How Indian Banks are Adopting Artificial Intelligence?

ARTICLE

How Indian Banks are Adopting like machine learning, deep learning, unsupervised
learning, and neural networks are often used in
Artificial Intelligence? conjunction with artificial intelligence, they can be
by Shobhit Goel, Dirghau K. Raut, viewed as subsets of the broad field of AI (Chart 1)1. The
rapid rise in computing power combined with the rise
Madhuresh Kumar and Manu Sharma^
of unstructured data has supported the development
Artificial Intelligence (AI) is being increasingly of machine learning and other AI systems. Further, the
rise of smart devices powered by AI and the Internet
adopted in the banking and financial services industry,
of Things (IoT) has in turn created new swathes of
with multiple use cases including fraud detection,
data which can be used to create the next generation
customer segmentation, and chat automation. This
of AI solutions.
article attempts to construct a quantitative measure of
AI adoption in the Indian banking system using a text- Artificial intelligence (AI) and machine learning
mining approach. It also attempts to identify the impact of (ML) have gained prominence, especially with
bank-specific characteristics in driving AI adoption using advancements in generative AI models based on large
a panel fixed effects model. The results indicate that AI language models (LLMs). The adoption of AI and
adoption is gaining momentum led by private banks, with related technologies in the banking, financial services,
asset size and CRAR influencing the rate of adoption. and insurance (BFSI) sector is rapidly transforming
the landscape of financial services both globally and
Introduction in India (RBI, 2023). The banking sector is integrating
The banking and financial sector has a long AI into banking operations to enhance efficiency,
and storied history of adapting to and adopting accuracy, and customer experience, paving the way
technological innovations, starting from the invention for a more innovative and customer-centric financial
of paper currency in the 7th century AD to the ecosystem (BIS, 2024).
introduction of the internet and mobile banking in the Chart 1: Subsets of Artificial Intelligence
21st century. In each case, the industry initially faced Emerging AI technologies

scepticism and challenges in embracing these new AI TECHNOLOGIES ILLUSTRATIVE SOLUTIONS

Computer Vision
technologies, but eventually, they became integral to Virtual
Audio Processing Agents
its functioning and productive efficiencies. Sense
Identity
Analytics
Artificial Intelligence (AI) today can be broadly
defined as the branch of computer science that aims Cognitive
Robotics
Natural Language Processing
to create machines and systems that can perform tasks
Knowledge Representation
which normally require human intelligence, such as Comprehend
Speech
Analytics

reasoning, learning, decision-making, and natural

Recommendation
language processing (Holden, 1991; McCarthy, 1960, Systems

1989, 2007; Russell and Norvig, 2010). While terms Machine Learning
Data
Expert Systems Visualisation
^ The authors are from the Department of Economic and Policy Research, Act
Reserve Bank of India. The authors would like to thank Dr. Snehal
Harwadkar and Sonali Goel for providing data and inputs for the article. Source: Accenture.
The authors are also thankful to Shri Rajib Das and Dr. Sarat Dhal for
their useful comments on the paper. The views expressed are those of the
1 A glossary of terms commonly associated with AI is provided in Annex 1.
authors and do not represent the views of the Reserve Bank of India.

RBI Bulletin October 2024 219

ARTICLE How Indian Banks are Adopting Artificial Intelligence?

The potential adoption and usage of the AI banking sector and the broader financial industry’s
models has necessitated policy discussion and relentless pursuit of efficiency, innovation, and
research around multiple macro financial and enhanced decision-making capabilities, others have
other issues including, quantitatively measuring argued that the adoption of these technologies
the adoption of these technologies, identifying has become a necessity as the wider economy has
the factors driving their adoption, and evaluating embraced them. The banking industry is realising the
the impact of these technologies. While there potential of not only the traditional data analytics and
exists literature on the adoption of AI it is mostly machine learning but it is increasingly recognising the
qualitative in nature. This paper attempts to provide role of new and emerging advanced AI technologies
a quantitative measure of the pace of adoption of AI like deep neural networks and large language models
in Indian banks. It employs text-mining techniques to to derive value in traditional areas like marketing and
analyse 32 Indian commercial banks’ annual reports 2
sales, risk management, finance, and IT. According
for eight years from FY 2015-16 to FY 2022-23 to to McKinskey (2021), the potential for value creation
identify their adoption and usage of AI and related through AI and related technologies in global banking
technologies and their growth in recent years. It is more than US$1 trillion annually mainly in the
then explores the role of bank’s financial health in domain of marketing and sales, and risk management
influencing AI adoption by utilising a fixed effects (Chart 2).
panel data model.
From the broader business drivers’ standpoint,
The rest of the paper is divided into four within the Indian financial sector, the improvement
sections. Section II provides stylised facts such of customer experience, revenue augmentation, and
as applications and business drivers of AI and its the creation of new products – all of which can be
potential value creation in the banking sector. Data broadly seen as endeavours in the realm of marketing
sources and methodology are explained in Section III and sales – stand out as significant motivation for
while outcomes of text mining and empirical results
on drivers of AI adoption are provided in Section IV. Chart 2: Potential Value of AI and Analytics for
Section V concludes the article. Global Banking

II. Artificial Intelligence and the Financial Sector

The digital age has witnessed a monumental

shift in the pace and scale of innovation, with
convergence of big data, advanced algorithms, and
computational power. This rapid evolution of deep
learning, machine learning, and other AI technologies
has been accompanied by its increasing adoption and
integration in the financial sector. While some view
it as a natural evolution in this era marked by the
2 All 12 public sector banks and 20 out of 21 private sector banks have
been considered for analysis. The banks under consideration account for
100 per cent of total assets in the case of public sector banks and 99.3 per
Source: McKinsey (2021).
cent of total assets in the case of private sector banks

220 RBI Bulletin October 2024

How Indian Banks are Adopting Artificial Intelligence? ARTICLE

Chart 3: AI in Indian Financial Services Sector

Sources: PwC India; and FICCI.

