The Impact of AI on the Banking Industry

NOAH KOERSELMAN, University of Twente, The Netherlands

ABSTRACT always kept improving throughout the years. Keeping up with the
Technological innovations are emerging at a rapid pace. Throughout history, newest innovations. These new innovations within the financial
these technological advancements have increased productivity within busi- sector are often called FinTech. This stands for financial technology
nesses. The latest advancements within businesses consist of automation, and can be seen as the new technologies that improve and automate
industrial robots and artificial intelligence. The usage of the latter is on the financial services. Artificial Intelligence, Block Chain, Cloud
the rise in the last few years and is transforming the way of working in a
Computing and Big Data (ABCD) constitute the core of FinTech
wide variety of different industries. One of these industries is the banking
industry. Over the years banks have always improved their methods, by [26]. Banks have integrated these innovative solutions as early as the
applying FinTech solutions, to differentiate themselves and to stay ahead 1960s by introducing ATMs [8]. Adapting to online banking in the
of their competitors. Currently, the new FinTech solution to which banks 2000s and introducing mobile based banking in the 2010s [8]. Now
have to adapt is artificial intelligence, also known as AI. It has the potential banks are innovating through the implementation of AI. The use of
to improve most operations of the banks. In this research, a systematic lit- AI can potentially improve efficiency, reduce cost, reduce errors and
erature review will be performed to find the effect that AI has had on the many more within a bank. The integration in the financial services
banking industry. The findings of the systematic literature review will then sector, of which banks are a big part, is one of the sectors that is
be analysed. furthest advanced with a high percentage of integration of AI within
Additional Key Words and Phrases: Artificial Intelligence, Banking Industry, the companies. AI is integrated by more than 60% of the companies
FinTech, Impact, Algorithms within the financial services sector [11]. With the capabilities of AI
still improving, the banks that already have integrated it into one
1 INTRODUCTION or more processes still have a lot of growth potential using newer
With the start of the new industrial revolution, industry 4.0, in 2011 and better forms of AI.
sectors shifted their priorities to advanced digitalisation [14]. A big In order to deeply understand the impact of AI in the banking
part of the new technologies that were getting implemented was industry a systematic literature review is conducted in this research.
AI. John McCarthy, one of the founders of AI, once defined it as The main research question to be answered is then as follows:
"getting a computer to do things which, when done by people, are
said to involve intelligence" [45]. Even though the idea of AI has What is the impact of Artificial Intelligence on the banking industry?
been around for more than half a century, with the paper of John
McCarthy being published in 1958 [29], only recently it is changing The primary research question is answered by considering the
the job market at an unprecedented pace. The integration of AI by following sub-research questions:
a company in at least one function was only 20% in 2017 but has
SRQ1: What are the potential different use cases for AI in the
increased to 50% in 2022 [13]. This shows the scale at which AI is
banking industry?
transforming companies.
SRQ2: What negative impacts can the implementation of AI in
The recent rise in the usage of AI within companies is due to an in-
the banking industry have?
crease in the possibilities of the different techniques of AI. The field
SRQ3: What methods of AI are used in the Banking Industry?
of AI includes several techniques such as robotics, computer vision,
natural language processing, deep learning and machine learning. The analysis towards these questions can be found later in the
These techniques in general can be described as simulating the paper. The remainder of the paper is structured as follows. Section
thought processes of a human by a computer. The computer uses two presents the methodology of the systematic literature review.
algorithms that allow the computer to make decisions. The decision The results of this systematic literature review are presented in sec-
making of the computer is based on processing vast amounts of data tion three. Section four presents the literature review itself. Lastly,
and learning to recognize patterns within the data sets. This has sections five and six are respectively the discussion and the conclu-
the potential to improve and take over a wide variety of functions. sion.
Among others, AI can help by automating repetitive tasks, mak-
ing business decisions and enhancing customer experience. This 2 METHODOLOGY
ultimately will help a business grow. To perform this systematic literature review the methodology as
These forms of integration of AI can be found in nearly every described below is used. The methodology explains the way this
industry. The banking industry is no exception to this. Banks have research is carried out. It is a systematic plan of collecting, criti-
cally evaluating, integrating, and presenting findings with as final
TScIT 39, July 7, 2023, Enschede, The Netherlands goal to answer the main research question with its sub-research
© 2023 University of Twente, Faculty of Electrical Engineering, Mathematics and
Computer Science.
questions [36]. Explaining how the data is collected and analysed
Permission to make digital or hard copies of all or part of this work for personal or through the use of this methodology helps to ensure the reliability
classroom use is granted without fee provided that copies are not made or distributed of this research. To start this methodology an overview is given of
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. To copy otherwise, or republish, to post on servers or to redistribute the databases that have been used during this research including a
to lists, requires prior specific permission and/or a fee. short description of the characteristics of these specific databases.

