Artificial Intelligence (AI)

powered decision making for

the bank of the future

Pankaj Baid

1
Banks that leverage machine-learning models have potential to increase value by

▪ Banks gain an edge by creating superior customer experiences with end-to-end automation
Stronger customer and using advanced analytics to craft highly personalized messages at each step of the
acquisition customer acquisition journey.

▪ Banks can increase the lifetime value of customers by engaging with them continuously and
Higher customer intelligently to strengthen each relationship across diverse products and services.
lifetime value

▪ Banks can lower costs by automating as fully as possible document processing, review, and
Lower operating decision making, particularly in acquisition and servicing
costs

▪ To lower credit risks, banks can adopt more sophisticated screening of prospective
Lower credit risk
customers and early detection of behaviors that signal higher risk of default and fraud

SOURCE: Multiple Sources from Internet 2

Banks can benefit from organizing their automation efforts around these significant elements

Leveraging advanced analytical / machine learning models for automated, personalized decisions
across the customer life cycle

Building and deploying advanced analytics and machine learning models at scale

Augmenting advanced analytical models with capabilities1 to reduce costs, streamline customer
journeys, and enhance the overall experience

Building an enterprise wide digital-marketing engine to translate insights generated in the decision-
making layer into a set of coordinated messages delivered through the bank’s engagement layer

1 Next generation technologies like Natural Language Processing (NLP), facial recognition, block chain, Robotic process automation and behavioural analytics 3
2 AA – Advanced Analytics, ML – Machine Learning
Technologies which are important today for the large banks (2020 Fintech survey results)

Primary Applications of AI Regional Highlights

• This is consolidated survey
(% responses) (% responses) responses from more than 700
senior decision makers across the
accounting, banking, financial
Predictive Analy... 24% services, investment and insurance
US 28% industries in the United States,
United Kingdom, continental
Data Mi... 23%
Europe and Asia.

Lending/Cr... 17%
UK 28% • About a quarter of all respondents
indicated that they use Predictive
Analytics or Data Mining / Analytics,
Robo Advi... 15%
with respondents from the US and
UK leading the way.
O 8% Continental Eu... 14%
• Meanwhile, about one in seven
(15%) of respondents signaled that
Quant Tra... 7% they are using AI for Robo Advisory,
a new class of tools that manage
14%
assets with minimal human
Biomet... 6%
intervention.

SOURCE: Goodwin’s Fintech 2020, A Global Survey on the State of Financial Technology 4
Analytical techniques for various problems in the banks (Survey results)

Total AI value potential that could be unlocked by problem • This illustrates the relative total
types as essential versus relevant to use cases (%) value of these problem types
Problem Types Sample Techniques across Mckinsey database of use
Essential Relevant
cases, along with some of the
sample analytics techniques that
Classification Logistic Regression 44 29 can be used to solve each problem
type.
Continuous Estimation Linear Regression 37 29
• The most prevalent problem types
are classification, continuous
estimation, and clustering,
Clustering K-Means 16 39 suggesting that developing the
capabilities in associated
techniques could have the widest
Optimization Genetic Algorithms 17 21 benefit
• Some of the problem types that
Anomaly Detection K nearest neighbors 19 6 rank lower can be viewed as
subcategories of other problem
types—for example, anomaly
Ranking Ranking SVM 9 8 detection is a special case of
classification, while
Recommender systems Collaborative Filtering 14 1
recommendations can be
considered a type of optimization
problem—and thus their associated
Data generation Markov Models 07 capabilities could be even more
relevant

SOURCE: Mckinsey 5
Banks should prioritize using advanced analytics (AA) and machine learning (ML) in decisions across the
customer life cycle

Customer Credit Decisioning Monitoring and Deepening Smart servicing

acquisition • Credit qualification collections Relationships • Servicing Personas
• Personalized offers • Limit assessment • Early-warning signals • Intelligent offers • Dynamic customer
• Pricing Optimization • Probability of Default • Churn reduction routing
• Customer retargeting
• Fraud prevention • VAR- based customer • Channel propensity • Real time
• Propensity to buy recommendation engine
segmentation1 • Fatigue rule engine
scoring • AI- enabled agent review
• Agent-customer mapping
• Channel mapping and training