the utilisation of AI (Chart 3a). Indian banks are of their trading operations. These strategies are
increasingly focusing on deployment of AI-driven increasingly driven by predictive analytics, sentiment
chatbots and virtual assistants to enhance customer analysis, and other AI-powered tools, with the
experience and provide personalised service (Chart potential to outperform traditional investment
3b) (PwC India and FICCI, 2022). These technologies approaches (Boukherouaa et al., 2021; Goudarzi et al.,
provide customers with instant responses to 2018). Some of the areas for which banks are using
queries, streamline account management, and AI to save cost are: conversational banking, anti-fraud
offer personalised financial advice (Alt et al., 2021; and under-writing3.
Boukherouaa et al., 2021; Goudarzi et al., 2018; Orçun
Financial sector in India is also trying to leverage
Kaya, 2019). AI technologies are reshaping traditional
the AI tools for regulatory compliance. AI solutions
banking and expanding financial services to the
could facilitate compliance by automating reporting,
underserved populations. Mobile banking apps, AI-
monitoring transactions for suspicious activities, better
driven credit scoring, and blockchain-based solutions
understanding of the regulatory requirements and
are examples of some of the powerful and enabling
ensuring adherence to evolving regulatory standards,
vehicles of greater financial inclusion by extending
reducing the risk of regulatory misconducts, penalties
access to banking services to previously unbanked or
and fines (Boukherouaa et al., 2021).
underbanked individuals (Bazarbash, 2019; Gensler
and Bailey, 2021). As seen earlier, AI solutions have huge potential to
enhance risk management capabilities, by undertaking
Banks and FIs are using Machine Learning (ML)
analysis of massive datasets, finding solutions not
algorithms as part of their proprietary trading desks,
available earlier and getting better perspectives on
forex trading desks and even fixed income trading to
3 https://www.businessinsider.in/finance/news/the-impact-of-artificial-
deploy sophisticated trading strategies on a real-time
intelligence-in-the-banking-sector-how-ai-is-being-used-in-2020/
basis, with the goal of increasing the profitability articleshow/72860899.cms

RBI Bulletin October 2024 221

ARTICLE How Indian Banks are Adopting Artificial Intelligence?

Chart 4: Applications of AI by Banks

Source: The Economist Intelligence Unit Survey, 2022.

the real-time market conditions; thereby allowing for to identify potential threats and vulnerabilities (Aziz
more accurate and timely business risk predictions, and Dowling, 2018; Donepudi, 2017; Goudarzi et al.,
contributing to improved portfolio management, 2018; Milojević and Redzepagic, 2021; Orçun Kaya,
loan underwriting, and fraud detection. According 2019).
to a survey of the IT executives in the banking sector III. Data Sources and Methodology
globally, banks are deploying AI based applications
Most of the existing literature on the impact of
for risk management, the most prominent of which
AI/ML on financial sector has limited quantitative
are for fraud detection, optimising IT operations, risk
information on the actual usage and adoption of
assessment and personalising credit scoring (Chart 4).
artificial intelligence, machine learning and other
AI-powered fraud detection systems can identify automation technologies in the Indian banking sector.
suspicious activities and transactions in real-time, The existing literature has mainly adopted three broad
thus offering a robust defense against financial crimes. approaches. The first approach has been a completely
Some experts argue that future of fraud detection qualitative one, where analysis of academic journals,
has to be AI powered as financial transactions are industry reports and leading business publications is
increasingly occurring in the digital realm, making the undertaken to identify use cases of AI in the industry.
battle against fraud more complex (Boukherouaa et al., However, the main shortcomings of this approach are:
2021; Goudarzi et al., 2018; Milojević and Redzepagic, (1) Possibility of researchers’ bias which can lead to
2021; Orçun Kaya, 2019). With the growing volume an overemphasis on certain themes or findings, and
of sensitive financial data, cybersecurity has become underrepresentation of others; (2) Concerns over
a paramount concern. Banks and FIs are increasingly rigour and reliability of reports; (3) Limited ability
adopting AI-powered cybersecurity tools for automated to generalise findings and allow spatial and inter-
anomaly detection and predictive behavioural analysis temporal comparisons (Carter et al., 2021; Rahman,

222 RBI Bulletin October 2024

How Indian Banks are Adopting Artificial Intelligence? ARTICLE

2016). The second approach has been to conduct to successfully develop or operationalise a technology
surveys/questionnaires and provide qualitative solution or may be utilising cost-effective open source
overview of AI adoption or convert the responses into technologies to develop solutions.
a quantitative metric. However, this approach too
In view of this, we propose a text mining-based
suffers from issues like (1) difficulty in capturing the
approach to measure the adoption of AI/ML and
complexity with adoption of a frontier technology like
related technologies in the Indian banking sector.
AI; (2) unreliability and bias as the respondent may
We perform text mining on the annual reports of
not be aware of all aspects or use cases of technology
the Indian banks as they are highly likely to reflect
within the firm and may also overstate and over
the usage and adoption of new-age technology like
emphasise their usage of AI to appear more innovative
AI/ML. Further, the annual reports are a trustworthy
or may understate the response due to privacy
and publicly available source of information and are
concerns; (3) Comparison issues especially across
also available across years thereby allowing for inter-
time which is a very important aspect as technology
temporal analysis.
adoption is a dynamic process and a firm’s use of
AI and other technologies can evolve rapidly (Carter We have adopted a dictionary-based approach
et al., 2021; Survey Research and Questionnaires, after parsing the annual reports to extract individual
n.d.).The third approach has been to use a proxy words/phrases and evaluate the instances of usage of
variable as a measure of AI adoption including using these words. While the earliest work utilised existing
some proportion of IT investment expenditure/capital dictionaries like Harvard word dictionary, subsequent
expenditure/M&A expenditure of technology related work created their own unique dictionaries according
firms. The main shortcomings of this approach are (1) to the specific need. These text mining-based
these metrics measure not just AI based technologies approaches have been used to measure otherwise
but also other types of IT expenditure (2) expenditure hard to evaluate attributes like corporate governance,
may not directly lead to adoption as firms may fail ESG activity and FinTech adoption (Table 1).