1
TScIT 39, July 7, 2023, Enschede, The Netherlands Noah Koerselman

Databases with a broad search field are used to capture as many Table 1. Results of the initial search strings SRQ1
relevant articles as possible.
Database Search String Results
Overview of the databases: Google Scholar AI Banking Industry Use Cases 17,900
• Google Scholar1 : a database used for broad searches that (2014-2023)
shows scholarly articles from different publishers. AI Banking Industry Possibilities 17,400
• Web of Science2 : a large database suitable for searching arti- (2014-2023)
cles from high-impact research journals. AI Banking Industry Applications 17,100
• Scopus3 : a database for searching research literature in the (2014-2023)
science and technology field. Web of Science TS=((AI AND Banking AND 33
Industry) AND ( Possibilities OR
2.1 Procedure Use Cases OR Applications)) AND
PUBYEAR AFT 2013
By using a specific set of keywords to search in these databases,
Scopus TITLE-ABS-KEY(ai* AND banking 458
this literature review was conducted. The procedure used for this is
AND industry AND (possibilities*
described in this section.
OR (use AND cases) OR
In their study Xiao and Watson [59] established eight common
applications*)) AND PUBYEAR >
steps to conduct a systematic literature review. These steps are as
2013
follows:
(1) formulating the research problem; Table 2. Results of the initial search strings SRQ2
(2) developing and validating the review protocol;
(3) searching the literature; Database Search String Results
(4) screening for inclusion; Google Scholar negative impact artificial 18,500
(5) assessing quality; intelligence (2014-2023)
(6) extracting data; limitations artificial intelligence 17,900
(7) analyzing and synthesizing data; (2014-2023)
(8) reporting the findings. disadvantage artificial intelligence 121,200
Using these steps this procedure is established. A list of studies (2014-2023)
was retrieved by entering a specific search string in each of the Web of Science TS=(((negative AND impact) OR 6,747
databases. This was done for each of the research questions. The limitations OR disadvantage) AND
initial search string consisted of a combination of keywords to Artificial AND Intelligence) AND
find specific results. The initial search strings for every research PUBYEAR AFT 2013
question per database and the number of papers that it produced Scopus TITLE-ABS-KEY(((negative AND 14,032
can be found in tables 1, 2, 3 and 4 for the research questions 1, 2, 3 impact) OR limitations OR
and 4 respectively. Not all papers were found using the initial search disadvantage) AND artificial AND
string. After using the initial search string further papers were found intelligence) AND PUBYEAR >
by refining the initial search string or by using the reference list 2013
from initially found papers. This method of using the reference list
of papers to find new papers is called snowballing [58]. Table 3. Results of the initial search strings SRQ3
To find relevant research the databases were scanned based on
the titles of the research. In case the title seemed to be relevant Database Search String Results
to the topic, the abstract of the research and the keywords were Google Scholar Artificial Intelligence methods 16,600
read. After reading the abstract the conclusion was read in case banking (2014-2023)
the research still seemed to be relevant. The relevant studies are Artificial Intelligence techniques 18,000
included in this literature review. A database shows the research banking (2014-2023)
that is the most closely related to the search terms first. At a certain Web of Science TS=(((Artificial AND Intelligence) 448
point the relevancy of the research presented by the database drops. OR AI) AND (methods OR
Therefore, it was decided to stop the search using a specific search techniques) AND banking) AND
string after around ten unrelated studies in a row had been found. PUBYEAR AFT 2013
This was done when a combination of the search string and a specific Scopus 446
TITLE-ABS-KEY(((artificial AND
database gave more than a few hundred results. After finding a
intelligence) OR ai) AND (methods
sufficient amount of studies to answer and substantiate the findings
OR techniques) AND banking)
the search towards this topic was ended. After finishing the search
AND PUBYEAR > 2013
1 Google Scholar: https://scholar.google.com/
2Web of Science: https://www.webofknowledge.com/
3 Scopus: https://www.scopus.com/