Credit approval
Monthly customer turnaround time, % of
Average days past due, Deposit/AUM3 attrition Net promoter score, cost
acquisition run rate NPA2 rate of servicing
applications approved

Key Metrics

1
VAR is value at risk, 2 Non Performing Asset
6
3
AUM is assets under management
AI is becoming critical as banking frauds are on the rise in the Indian Banks

FICO Credit Score Range: 300 - 850

8% delinquency rate 19.9%

18.2%
• The credit score chart below is
based on FICO’s data and shows
28% delinquency rate what percentage of the population
fall into certain FICO score ranges

16.6% • Delinquency rates are higher

around 61% with consumers having
61% delinquency rate 13%
FICO scores 599 and lower and
10.3% 28% delinquency rates for
consumers having FICO scores
7.6% 9.4% 599-699 and 8% with scores 700-
749 and 3% in the rest
4.9%
• Advanced Analytics is gaining
popularity in domains like fraud
detection, KYC analytics, credit
monitoring and collections in banks

Scores ranging From Very Bad to Excellent

SOURCE: https://www.is650agoodcreditscore.com/fico-credit-score-chart/ 7
Case Study – Low performance risk rating models without advanced analytics should not be allowed into production

• This represents a typical multifactor

High risk customer risk-rating model for the
customers sent retail business of a large North
to enhanced due- 72 American universal bank
diligence units 57 • A manually conducted expert
(disguised real 100 review of the results revealed that
data example), 85 for every 100 customers rated high
indexed to 100 risk, 72 were actually medium to
low risk; furthermore, 57 of every
100 customers rated medium to low
risk by the model proved on review
to have a high-risk profile
High risk ; High risk ; • To put this into perspective, a
customers credit-risk model with this kind of
according
to customer performance would never be
risk rating allowed into production
model

SOURCE: https://www.mckinsey.com/business-functions/risk/our-insights/the-investigator-centered-approach-to-financial-crime-doing-what-matters 8
Case Study – Bank used enhanced data and analytics to dramatically reduce the money laundering activities

• At one large US bank, the false-positive rate in

company transfers

primary team and

secondary team

secondary team
anti–money laundering (AML) alerts was very

Filed as SAR1
Reviewed by
Known intra-

Reviewed by
Total alerts

Closed by
high. The remedial process involved a two-stage

closed
investigation. One team would determine
whether an alert was truly triggered by
suspicious activity. It would eliminate clearly false
positives and pass on the remainder to experts
for further investigation. Very few suspicious-
activity-report filings resulted.
0 90
• The bank rightly felt that this elaborate procedure
and meager result was overtaxing resources. To
Before
improve the specificity of its tests so that AML
enhanced data 100 expertise could be better utilized, the bank
and analytics,% looked at the underlying data and algorithms. It
discovered that the databases incompletely
identified customers and transactions. By adding
10 8 more data elements and linking systems through
2 machine-learning techniques, the bank achieved
a more complete understanding of the
transactions being monitored.
50
• It turned out that more than half of the cases
After enhanced 100 alerted for investigation were perfectly innocuous
data and intracompany transactions. With their more
analytics,% 45 sensitive database, the bank was able to keep
the process from issuing alerts for these
5 3 2 transactions, which substantially freed resources
for allocation to more complex cases

1
Suspicious Activity Report 9
SOURCE: https://www.mckinsey.com/business-functions/risk/our-insights/the-neglected-art-of-risk-detection
Advanced Analytics in Credit Decisioning

Machine learning models to automate the process for

determining the maximum amount a customer may borrow.
Limit These loan-approval systems, by leveraging optical character
recognition (OCR) to extract data from sources such as bank
Assessment
statements, tax returns, and utilities invoices, can quickly
assess a customer’s disposable income and capacity to make
regular loan payments.