Table 1: Literature Utilising Text Analytics

Loughran et al. (2009) Undertook text mining to examine the occurrence of ethics-related terms in firms’ 10-K reports. Found that firms using ethics-
related terms are more likely to be (i) “sin” stocks, (ii) the object of class action lawsuits and (iii) to score poorly on measures
of corporate governance.
Shirata et al. (2011) Adopted a text mining approach to analyse Japanese firms’ financial reports and extracted key phrases/descriptions which
could be able to predict bankruptcy.
Feng et al. (2013) Undertook textual analysis of firms 10-K reports to create a word count based measure of industry competition. This measure
is related to existing industry-level measures of disclosure (e.g., Herfindahl index), but provide additional insight into both
across-industry variation and within-industry variation.
Baier et al. (2020) Undertook textual analysis of firms’ 10-K reports for measuring environmental, social and governance (ESG) activity of the
firms. Undertook a word count based measure using a uniquely created dictionary for ESG related words.
Hong et al. (2023) Constructed a new measure of FinTech adoption by firms, using a word count approach with a uniquely created dictionary for
FinTech on firms’ 10-K reports. Further, examined how product market competition affects FinTech adoption at the firm level
and found that firms operating in more competitive environments have a higher incentive to adopt FinTech.
Sharma et al. (2024) Undertook natural language processing (NLP) based text mining approach to analyse news articles, speeches, interviews and
other publicly available documents for central banks and multi-lateral institutions. They further attempt to decipher the policy
focus and priority of the central banks and multilateral institutions in the areas of FinTech and the shift in these trends over
time.

RBI Bulletin October 2024 223

ARTICLE How Indian Banks are Adopting Artificial Intelligence?

The key step in performing a keyword and which need to be acknowledged9. The quantitative
named entity matching is using a suitable dictionary mentions of AI-related terms in the dictionary-
which contains a list of words or phrases related to based frequency measures may not account for
specific categories or concepts of interest. In line with the contextual meaning (Loughran and Mcdonald,
existing literature, we started by using the popular 2016). The presence of AI related keywords may be
dictionaries/glossary related to AI and ML including indicating discussion about AI related opportunities
those by Google Vertex AI4, Google Developers5, IBM6, and risks rather than its deployment (Chen et al.,
NHS AI Lab7 and Council of Europe8 (RBI, 2023). We 2023). On the other hand, it may fail to account for
utilise the popular LLM models of ChatGPT and Bard
banks that have implemented AI extensively but
to create a dictionary of AI related terms pertaining to
have not emphasised it as frequently in their reports,
financial sector. The advantage of using LLM models
leading to underestimation of AI adoption. Thus, the
to create a dictionary comes from the fact that LLM
methodology captures banks discussion and focus on
models have been trained on an extensive dataset
AI and related technologies (which could be the part of
and therefore better equipped to identify keywords
their preparedness/evaluation of these technologies
related to AI in the context of financial sector. A
before its deployment) and not just the deployment
comprehensive dictionary has been formulated by
combining all these dictionaries. We have employed of these technologies.
a combination approach for creating the dictionary Annual reports for 32 commercial banks10  11 for 8
where both keyword matching and named entity years, starting from FY 2015-16 to FY 2022-23, were
recognition are used. This approach is based on analysed. In the first step, the annual reports of the
searching for presence of AI related keywords such said banks were downloaded in portable document
as “artificial intelligence,” “machine learning,” format (pdf). In the text pre-processing stage, these
“neural networks,” “deep learning,” “data science,” pdf files were checked for machine readability and
and other relevant terms. We also match for named
made machine readable, wherever possible, using
entities related to AI, like AI applications (e.g., natural
techniques like optical character recognition (OCR).
language processing, computer vision), or AI-related
organisations (e.g., ChatGPT). A higher frequency of 9 Alternative topic modelling algorithms, such as Latent Dirichlet
AI-related keywords indicates a higher likelihood of Allocation (LDA) or Non-Negative Matrix Factorization (NMF) were
attempted. However, in our case with banks annual reports which are geared
deployment or focus on AI. Words which may be used toward topics related to financial sector with extremely limited space
devoted to discussion around AI/ML and related technologies, these
in an alternate usage sense in financial sector have techniques did not provide useful results. Other approaches include Support
been removed along with terms which were found to Vector Machines (SVM) and Convolutional Neural Networks (CNN) can
identify context and semantic patters. However, as they require high amount
have negligible mention in the test case to improve of labelled data and were therefore not used. Recent studies have explored
usage of transformers like BERT (Bidirectional Encoder Representations
accuracy and parsimony gains, respectively.
from Transformers), GPTs and other LLMs which excel in contextual
understanding, but they not only require large datasets but also are
As is the case with any empirical methodology, extremely computationally extensive mechanisms. This combined with
this methodology also has certain limitations their black box kind of nature limited their suitability for this study.
10 Public Sector Banks witnessed mergers during the period under

4 consideration. The acquiring entities annual reports which were available

https://cloud.google.com/vertex-ai/docs/glossary
for the whole study period were considered. Thus, 12 public sector banks
5 https://developers.google.com/machine-learning/glossary which existed post mergers have been included in the study.
6 https://www.ibm.com/cloud/architecture/architecture/practices/ 11 A total of 21 private sectors scheduled commercial banks were considered,
cognitive-glossary/ including IDBI Bank which was categorised as a private sector bank in 2019.
7 https://nhsx.github.io/ai-dictionary
One private sector bank had to be excluded as its annual reports were not
8 https://www.coe.int/en/web/artificial-intelligence/glossary machine readable.