2
The Impact of AI on the Banking Industry TScIT 39, July 7, 2023, Enschede, The Netherlands

Table 4. Results of the initial search strings RQ

Database Search String Results

Google Scholar 18,300
Impact of AI Bank (2014-2023)
Impact of AI Banking Industry 18,400
(2014-2023)
Web of Science 101
TS=(AI AND Impact AND Bank)
AND PUBYEAR AFT 2013
TS=(AI AND Impact AND Banking 24
AND Industry) AND PUBYEAR
AFT 2013
Scopus TITLE-ABS-KEY(ai AND impact 103
AND bank) AND PUBYEAR >
2013
TITLE-ABS-KEY(ai AND impact 41 Fig. 1. The distribution of the papers over the research questions
AND banking AND industry) AND
PUBYEAR > 2013
From the inclusion criteria, the following exclusion criteria follow
logically:
towards all the sub-research questions, the main research question
was answered. • Studies that do not directly address one of the research ques-
tions.
2.2 Screening Procedure • Opinion pieces or other forms of subjective articles.
For the screening procedure of the studies, the best practices pro- • Studies that do not meet the inclusion criteria.
posed by Kitchenham [21] were used, to the extent that was possible
for this study. Rules to include and exclude studies help to filter 3 RESULTS SYSTEMATIC LITERATURE REVIEW
unwanted studies from the vast amount of studies. The inclusion The search phase of the systematic literature review resulted in
criteria that were applied during the literature review were: many papers. Most of the papers were not considered relevant for
• Research Alignment this research. The number of papers that were considered relevant
– Studies that explore the different use cases of AI in banking. was 49, these were found either by the initial search strings or by
– Studies that focuses on the negative impacts of AI in bank- snowballing. The paper selection mechanism can be found in Figure
ing. 2.
– Studies that researched the effectiveness of different AI Using Google Scholar 24 relevant papers of good quality were
methods for different use cases in banking. found.
– Studies that directly address the impact of AI in banking. Web of Science contained 19 relevant papers. Among these papers
• Publication Type were 5 duplicates of what was found using Google Scholar. The other
– The studies should be peer-reviewed. papers were considered to have new added value.
– The studies should be a research paper, article or other Scopus contained 12 relevant papers. One of these papers was
scholarly document. already found using the other databases the rest of the papers were
• Study Design new relevant papers.
– Qualitative studies addressing one of the research ques- The papers that were found helped to answer the research ques-
tions, for example a case study or systematic literature tions. The way the papers are distributed over the different research
review. questions can be found in Figure 1.
– Quantitative studies addressing one of the research ques-
tions, for example a benchmarking study, statistical analy-
4 FINDINGS
sis, experiment or survey. By following the methodology the information to answer the re-
– Qualitative studies addressing one of the research ques- search questions was found. In this section, these findings of the
tions. systematic literature review are synthesized. It provides a summary
– Studies that make use of both techniques. of the current state of knowledge on this topic.
• Time Frame
– The publication date of the studies is after 2013. 4.1 Use Cases for AI in the Banking Industry
• Language A wide variety of use cases for which AI can potentially be adopted
– Studies that are written in English. in the banking industry was found [1, 28, 51, 56]. It was chosen
to describe here the applications of AI that have or can have a

3
TScIT 39, July 7, 2023, Enschede, The Netherlands Noah Koerselman

substantial role in the daily banking industry. Since these use cases
are spread over a wide variety of different tasks they need to be
classified. It was found that in most studies the different tasks can
be classified into the following four broad categories [1, 22, 51]:
• Customer Services
• Financial management
• Fraud Detection
• Compliance with Regulations

4.1.1 Customer Services.

Chatbot
A number of potential tasks that can be fulfilled by the use of AI can
be classified in the customer service category. The most obvious is
the use of chatbots. A chatbot is a computer program that helps a
customer by answering questions a customer has in a human-like
way through a chatting format, using AI techniques such as Natural
Language Processing (NLP) [5]. The introduction of chatbots by
a bank helps ease the banking process for the employee [43]. A
chatbot advisor has multiple advantages over a human advisor. So
has a chatbot the advantage on the factors of price, convenience, ef-
ficiency and 24/7 access [32, 37]. By being available round the clock
giving instant answers to customers the use of a chatbot reduces
the turnaround time (TAT) and increases the efficiency within the
customer service process [32]. Still, human advise is preferred when
it comes to accountability and effectiveness [37]. For this reason
in case a chatbot does not provide the answers a customer wants,
the customer can in most cases still be redirected to someone from
customer service.

Customer Onboarding
Another important process in which AI can play an important role is
the customer onboarding process. This process includes all activities
that are involved in introducing a new customer to the bank. This
includes among others data and document collection, identification
checks and know your customer (KYC) procedures. This can be a
time consuming process, which can take up to twelve weeks [20].
The steps taken in this process are crucial as they can make or break
the ongoing relationship with the customer. Introducing AI can
speed up this process and increase the security to start the new
relationship with the customer on good terms.

