º andar Pl. 13 33 Seorin-dong, Jongno-gu
Market Place Tower I Madrid 28046 Seoul 110-752
São Paulo SP 04578-903
Tel.: +34.91.531.79.00 Tel.: +82.10.3019.1417
Tel.: +55.11.5501.1100 Fax: +34.91.531.79.09 Fax: +82.2.399.5534
Fax: +55.11.5501.1110
Switzerland
Canada
Tessinerplatz 5
1981 McGill College Avenue Zurich 8027
Montréal, Québec H3A 3T5
Tel.: +41.44.5533.333
Tel.: +1.514.499.0461