224 RBI Bulletin October 2024

How Indian Banks are Adopting Artificial Intelligence? ARTICLE

The machine-readable pdfs were then processed IV. Results

individually, by extracting the text content from each IV.1 Artificial Intelligence Adoption
page and storing it in a data-frame. In the next step,
The results show that initially (FY 2015-16), public
tasks such as lowercasing, removing special characters,
sector banks were proactively considering AI/ML and
tokenisation stop-word removal, and lemmatisation
related technologies, showing a similar mention of AI
were performed to make the text suitable for dictionary
related keywords (AI score) compared to their private
matching. In the last step, keyword matching was done
sector counterparts. However, during the period from
by checking each keyword for its presence in the text
2016 to 2021, the AI-related word count in the annual
data, and the frequency of occurrence was recorded. reports of private sector banks has picked up. This may
The key banking characteristics including total be due to a combination of recognition of additional
assets, CRAR (capital to risk weighted assets ratio), use cases of existing AI based technologies along with
return to total assets, efficiency ratio, GNPA (gross more agility in adopting the newer and advanced AI
NPAs to gross advances) ratio, NNPA (net NPAs to net techniques and models.
advances) ratio and retail lending ratio (gross retail The usage of AI related keywords in the
loans to gross advances) for both public and private annual reports of private sector banks increased by
sector banks have been outlined in Table 2. We find approximately six-fold in 2022-23 reports as compared
that there are considerable differences at aggregate to 2015-16 level. Even in case of public sector banks
level between private and public sectors in terms of (PSBs), the emphasis on new age technologies like
these banking characteristics. AI in their annual reports has increased more than 3

Table 2: Key Banking Characteristics of Public and Private Sector Banks in India
Year Total Assets Capital Ratio Return on Efficiency Gross NPAs to Net NPAs to Net Retail Lending
(in ₹ lakh crore) (CRAR) Total Assets (Cost to Income Gross Advances Advances Ratio
(annualised) Ratio) (Per cent) (Per cent) (Per cent)
Public Sector Banks
2015-16 81.01 11.82 -0.27 53.02 9.83 6.13 17.4
2016-17 88.15 12.14 -0.16 50.26 12.47 7.39 19.6
2017-18 91.79 11.66 -0.94 52.32 15.52 8.58 22.4
2018-19 93.73 12.20 -0.77 55.18 12.25 5.12 23.4
2019-20 99.88 12.85 -0.29 53.33 10.79 4.00 25.0
2020-21 109.13 14.04 0.29 51.46 9.36 3.23 26.8
2021-22 119.87 14.62 0.55 53.09 7.57 2.33 27.8
2022-23 130.39 15.53 0.75 51.77 5.22 1.34 29.2
Private Sector Bank
2015-16 28.91 15.68 1.64 44.11 2.70 1.27 24.8
2016-17 34.39 15.53 1.45 43.19 3.51 1.84 25.5
2017-18 41.17 16.43 1.27 43.83 4.01 1.97 25.4
2018-19 51.17 16.07 0.82 46.60 4.81 1.89 26.8
2019-20 56.77 16.55 0.51 43.66 5.11 1.43 30.2
2020-21 63.45 18.42 1.22 41.50 4.74 1.41 31.7
2021-22 72.40 18.78 1.44 44.83 3.73 0.96 30.5
2022-23 83.01 18.61 1.60 48.38 2.17 0.55 35.2
Source: Database on India Economy, RBI.

RBI Bulletin October 2024 225

ARTICLE How Indian Banks are Adopting Artificial Intelligence?

reduce human interventions (Chart 6). Data analytics

Chart 5: AI Word Count for Public and
Private Sector Banks is another major thrust area with possible usage in
fraud detection and predictive analytics. While cloud
computing and big data continue to be the major
technologies which banks are employing, there is
an increasing recognition of potential usage of the
newer AI and ML technologies like Robotic Process
Automation (RPA), Internet of Things (IoT) and Natural
Language Processing (NLP).

IV.2 Drivers of AI Adoption

A panel fixed effect model was used to evaluate

how the AI score, as measured by text analytics
(explained above in section IV.1), may be influenced
Source: Authors’ calculations. by bank-level characteristics like assets, capital to
risk weighted assets ratio (CRAR), return on assets
times between 2015-16 and 2022-23 (Chart 5). In a few (RoA) and gross non-performing assets (GNPA).
public-sector banks, the enthusiasm towards AI based The data on bank specific variables has been taken
technologies is broadly at par with their private sector from RBI’s Database on Indian Economy. The choice
peers, especially in recent years. of a fixed effects model was made as the time-
invariant unobserved characteristics of each bank
Further, the word cloud reveals interesting
could influence both the AI adoption score and the
insights with most banks focussing on automation
bank-level characteristics. For instance, a bank’s
which may be due to a push for efficiency gains and
management quality or corporate culture could
influence its AI adoption score. By using a fixed
Chart 6: Word Cloud of AI Related Keywords
Mentioned in Annual Reports effects model, we can control for these unobserved
characteristics and isolate the relationship between AI
adoption and the observed bank-level characteristics.
Before estimating the model, we performed two
diagnostic tests: the Hausman test and the Redundant
Fixed Effects test. The Hausman test is used to decide
between a fixed effects model and a random effects
model. The null hypothesis under this test is that
the preferred model is random effects. However, in
our case, we rejected the null hypothesis, suggesting
that the fixed effects model is more appropriate. The
Redundant Fixed Effects test checks whether the fixed
effects (or individual-specific effects) in the model are
Source: Authors’ calculations.
significant. The null hypothesis of this test is that