4.1.2 Financial management.

Credit Risk Management
Credit risk management is one of the cornerstones of the banking
industry. In this process, the creditworthiness of the customer is
determined. This is a prediction of the possibility that the customer
will pay back the loan. By determining this creditworthiness the
bank can based on the results decide if credit should be extended
to the customer. These decisions that are made by bank staff are
predominantly based on financial information, as was found by
Trönnberg and Hemlin [52]. Even though decisions were predomi-
nantly based on financial information, they also found that decisions
were also regularly intuitive based. In the traditional credit risk man-
agement methods mainly the experience and the judgment level of

4
Fig. 2. paper selection mechanism
The Impact of AI on the Banking Industry TScIT 39, July 7, 2023, Enschede, The Netherlands

the employee determine the outcome [19]. Using a set of variables to read the new compliance requirements and adapt the processes
and past data these predictions can also be made by an algorithm. accordingly.
For this AI can be used. In recent studies is already been found that
AI models are performing better than the traditional models [55]. 4.2 Limitations of the use cases
One of the main obstacles to the wider adoption of AI models in While the banking industry can benefit from AI in the different
credit risk management is the lack of transparency of the models use cases mentioned above it can also impact a bank negatively.
[18]. AI has its limitations and disadvantages. It is important for a bank
to understand this so a proper infrastructure can be set up for the
Asset Management implementation of AI. This way a bank gets the most out of the
An asset manager of a bank is responsible for managing the in- functionalities of the implemented AI.
vestments of a client, with the goal to increase the client’s assets. Some general disadvantages were found that apply to all im-
This is done by making investment decisions based on analyses and plementations of AI that are applicable in the banking industry
research of the market. The decisions that are made are based on [7, 12, 35]. These disadvantages are:
the objectives of the client. Instead of an asset manager, AI can be
used to make these decisions. AI can automate the process of asset • The high costs of the implementation;
management by analysing the data and making trading decisions • the high costs of maintenance and repair;
based on this. The use of AI in asset management does not mean that • potential bugs in the system;
asset managers are unnecessary. Asset managers are still needed to • the requirement of technical expertise.
monitor the AI to make sure that it is in line with the objectives of Looking at the four earlier mentioned categories there are also
the client [10]. some specific disadvantages of AI. A specific disadvantage of the
use of chatbots is the potential inaccuracy of some answers. The
Document Processing chatbots are trained with a specific data set, if this contains data
Another process that AI can potentially automate within banks is that is not accurate the chatbot will provide inaccurate answers [30].
that of document processing. This is the process of scanning a doc- Further, chatbots have difficulties with handling complex queries
ument to find important information that can be used for further [30]. Therefore customer service should still have employees since
business processes. Alhaddad [31] defines two possible methods for AI cannot be expected to handle all possible queries of the customers.
this in his study. In the first method, this process is executed by an Using AI in asset management also has its limitations and poten-
employee, reading through these documents and then manually in- tial risks. When using AI in the asset management these should be
putting the important information into the bank’s individual portals. taken into account. A big potential risk of using AI when making
The second method represents the use of AI to automatically extract trading investments is over-fitting, using too much known infor-
important data from the documents that are then forwarded to the mation to create the algorithm [34]. In more dept, this concept of
bank’s portals. The use of AI as mentioned in the second method is over-fitting means that the algorithm is trained to exactly match
more efficient compared to the first method. the training data, with the result that the algorithm is not accurate
for unseen data. If this happens the AI is useless for making trading
4.1.3 Fraud Detection. decisions. Using historical data to train these algorithms also makes
AI technologies also have possibilities in the financial crime section them particularly ineffective during a crisis [6]. The reason for a
for banks. So can AI be used in the fraud detection, tracking down crisis cannot always be traced back to historical data. The rise of
credit card fraud and money laundering. AI can be used to identify the crisis can be quite sudden. This makes the market more volatile
and block these fraudulent activities in real-time and predicting and more unpredictable than normal.
future possible suspicious actions [47]. The data set consisting of Risks of the implementation of AI can also be found in financial
the vast amount of credit card transactions that have been made crime protection. A big risk is that a bank can lose its credibility
helps to train the algorithm. These transactions can then be used when customers are mistakenly seen as fraudsters [39]. This type
to categorize the credit card holder in fraudulent usage or normal of mistake is also known as a type one error or a false positive. This
usage. Where the usage of AI to detect credit card fraud has matured mistake can be made by the system due to a racial bias within the
over the last years, still a lot of profit can be made in the detection algorithm. A type two error can also happen when a fraudster is
of money laundering and terrorism financing [54]. not detected [39]. The risk of bias is not only relevant for fraud
detection systems. This can also happen when AI is used to give
4.1.4 Compliance with Regulations. credit risk scores, to determine if credit should be extended to the
The use of technology in this area is often called RegTech which is customer.
a portmanteau of Regulatory Technology. RegTech helps a bank by Also for the use of AI for regulatory requirements, the general
managing regulatory requirements by using automated processes, disadvantages apply. Further, the implementation of these RegTech
which improves the reporting capabilities, increases compliance and solutions brings multiple challenges. One of these challenges is
reduces costs [57]. Further, using AI in the area of RegTech can help the inconsistency in regulations between different national and
by monitoring the bank’s processes [3]. Banks need to adapt their international regulators [49]. Another challenge is that of replacing
processes to the changes in regulations and since the regulations are legacy systems that do not integrate with the new technologies [49].
changing frequently banks are constantly adapting. AI can be used The implementation of RegTech is often delayed due to these legacy