Analytics models can use their decisioning capabilities to

quantify the customer’s propensity to buy according to the
Pricing customer’s use of different types of financial products. Some
even leverage natural-language processing (NLP) to analyze
unstructured transcripts of interactions with sales and service
representatives.

Fraud AI-driven credit decisioning can build the business while

Management lowering costs. Sharper identification of risky customers enables
banks to increase approval rates without increasing credit risk.
Advanced analytical models can predict fraud related instances.

SOURCE: EY Global - https://www.ey.com/en_gl/consulting/how-data-analytics-is-leading-the-fight-against-financial-crime 10

The combination of AI and analytics enhances the onboarding journey for each new customer
Name: Joy
Age:32 years
Occupation: Working professional
Propensity-to-buy
model to identify
whom to retarget.
Channel propensity
to identify right
outreach channel
• Joy’s landing page
shows her
personalized offers:
personal loan for
travel, 5% off on • Joy receives a
travel insurance WhatsApp reminder
for personal loan for
travel at zero
Analytics-backed processing charges
hyper personalized
offers based on
customer
microsegment
Family: Married, no children
Profile attribute: Avid traveler
AA-enabled real
time credit
underwriting, limit
assessment, and
pricing
• Joy receives a call
to assist in journey;
caller also informs
Joy about custom
travel services • Joy completes a
streamlined 3-click
journey to see the
offer terms and
Analytics-enabled conditions
customer–caller
mapping with
specific cues
provided to caller
Hyper-personalized
cross-sell and upsell
offers
• Joy completes
online KYC form,
provides details for
employment
verification, and • Loan disbursed to
sets up an online Joy; curated
payment mandate catalogue of offers
ahead of Joy’s
AI capabilities to travel sent by email
conduct relevant
fraud checks (eg,
facial recognition
with KYC docs)
11
Artificial Intelligence in Monitoring & Collections comes to rescue to reduce NPA in banks

AI helps build a 360-degree view of a customer’s financial position, helping banks to recognize early-warning signals that
a borrower’s risk profile may have changed and that the risk of default should be reassessed.

Contact Strategy
▪ To determine an appropriate contact strategy for customers at risk of default, banks can segment
accounts according to value at risk (VAR), which is the loan balance times the probability of default.
▪ This allows banks to focus high-touch interactions on borrowers that account for the highest VAR; banks
can then use low-cost channels like telephoning and texting for borrowers posing less risk.
▪ Banks have used this approach to reduce both the cost of collections and the volume of loans to be
resolved through restructuring, sale, or write-off. (detailed in the next slide)

Treatment Strategy
▪ Customers with high willingness but limited ability to pay in the short term may require restructuring of the loan
through partial-payment plans or loan extensions.
▪ In cases where the customer exhibits both low willingness and limited ability to pay, banks should focus on
early settlement and asset recovery.
▪ Advanced analytics, enabled by unstructured internal data sources such as call transcripts from collections
contact centers and external data sources such as spending behavior on other digital channels, can
improve the accuracy of determinations of ability and willingness to pay.
12
AI and ML can classify customers into microsegments for targeted interventions

Absent True high- Unable to

Customer Type True low-risk Dialer-based Cure
minded touch
• Use least • Use • Match • Focus on • Offer debt
experienced interactive agents to customers restructuring
agents voice customers; able to pay settlements
provided message send live and at high early for
with set (segment prompts to risk of not those who
Description scripts will probably agents to paying truly
self-cure) modify underwater
scripts

Onscreen 10% of time Matching and Added focus Significant

prompts guide saved, prompts can addresses increase in
agent–client allowing for increase higher restructuring &
conversation reassignment sense of probability of settlements
Impact based on of agents to connection default rates in increases
probability of more difficult and this segment chance of
breaking customers likelihood of collecting at
promises and specific paying least part of
campaigns debt

SOURCE: Ignacio Crespo and Arvind Govindarajan, “The analytics-enabled collections model,” McKinsey on Payments, August 2018, McKinsey.com 13
Artificial Intelligence in nurturing customer relationships to maximize customer value

Strong customer engagement is the foundation for maximizing customer value, and leaders are using advanced
analytics to identify less engaged customers at risk of attrition and to craft messages for timely nudges.