226 RBI Bulletin October 2024

How Indian Banks are Adopting Artificial Intelligence? ARTICLE

these effects are redundant. Again, we rejected the RoAit–1 is return on assets expressed in percentage
null hypothesis, indicating that these fixed effects are which is a measure of how productively the bank has
significant and should be included in our model. been able to utilise its assets and thus a measure of
its capital efficiency in general (Adhitya and Sembel,
Estimation results of panel fixed models are
2020; Chhaidar et al., 2023; Doran et al., 2022).
provided in Table 3. In the baseline model (Model 1),
AI score is regressed upon assets, capital position and GNPAit–1 stands for gross non-performing assets.
dummies representing merger year and private-sector Banks with higher GNPA may derive greater benefits
banks. In the subsequent estimation, we augment from utilising AI related technologies in credit
our model to include other banking indicators such underwriting and recovery process (Bazarbash, 2019;
as GNPA (Model 2), RoA (Model 3), and efficiency Chhaidar et al., 2023; Doran et al., 2022; Goudarzi
ratio (Model 4) and retail lending ratio (Model 5). The et al., 2018; Seth and Bhavika Gandhi, 2023). However,
specification for the fully augmented model (Model 5) banks with higher GNPAs could be more focussed on
is as follows: cleaning books and also have had to allocate more
resources toward provisioning needs, thus reducing
AISCOREit =β0+β1*Ln(Assets)it–1 + β2*CRARit–1 + β3
the resources available for investing in AI and related
*RoAit–1 + β4*GNPAit–1 + β5*EffRatioit–1 +
technologies. In model 2a, net non-performing assets
β6*RetailLendingit–1 + β7 MergerDummyit +
(NNPA) has been taken instead of GNPA as an alternate
β8*PrivateDummyit + αi + εit
specification.
In this equation AISCOREit is the dependent
EffRatioit–1 is efficiency ratio which represents
variable. Ln(Assets)it–1 is the natural log of assets of
the cost to income ratio for a bank. Efficiency ratio can
a bank (in crores of rupees)12 and has been mean-
be indirect proxy for measuring the potential benefits
centered to account for the scale issue. Asset size
of using AI related technologies to reduce costs (Alt
can be viewed as proxy of size of banks and it can
et al., 2021; Boukherouaa et al., 2021; Orçun Kaya,
potentially influence adoption of technologies like
2019).
AI as it affects the amount of resources available, risk
taking capacity, exploitation of economies of scale, RetailLendingit–1 is the retail lending ratio which
potential benefits and even executional challenges represents how retail centric is a bank. With a lot
(Bordonaba-Juste et al., 2012; Burke, 2005; Hall and of AI technologies focussed towards improving the
Khan, 2003; Lee and Xia, 2006; Na et al., 2023). retail customer experience like creation of chatbots,
wealth management solutions, recommendation
CRARit–1 is capital to risk-weighted assets ratio
models among others. It is possible that banks which
and measures the capital adequacy, thus a proxy
have higher proportion of retail loans are more retail
for the financial health of the banks which in turn
banking focussed and thus have higher adoption of AI
indicates availability of capital to afford upfront costs
and related technologies.
of acquiring new technologies which can be significant
in case of complex and capital intensive technologies All the bank specific variables have been taken in
like AI (Chhaidar et al., 2023; Hall and Khan, 2003; lag form to account for the possibility that the impact
Nugroho et al., 2017). of these variables on the AI adoption score may not be
immediate, but rather delayed. Moreover, using lagged
12 Due to merger of some PSBs during the study period, the assets of the independent variables can also help mitigate potential
merged banks have been spliced to remove the impact of merger on asset
size. endogeneity problems. MergerDummyit takes value

RBI Bulletin October 2024 227

ARTICLE How Indian Banks are Adopting Artificial Intelligence?

1 for the bank in the year it has been merged and 0 banks. This finding aligns with resource-based theory,
otherwise, thus allows us to estimate the short-term which posits that organisations with greater resources
impact of the merger13. PrivateDummyit is the dummy are more inclined to invest in innovation and modern
variable taking value 1 for private sector banks and technologies like AI (Barney, 1991) and survey results
0 otherwise, which allows to control for potential showing higher AI adoption rate among banks having
differences between these two categories. αi represents larger asset size14. Further, larger banks owing to their
the individual-specific (bank-specific) effects and εit is difficulties in coordination across verticals are likely
the error term. The subscript i represents the cross- to achieve higher net gains from adoption of such
technologies and data integration, thereby increasing
section (i.e. bank) and t represents the time period.
the motivation for adoption of AI. It may also be
We find consistent results across all the models indicating that the adoption of technologies such as
considered (Table 3). AI is relatively difficult for smaller banks due to larger
Across all models, we find that AI score is fixed cost and absence of economies of scale.
positively related to the asset size of the banks as Capital-to-risk weighted asset ratio (CRAR) which
evident from positive and statistically significant is a proxy for the capital adequacy of the bank and
coefficient, suggesting higher adoption by larger thus a reflection of the financial health of the bank,

Table 3: Factors Impacting Adoption of AI in Indian Banks

Model 1 Model 2 Model 2a Model 3 Model 4 Model 5
Ln_assets 49.34*** 48.62*** 49.09*** 48.40*** 48.72*** 50.13***
CRAR 1.68*** 1.78*** 1.90*** 1.88*** 1.98*** 1.98***
GNPA ratio 0.31 0.19 0.25 0.26
RoA -0.79 -0.04 0.01
Efficiency ratio 0.17 0.20
Retail Lending ratio -0.10
Merger dummy -6.58 -6.70 -6.40 -6.70 -6.19 -6.50
Private dummy 20.51* 17.05 21.56* 17.23 17.06 18.14
NNPA ratio 0.62
Cross-Sections Included 32 32 32 32 32 32
Total Observations 256 256 256 256 256 256
Constant Included Included Included Included Included Included
R2 0.76 0.76 0.76 0.76 0.76 0.76
Adjusted R 2
0.72 0.72 0.72 0.72 0.72 0.72
F-statistic 19.49*** 18.98*** 19.03*** 18.41*** 18.05*** 17.54***
Akaike info 8.39 8.39 8.39 8.40 8.40 8.40
Schwarz 8.88 8.90 8.90 8.92 8.94 8.96
Hannan-Quinn 8.58 8.60 8.60 8.61 8.61 8.62
Durbin-Watson 1.24 1.23 1.24 1.24 1.28 1.29
***, ** and * denotes statistical significance at 1 per cent, 5 per cent and 10 per cent probability level, respectively.