5
TScIT 39, July 7, 2023, Enschede, The Netherlands Noah Koerselman

systems. The replacement of these systems is often expensive and In the field of fraud detection, NNs have proven to be very suc-
time consuming. cessful. Multiple studies researched the use of NNs when it comes
There is also an important risk of using AI that needs to be con- to fraudulent transactions [41, 44, 48]. By using NNs all achieved an
sidered when implementing AI in the processes. That is the risk accuracy of approximately 100%, while still showing great results
of the AI becoming a black box. This means that the way the AI on other performance metrics such as precision, recall and f-score.
operates gets so complex that humans, even data scientists, are un-
able to understand the decision making process. Edelman [2] found 4.3.2 Natural Language Processing. Natural language processing
that the lack of transparency is the main reason people do not feel (NLP) is used to analyse the way humans speak or write. It ingests
comfortable with technology. In case of these opaque technologies, unstructured data in the form of spoken words or text and converts
the trustworthiness is mostly dependent on our trust in the expert. this data into structured data that can be used. This process consists
A second problem with black box AI raises when there are flaws in of several steps. Starting with preprocessing the text by breaking
the decision making process. These flaws are difficult to fix since it it apart into separate sentences. These sentences are then broken
is not easy to understand what is going on [38]. apart into words. Next, the words are brought back to their root
form and the meaning in the text is learned. After preprocessing
4.3 AI methods used in the Banking Industry the computer can process the text dependent on the task NLP has
AI is one of the major fields of computer science and can be seen to fulfill. The most significant use of NLP is in the implementation
as an umbrella term for all the computer programs that act as an of chatbots. Further, NLP also gets used for credit scoring, trading
intelligence. A bank that has implemented AI makes use of multiple and fraud detection where it gets combined with a deep learning
AI methods. The methods offer different approaches to solving the method. Kulkarni et al. [23] have developed a chatbot for banks
problems. By understanding and utilising these methods, banks can using NLP. The chatbot can be asked all bank related questions,
take advantage of the possibilities of these AI methods. The most such as information about loans, pension, account, mobile banking
popular techniques for banks are neural networks, natural language etc. The questions can be asked in different ways. The chatbot then
processing and support vector machines. uses NLP to understand this question and to accurately respond to
give the user the intended information.
4.3.1 Neural Networks. Neural Networks (NN) are inspired by the
way biological neurons signal to one another in the human brain. 4.3.3 Support Vector Machine. A support vector machine (SVM) is
The NN consists of multiple layers. These layers are an input layer, a learning algorithm that is mainly used for classification problems.
one or multiple hidden layers and an output layer. Each layer con- SVM tries to find the optimal decision boundary between different
sists of nodes which are the artificial neurons. These nodes are categories. This boundary is called the hyperplane. The SVM goal
connected with the nodes of the following layer by edges. All nodes is to maximise the margin between the points of the categories
in the input layer get a particular value assigned. Weights are used closest to the hyperplane. The SVM is trained by using a training
in the network to control the strength of the signal that is trans- set with each example belonging to one of the given categories.
mitted between two nodes. The nodes in each layer processes the An SVM requires a lot less input data than a neural network to
information and feeds the processed information to the next layer. perform accurately. SVM can be used for the same problems as
The NN learns by comparing the predicted output with the actual neural networks in the banking industry. So, it can be used for
output. This information is used to reduce the error by transferring credit assessment, making trading decisions and fraud detection.
the information back into the network and adjusting the weights. Baesens et al. [4] studied various classification techniques for
This process is performed until the NN has a certain accuracy. For credit scoring. They found the best performances with SVM and
this process, a large number of data is needed to train the NN to neural network techniques. However, they also stated that simpler
make more accurate decisions. NNs can be used for multiple func- techniques also showed very good results. These simpler techniques’
tions. It can be used for credit assessment, making trading decisions performances were in most cases not statistically different from
and fraud detection. SVM and neural networks. This was later further investigated by
Lessmann et al. [25] concluded that NNs were the most effective Lessmann et al. [25] who also found good performances for SVM,
individual classifier for credit assessment. They also found that in- but not as good as neural networks or ensemble classifiers.
dividual classifiers, including NNs, were outperformed by ensemble By reviewing machine learning techniques Strader et al. [46] con-
classifiers, which make use of various classification algorithms. cluded that neural networks are best used for predicting numerical
Asset managers can also benefit greatly from NNs. Shihao et al. stock market index values, but SVM best fits for predicting a rise or
[17] compared different AI methods for trading. They found that fall of the stock market. In Nti et al. [33] extensive literature review
NNs were the best performing method. Lam [24] compared the per- showed that neural networks and SVM are the most used algorithms
formance of NNs with that of the overall market average return, for stock market predictions. They also found that they are better
the minimum benchmark, and that of the top one-third returns, the performing than other machine learning algorithms. Similar results
maximum benchmark. Here the minimum benchmark represents were found by the review of Kumbure et al. [27].
the highly recommended strategy of diversified portfolios. The re- Raghavan and Gayar [40] compared various machine learning
sults showed that NNs were able to consistently and significantly techniques to detect fraudulent transactions, to find insight into
outperform the minimum benchmark. However NNs were not able which methods companies can best use. They found similar results
to significantly outperform the maximum benchmark. for SVM and neural network methods. While the results of these