▪ As with any customer communication in a smart omnichannel service environment, each

personalized offer is delivered through the right channel according to the time of day.
▪ Deeper relationships are predicated on a bank’s precise understanding of a customer’s
unique needs and expectations. A bank can craft offers to meet emerging needs and deliver
Deepening them at the right time and through the right channel. By doing so, the bank demonstrates that it
understands customers’ current position and aspirations.
Relationships
▪ For example, by analyzing browsing history and spending patterns, a bank might recognize a
consumer’s need for credit to finance an upcoming purchase of a household appliance.
▪ Ping An, for example, has developed a prediction algorithm to estimate the ideal product-per-
customer (PPC) ratio for each user, based on individual needs.

▪ AI-powered decisioning can enable banks to create a smart, highly personalized servicing
experience based on customer microsegments, thereby enabling different channels to deliver
superior service and a compelling experience with interactions that are simple and intuitive.
Servicing and ▪ Banks can support their relationship managers with timely customer insights and tailor-made
offers for each customer. They can also significantly improve agents’ productivity with
engagement streamlined preapproved products crafted to meet each customer’s distinct needs.
▪ Models that analyze voice and speech characteristics can match agents with customers based
on behavioral and psychological mapping. Similarly, transcript analysis can enable prediction
of customer distress and suggest resolution to the agent.

14
Augmented AA/ML models with cutting edge capabilities
The rapid improvement of AI-powered technologies spurs competition on speed, cost, experience, and intelligent propositions.
To maintain its market leadership, an AI-first institution must develop models capable of meeting the processing requirements of
edge capabilities, including natural-language processing (NLP), computer vision, facial recognition, and more.
Some edge technologies already afford banks the opportunity to strengthen existing models with expanded data sets.

AI enabled Use Case Description of an analytical use case in a customer-service journey

• Many interactions with customers — via telephone, mobile app, website, or increasingly, in a branch begin with a conversational
interface to establish the purpose of the interaction and collect the info required to resolve the query or transfer it to an agent.
• A routing engine can use voice and image analysis to understand a customer’s current sentiment and match the customer with a
suitable agent.
• The models underpinning virtual assistants and chatbots employ NLP and voice-script analysis to increase their predictive
accuracy as they churn through vast unstructured data generated during customer-service and sales interactions.
• While each customer-service journey presents an opportunity to deepen the relationship with the help of next-product-to-buy
recommendations, banks should constantly seek to improve their recommendation engines and messaging campaigns (details
on next slide)

Cutting edge capabilities deployed as part of an enterprise strategy to enhance the AI bank’s value proposition
have the potential not only to improve credit underwriting and fraud prevention but also to reduce the costs of
document handling and regulatory compliance

NLP – Natural Language Processing 15

Cutting edge capabilities enhance customer-service journeys
Digital self-cure channels for
customers (eg, WhatsApp,
mobile app, website)
• Customer may seek immediate
resolution of queries through
digital self-service channels
Feedback loop via
engagement channels
• Post-call feedback and
automated follow-up occur
via digital channels
1 Interactive Voice response
2 Natural language processing enabled
Voice-recognition enabled
IVR,1 frontline bots handling
30–50% of queries
• If query is not resolved through
automated channels, customer
may contact bank via chat or
request call with live agent
AI-enabled service-to-
sales engine
• Customers are prompted
for any specific offers
preapproved for them
AI-enabled customer
profiling, customer–agent
matching
• Customer connected to the
appropriate agent (chat or call)
based on type of query and
customer profiling
Voice-analytics and NLP-
enabled2 customer
sentiment analysis
• Contact-center agents supported
by live feedback and prompts to
sustain superior customer
experience
16
Appendix

17