13 As only the acquiree banks were considered in the study for which AI score has been calculated for all years under consideration, a balanced panel was
utilised.
14 https://www.businessinsider.in/finance/news/the-impact-of-artificial-intelligence-in-the-banking-sector-how-ai-is-being-used-in-2020/

articleshow/72860899.cms

228 RBI Bulletin October 2024

How Indian Banks are Adopting Artificial Intelligence? ARTICLE

is positively related to AI score. This result resonates offline channels, especially in rural and semi-urban
with the viewpoint that well-capitalized banks are areas. However, with the rapid advancements in AI,
better positioned to take the investment risks in new especially generative AI and LLM based models in last
technology in terms of adequate capital buffers and 2 years, which have been accompanied with public at
confidence to pursue AI solutions. large being able to access and thus subsequently draw
comfort with AI based solutions, public sector banks
Other bank specific indicators such as RoA,
also appear to be increasing their usage of AI based
GNPA/NNPA, efficiency ratio and retail lending are not
solutions.
statistically significant for explaining AI score15. The
merger dummy is also not found to be a statistically V. Conclusion
significant determinant of AI score of banks. While AI is being increasingly explored in every
The coefficient for the dummy variable industry, its usage is expected to have a profound
representing private sector banks though positive is impact on various operations of banking and finance,
only significant at 10 per cent level in Model 1 and including risk assessment, fraud detection, customer
2a. This result needs to be seen in light of analytical service, investment strategies, regulatory compliance,
findings where both private and public-sector banks and more. As the capabilities of these technologies
broadly started at similar level of AI score in 2015-16 continue to grow, so does their influence on the
but then saw private sector banks AI score increase decision-making processes. AI is expected to have
rapidly in the next 5 years. However, in the last 2 years, the potential to reduce the inefficiencies, through
public sector banks also appear to have enhanced their automation by minimising errors in the human
AI adoption, as observed in their AI score, possibly decision making and by providing cost effective
reflecting the broad-based ongoing expansion of solutions. It is also expected to make banking services
digital technologies in the Indian banking sector. accessible to the population at the bottom of the
pyramid. While the integration of AI into banking
The greater adoption of AI in private sector banks
and finance offers immense opportunities, it also
could be due to a larger proportion of their clientele
presents challenges such as possibility of bias, lack
being better equipped to access digital services and
of transparency, and issues surrounding the ethical
more comfortable with usage of modern technology-
use of data which requires an in-depth evaluation
based solution (ET BFSI, 2021; Malladi et al., 2021).
in view of its implications for financial sector and
Also, private banks often cater to more financially
the overall economy. The recent policy initiatives by
aware and affluent customers and therefore could
India including the National Strategy for Artificial
see higher potential for leveraging AI based solutions
Intelligence, IndiaAI Mission, AI for India 2.0 and Skill
like customer segmentation, robo-advisory, robo-
India AI Portal aim to reap the potential offered by AI
wealth management tools to cross-sell or provide
and related technologies while being cognisance of the
other financial services. Further, private sector banks
risks and challenges presented by it. The Reserve Bank
especially those with a smaller branch network, are
of India has also recognised the potential of AI/ML
much more likely to adopt AI based solutions to
and related technologies and encouraged the banks to
gain new customers or cross-sell different products,
appropriately adopt these technologies for conducting
as it represents a more cost-effective solution. On
ongoing due diligence and effective monitoring for
the other hand, PSBs already have well established
KYC/AML norms16.
15 Results of model incorporating staff expenses and Return on Equity (RoE)
are not shown in the Table 2 but they can be made available on request/ 16 RBI Master Direction - Know Your Customer (KYC) Direction, 2016
demand. (Updated as on January 04, 2024).

RBI Bulletin October 2024 229

ARTICLE How Indian Banks are Adopting Artificial Intelligence?