6
The Impact of AI on the Banking Industry TScIT 39, July 7, 2023, Enschede, The Netherlands

methods were close, they significantly outperformed other meth- An AI can help with analysing this vast amount of data. Based on
ods. The best results were found by combining SVM with neural these analyses and the trading preferences of the asset manager, for
networks. In other research done by Ryman-Tubb et al. [42], in example, risk tolerance and preferred investment sector, the AI can
which they undertook a complete survey of key published work, make recommendations. AI can also be used to further reduce the
they found neural networks in the top ranked methods, but did not need for human labour. This can be done by taking over the entire
find sufficient evidence to conclude that neural networks were the trading process, from making decisions to actually performing the
best performer. transaction [10]. By increasing the efficiency of the processes AI can
lower the cost of asset management. The vast amount of data that
needs to be processed needs to be checked on quality. Improving the
4.4 The impact of AI on banks quality of the data reduces risks and costs. A main AI technique to
AI has already impacted a wide variety of operations within the improve the quality of the data is NLP. This is done by transferring
banking industry. Changing operations such as customer services, the relevant information into a more understandable report. This has
credit risk management, asset management, document processing led to a 60% reduction in the report-making [16]. When it comes to
and fraud detection. The impact of AI offers new opportunities to the customer experience Byrom [10] mentioned several benefits for
the banking industry. Within this industry, a great number of banks the use of AI. First of all, AI cuts customer fees by reducing the costs
are already benefiting from this new technology. of asset management. A disadvantage of this is that the customer is
AI has impacted the customer service department of banks in a liable for its own portfolio. AI can also be used for customization.
positive way. According to customers AI already has greater prob- AI helps to invest based on customer settings, such as risk tolerance,
lem solving capabilities than humans when it comes to less complex preferred investment sector, income etc. The investments of the
tasks [60]. Still, human customer service is needed since human customer can then be displayed in an AI-based interface.
customer service exceeds AI capabilities when it comes to the more For compliance with regulations, banks can benefit from RegTech
complex tasks [60]. Further, Boustani [9] found, regarding customer solutions. These solutions improve the reporting capabilities, in-
service, that even though customers are satisfied with their expe- crease compliance and reduce costs [57]. The need for these so-
rience with AI and accept AI technologies, they do not accept it lutions to reduce compliance costs is high. In 2022 the RegTech
equally as human customer service. Instead of replacing human spending’s for the year where estimated at $68 billion and are ex-
customer service, AI will be more of an assistance. AI can help the pected to exceed $204 billion in 2026 [50]. Further, it is estimated that
employee with organizing and increasing the efficiency of the pro- in 2026, 50% of the compliance of regulations is done by RegTech
cesses, which increases the employee’s productivity. Also, fewer [50].
employees will be needed when AI gets integrated into the process.
The savings in 2019 for these front office tasks were estimated at
$199 billion [15].
AI and specifically machine learning is used by banks for making 5 DISCUSSION
predictions. Making predictions is important for tasks like credit This systematic literature review showed the impact of AI on the
risk assessment and fraud detection. To evaluate the functionality of banking industry. To broaden the knowledge on this subject the
these machine learning models metrics such as accuracy, precision, review was started by investigating the different use cases for AI
recall and F-measure are used. Van Thiel and van Raai [53] found in the banking industry. This was done by identifying the different
for their credit risk prediction model the following scores for the possibilities of AI for banks. This resulted in multiple papers that
aforementioned metrics: accuracy 0.95, precision 0.96, recall 0.88 classified the different use cases. These papers seem to generally
and F-measure 0.92. Ryman-Tubb et al. [42] found similar results classify the use cases in customer service, financial management,
in their survey of published work about fraud detection models. fraud detection and compliance with regulations. Then the different
These high values cannot be achieved by a human employee alone. use cases were researched to get a better understanding of the role
Therefore the implementation of these AI models will enhance the that AI can play here. Papers were not included if it seemed that they
accuracy, reduce the false positives and increase the efficiency. This would not add new knowledge to the topic. By firstly understanding
ultimately will reduce the costs of the credit risk management and the possibilities of AI in the banking industry the topic could be
fraud detection. Human supervision will still be needed for tasks further investigated.
such as training the model and reviewing the decisions that are It is further important for a bank to understand the limitations
made by the model. The savings for the year 2019 by implementing or possible negative impacts of AI before implementing AI in the
AI for fraud detection and anti-money laundering were estimated different processes. This was done by first searching the negative
at $217 billion and for credit risk management at $31 billion [15]. impacts of AI on banks. This resulted in some general disadvantages
The use of AI in the asset management has three main goals: of AI applications such as the high cost of implementation and main-
achieving higher portfolio returns, lowering the cost of asset man- tenance, the possibility of bugs in the system and the requirement
agement and enhancing the customer experience [10]. AI can help for technical expertise. Making use of the snowballing technique the
to achieve higher portfolio returns by optimizing the investment risk of the black box effect was also found. After having a general
process. Often these investments are investments in stocks. This overview of these disadvantages, the various use cases for AI in
trading process has been digitised over the years. This has resulted banking were investigated for negative impacts. This research only
in a large quantity of data sets that can be analysed by the trader. stated the different negative impacts of AI on the various processes

7
TScIT 39, July 7, 2023, Enschede, The Netherlands Noah Koerselman

in banking. Future research can be conducted to investigate to what methods, the risk of AI here is that customers can mistakenly be
matter this negatively influences the processes within banking. seen as fraudsters, due to for example racial bias.
Since AI contains various techniques, research was conducted on AI in the regulatory compliance’s can be seen in the form of
these techniques and their performances. Most of the investigated RegTech. By using RegTech banks can monitor and streamline reg-
studies showed high popularity and good performances of neural ulatory processes. The difficulty of using RegTech is due to the
networks, NLP and SVM. Therefore these were researched in more inconsistency in regulations. This needs to be considered when
depth. Literature reviews or benchmarking studies were deemed implementing RegTech. When implemented correctly RegTech can
more relevant than studies on a single AI technique. The studies improve the reporting capabilities, increasing compliance and reduc-
showed slightly better results for neural networks than most other ing cost. By using NLP to process the frequently changing textual
methods. Other techniques exist and can be researched in future regulations, RegTech can adapt the processes of the bank accord-
more specific studies. ingly.
With the knowledge gained from researching the sub-research
questions, the main research question was investigated. Based on REFERENCES
the different use cases it was investigated how AI has made its [1] 2017. Artificial intelligence and machine learning in financial services Market
developments and financial stability implications. Financial Stability Board (2017).
impact on the banking industry. [2] 2022. Special Report: Trust in Technology. Edelman (2022). https://www.
edelman.com/sites/g/files/aatuss191/files/2022-10/2022%20Trust%20Barometer%
20Special%20Report_Trust%20in%20Technology%20Final_10-19.pdf
6 CONCLUSION [3] Saqib Aziz and Michael Dowling. 2019. Machine Learning and AI for Risk Man-
agement. Palgrave Studies in Digital Business and Enabling Technologies (2019),
Even though AI is at the root of innovation it already made a signif- 33–50. https://doi.org/10.1007/978-3-030-02330-0
[4] B Baesens, T Van Gestel, S Viaene, M Stepanova, J Suykens, and J Vanthienen.
icant impact on the banking industry. Providing improvements in 2003. Benchmarking state-of-the-art classification algorithms for credit scoring.
the processes of client services, credit risk assessment, asset manage- Journal of the Operational Research Society 54 (2003), 627–635. https://doi.org/10.
ment, crime detection and regulatory compliance’s. AI has helped 1057/palgrave.jors.2601545
[5] K Bala, Mukesh Kumar, Sayali Hulawale, and Sahil Pandita. 2017. Chat-Bot
these different departments by the decision making, automating the For College Management System Using A.I. International Research Journal of
processes and streamlining the operations. Hereby banks achieved Engineering and Technology (2017). https://www.irjet.net/archives/V4/i11/IRJET-
higher efficiencies, lower costs, better customer experience and bet- V4I11367.pdf
[6] Söhnke M Bartram, Jürgen Branke, and Mehrshad Motahari. 2020. ARTIFICIAL
ter decision making. While these benefits outweigh the limitations INTELLIGENCE IN ASSET MANAGEMENT. (2020). https://doi.org/10.2139/ssrn.
of AI, they still need to be considered. These can be the high costs 3692805
[7] Sachin Bhbosale, Vinayak Pujari, and Zameer Multani. 2020. Advantages And Dis-
of AI, potential bugs and the requirement for technical expertise. advantages Of Artificial Intellegence. https://www.researchgate.net/publication/
Specific advantages and disadvantages are dependent on the way 344584269_Advantages_And_Disadvantages_Of_Artificial_Intellegence