Text mining of annual reports of Indian banks Bordonaba-Juste, V., Lucia-Palacios, L., and Polo-
during 2015-16 to 2022-23 reveals that both private Redondo, Y. (2012). The influence of organizational
and public-sector banks are increasingly emphasising factors on e-business use: Analysis of firm size.
on AI and related technologies; however, the pace of Marketing Intelligence and Planning, 30(2), 212–229.
increase is higher for private banks. Automation, data https://doi.org/10.1108/02634501211211984/FULL/
analytics, cloud computing and big data are the major XML
thrust areas, with increasing consideration for RPA,
Boukherouaa, E. B., Shabsigh, G., Deodoro, J., Farias,
IoT and NLP like technologies by banks especially in
A., Iskender, E. S., Mirestean, A. T., and Ravikumar, R.
recent years. Among different banking indicators, size
(2021). Powering the Digital Economy: Opportunities
and financial health of a bank is found to positively
and Risks of Artificial Intelligence in Finance. https://
influence the bank’s focus on AI, reflecting the
www.imf.org/en/Publications/Departmental-Papers-
impact of economies of scale and the availability of
Policy-Papers/Issues/2021/10/21/Powering-the-Digital-
investment on the technological advancement.
Economy-Opportunities-and-Risks-of-Artificial-
References Intelligence-in-Finance-494717
Adhitya, A., and Sembel, H. M. R. (2020). the Impacts of Burke, K. (2005). The Impact of Firm Size on Internet
Mobile Banking Technology Adoption on the Financial Use in Small Businesses. Electronic Markets, 15(2),
Performance and Stock Performance of Big Banks in 79–93. https://doi.org/10.1080/10196780500083738
Indonesia. South East Asia Journal of Contemporary
Business, Economics and Law, 22(1), 63–73. Carter, S. M., Shih, P., Williams, J., Degeling, C., and
Mooney-Somers, J. (2021). Conducting Qualitative
Alt, M.-A., Vizeli, I., and Săplăcan, Z. (2021). Banking
Research Online: Challenges and Solutions. Patient,
with a Chatbot – A Study on Technology Acceptance.
14(6), 711–718. https://doi.org/10.1007/s40271-021-
Studia Universitatis Babes-Bolyai Oeconomica, 66(1),
00528-w
13–35. https://doi.org/10.2478/subboec-2021-0002
Chen, J., Henry, E., and Jiang, X. (2023). Is
Aziz, S., and Dowling, M. M. (2018). AI and Machine
Learning for Risk Management. SSRN Electronic Cybersecurity Risk Factor Disclosure Informative?
Journal. https://doi.org/10.2139/ssrn.3201337 Evidence from Disclosures Following a Data Breach.
Journal of Business Ethics, 187(1), 199–224. https://
Baier, P., Berninger, M., and Kiesel, F. (2020).
doi.org/10.1007/s10551-022-05107-z
Environmental, social and governance reporting in
annual reports: A textual analysis. Financial Markets, Chhaidar, A., Abdelhedi, M., and Abdelkafi, I. (2023).
Institutions and Instruments, 29(3), 93–118. https:// The Effect of Financial Technology Investment Level
doi.org/10.1111/fmii.12132 on European Banks’ Profitability. Journal of the
Barney, J. (1991). Firm Resources and Sustained Knowledge Economy, 14(3), 2959–2981. https://doi.
Competitive Advantage. Journal of Management, org/10.1007/s13132-022-00992-1
17(1), 99–120. Donepudi, P. K. (2017). Machine Learning and Artificial
Bazarbash, M. (2019). FinTech in Financial Inclusion Intelligence in Banking. Engineering International,
Machine Learning Applications in Assessing Credit 5(2).
Risk (WP/19/109). Doran, N. M., Bãdîrcea, R. M., and Manta, A. G. (2022).
BIS. (2024). Digitalisation of finance (Issue May). Digitization and Financial Performance of Banking
www.bis.org Sectors Facing COVID-19 Challenges in Central and

230 RBI Bulletin October 2024

How Indian Banks are Adopting Artificial Intelligence? ARTICLE

Eastern European Countries. Electronics (Switzerland), Loughran, T., Mcdonald, B., Yun, H., Loughran, T., and
11(21). https://doi.org/10.3390/electronics11213483 Yun, H. (2009). A Wolf in Sheep’s Clothing: The Use of
Ethics-Related Terms in 10-K Reports. 89(2009), 39–49.
ET BFSI. (2021). Public Sector Bank: How PSU banks
https://www3.nd.edu/~tloughra/WolfSheep.pdf
are catching up in the digital world. Economic Times.
https://bfsi.economictimes.indiatimes.com/news/ Malladi, C. M., Soni, R. K., and Srinivasan, S. (2021).
banking/how-psu-banks-are-catching-up-in-the-digital- Digital financial inclusion: next frontiers—challenges
world/84541392 and opportunities. CSI Transactions on ICT, 9(2), 127–
134. https://doi.org/10.1007/s40012-021-00328-5
Feng, L., Lundholm, R., and Michael, M. (2013). A
measure of competition based on 10-k filings. Journal McCarthy, J. (1960). Programs with common sense.
of Accounting Research, 51(2), 399–436. https://doi. McCarthy, J. (1989). Artificial Intelligence, Logic and
org/10.1111/j.1475-679X.2012.00472.x Formalizing Common Sense. In Philosophical Logic
Gensler, G., and Bailey, L. (2021). Deep Learning and and Artificial Intelligence (pp. 161–190). Springer
Financial Regulation. SSRN Electronic Journal, 1–45. Netherlands. https://doi.org/10.1007/978-94-009-2448-
2_6
https://doi.org/10.2139/ssrn.3788662
McCarthy, J. (2007). From here to human-level AI.
Goudarzi, S., Hickok, E., and Sinha, A. (2018). AI in
Artificial Intelligence, 171(18), 1174–1182. https://doi.
Banking and Finance. In The Centre For Internet and
org/10.1016/j.artint.2007.10.009
Society.
Milojević, N., and Redzepagic, S. (2021). Prospects
Hall, B. H., and Khan, B. (2003). Adoption of a New
of Artificial Intelligence and Machine Learning
Technology (No. 9730; Issue May). https://doi.
Application in Banking Risk Management. Journal of
org/10.4018/978-1-5225-7086-8.ch001
Central Banking Theory and Practice, 10(3), 41–57.
Holden, C. (1991). Machines Who Think (Vol. 254, https://doi.org/10.2478/jcbtp-2021-0023
Issue 5036). A K Peters. https://doi.org/10.1126/
Na, S., Heo, S., Choi, W., Han, S., and Kim, C. (2023).
science.254.5036.1291-b
Firm Size and Artificial Intelligence (AI)-Based
Hong, L., Nikbakht, E., and Zhou, T. (2023). Does Technology Adoption: The Role of Corporate Size in
product market competition affect the adoption of South Korean Construction Companies. Buildings,
FinTech by non-financial firms? Finance Research 13(4). https://doi.org/10.3390/buildings13041066
Letters, 54(August 2022), 103730. https://doi. Nugroho, M. A., Susilo, A. Z., Fajar, M. A., and
org/10.1016/j.frl.2023.103730 Rahmawati, D. (2017). Exploratory Study of SMEs
Lee, G., and Xia, W. (2006). Organizational size and IT Technology Adoption Readiness Factors. Procedia
innovation adoption: A meta-analysis. Information Computer Science, 124, 329–336. https://doi.
and Management, 43(8), 975–985. https://doi. org/10.1016/j.procs.2017.12.162
org/10.1016/J.IM.2006.09.003 Orçun Kaya. (2019). Artificial Intelligence in Banking
Loughran, T., and Mcdonald, B. (2016). Textual Industry. Deutsche Bank Research. https://doi.
Analysis in Accounting and Finance: A Survey. Journal org/10.1002/9781119710301.ch19
of Accounting Research, 54(4), 1187–1230. https://doi. PwC India, and FICCI. (2022). Uncovering the ground
org/10.1111/1475-679X.12123 truth: AI in Indian financial services. February, 1–36.

RBI Bulletin October 2024 231

ARTICLE How Indian Banks are Adopting Artificial Intelligence?

Rahman, M. S. (2016). The Advantages and Sharma, M., Raut, D. K., Goel, S., and Kumar, M. (2024).
Disadvantages of Using Qualitative and Quantitative Evolution of FinTech and Central Banks : A Text
Approaches and Methods in Language “Testing and Mining- Based Survey. RBI Bulletin, LXXVIII(August),
Assessment” Research: A Literature Review. Journal 131–148.
of Education and Learning, 6(1), 102. https://doi. Shirata, C. Y., Takeuchi, H., Ogino, S., and Watanabe,
org/10.5539/jel.v6n1p102
H. (2011). Extracting key phrases as predictors
RBI. (2023). Report on Trend and Progress of Banking of corporate bankruptcy: Empirical analysis of
in India 2022-23. annual reports by text mining. Journal of Emerging
Technologies in Accounting, 8(1), 31–44. https://doi.
Russell, S. J., and Norvig, P. (2010). Artificial
org/10.2308/jeta-10182
Intelligence: A Modern Approach (Vol. 4). Prentice
Hall. https://doi.org/10.1109/ICCAE.2010.5451578 Survey Research and Questionnaires. (n.d.). Research
Connections. Retrieved November 3, 2023, from
Seth, S., and Bhavika Gandhi. (2023). Revolutionizing
https://researchconnections.org/research-tools/data-
risk assessment and lending decisions: The impact
collection/survey-research-and-questionnaires
of digitalization and data analytics in Indian
banks. In Advancements in commerce, economics The Economist Intelligence Unit. (2022). Banking on a
and management: A research compilation (Issue game-changer: AI in financial services. The Economist,
September, pp. 111–118). 1–9.

232 RBI Bulletin October 2024

How Indian Banks are Adopting Artificial Intelligence? ARTICLE

Annexure
Annex 1: Glossary of Major AI Related Terms
Term Definition
Algorithm A set of rules that a machine can follow to learn how to do a task.
Artificial AI refers to the simulation of human intelligence processes by machines, especially
Intelligence (AI) computer systems to perform cognitive tasks like thinking, perceiving, learning, problem
solving and decision making. AI can have ‘communication’ or ‘decisions making’ similar
to human.
Big data Datasets that are too large or complex which may be structured, semi-structured and
unstructured data that can be mined for information and used in machine learning
projects, predictive modelling and other advanced analytics applications.
Chatbot Chatbot is a program designed for communicating like humans with users/people through
text or voice command.
ChatGPT ChatGPT stands for ‘Chat Generative Pre-Trained Transformer’, which is a large-language-
model based AI chatbot that uses natural language processing to create humanlike
conversational dialogue. The language model can respond to questions and compose
various written content, including articles, social media posts, essays, code and emails.
Computer Vision An interdisciplinary scientific field that deals with how computers can gain high-level
understanding from digital images or videos.
Data Mining The process of discovering patterns in large data sets involving methods at the intersection
of machine learning, statistics, and database systems.
Data Analytics Data analytics is the science of analyzing raw data through tools, technologies, and
processes. It is used to summarise data and to find trends pattern or to identify anomaly
to improve decision-making, and foster business growth. Some types of data analytics
such as prescriptive analytics and cognitive analytics are associated with AI/ML models
such as predictive modelling, deep learning and natural language processing.
Data Science An interdisciplinary field that uses scientific methods, processes, algorithms and systems
to extract knowledge and insights from structured and unstructured data.
Deep Learning A subset of machine learning which is based on artificial neural networks that has networks
capable of learning complex patterns and relationships within data. Deep learning is a
branch of machine learning. It is capable of learning.
Internet of It is a network of devices facilitating communication such as exchanging data with other
Things (IoT) devices/clouds and systems over the internet.
Machine Machine Learning is a type of artificial intelligence (AI) that allows software applications
Learning (ML) to become more accurate at predicting outcomes without being explicitly programmed
to do so. It involves the use of algorithms to parse data and learn from it and making a
determination or prediction. The machine gets “trained” using large amounts of data and
algorithms, and in turn gains the capability to perform specific tasks.

RBI Bulletin October 2024 233

ARTICLE How Indian Banks are Adopting Artificial Intelligence?

Term Definition
Natural Language A subfield of linguistics, computer science, and artificial intelligence concerned with the
Processing (NLP) interactions between computers and human language. It is the ability of computers to
understand text and spoken words like human being.
Neural Network A neural network is an adaptive system that learns by using interconnected nodes or
neurons in a layered structure that resembles a human brain. A neural network can learn
from data, so it can be trained to recognize patterns, classify data, and forecast future
events.
Predictive The use of data, statistical algorithms and machine learning techniques to identify the
Analysis likelihood of future outcomes based on historical data.
Robotic Process The use of software with artificial intelligence (AI) and machine learning capabilities to
Automation handle high-volume, repeatable tasks that previously required humans to perform.
(RPA)
Sentiment The use of natural language processing, text analysis, computational linguistics, and
Analysis biometrics to systematically identify, extract, quantify, and study affective states and
subjective information.
Supervised A type of machine learning where the model is provided with labeled training data.
Learning
Text Mining Text mining is the process of extracting valuable information from unstructured text
data to analyze, understand, and derive insights. It uses techniques of natural language
processing and machine learning.
Unsupervised A type of machine learning where the model is not provided with labeled training data.
Learning

234 RBI Bulletin October 2024