AI is used. [8] Suparna Biswas, Brant Carson, Violet Chung, Shwaitang Singh, and Renny
Thomas. 2020. AI-bank of the future : Can banks meet the AI challenge ? McKinsey
In the client services sector, the main impact that is made by & Company September (2020).
AI is the integration of chatbots. The 24/7 availability of chatbots [9] Nada Mallah Boustani. 2022. Artificial intelligence impact on banks clients and
has significantly reduced the turnaround time and increased the employees in an Asian developing country. Journal of Asia Business Studies 16, 2
(2022). https://doi.org/10.1108/JABS-09-2020-0376
customer services processes. The advantages are mainly visible for [10] Joseph Byrum. 2022. AI in Financial Portfolio Management: Practical Consid-
the less complex queries, since chatbots struggle with handling erations and Use Cases. Springer Series in Supply Chain Management 11 (2022),
249–270. https://doi.org/10.1007/978-3-030-75729-8
complex queries. These are often still fulfilled by an employee. The [11] Arif Cam, Michael Chui, and Bryce Hall. 2019. Global AI Survey: AI proves its
main AI method for chatbots is NLP, to process the textual data. worth, but few scale impact. McKinsey & Company November (2019).
In the financial management, the prediction capabilities of AI [12] Ku Chhaya, A Khanzode, and Ravindra D Sarode. 2020. Advantages and
Disadvantages of Artificial Intelligence and Machine Learning: a Litera-
have taken over a large part of the processes. In the credit risk as- ture Review. International Journal of Library & Information Science (IJLIS)
sessment, AI has made its impact by analysing vast amounts of data 9, 1 (2020). https://www.academia.edu/44895767/ADVANTAGES_AND_
in a very short time, to gain insight into the creditworthiness of DISADVANTAGES_OF_ARTIFICIAL_INTELLIGENCE_AND_MACHINE_
LEARNING_A_LITERATURE_REVIEW
the customer. Hereby achieving higher accuracies than traditional [13] Michael Chui, Bryce Hall, and Helen Mayhew. 2022. The state of AI in 2022—and
methods. For asset management, the decision making skills of AI are a half decade in review.
[14] Giovanna Culot, Guido Nassimbeni, Guido Orzes, and Marco Sartor. 2020. Behind
mostly used to predict stock market trends, to advice which invest- the definition of Industry 4.0: Analysis and open questions. International Journal
ments to be made. AI helps the asset manager by achieving higher of Production Economics 226 (2020). https://doi.org/10.1016/j.ijpe.2020.107617
returns, lowering operational costs and enhancing the customer [15] Eleni Digalaki. 2022. The impact of artificial intelligence in the banking sector &
how AI is being used in 2022. https://www.businessinsider.com/ai-in-banking-
experience. These prediction models mainly use neural networks report?international=true&r=US&IR=T
or SVM methods to make a decision. Neural networks show the [16] Sudeep Doshi, Ju-Hon Kwek, and Joseph Lai. 2019. Advanced analytics in asset
best results of the individual methods, but combining methods gets management: Beyond the buzz. McKinsey & Company (2019).
[17] Shihao Gu, Bryan Kelly, and Dacheng Xiu. 2020. Empirical Asset Pricing via
even better results. These significant benefits of AI can only be Machine Learning. The Review of Financial Studies 33, 5 (5 2020), 2223–2273.
achieved if the risk of over-fitting is avoided. Over-fitting can make https://doi.org/10.1093/rfs/hhaa009
[18] Branka Hadji Misheva, Ali Hirsa, Joerg Osterrieder, Onkar Kulkarni, and Stephen
the algorithm useless for making accurate decisions. Fung Lin. 2021. EXPLAINABLE AI IN CREDIT RISK MANAGEMENT. (2021).
The fraud detection can benefit from similar methods as the finan- https://doi.org/10.2139/ssrn.3795322
cial management department. By using deep learning methods such [19] Sihem Khemakhem and Younes Boujelbene. 2017. Artificial Intelligence for Credit
Risk Assessment: Artificial Neural Network and Support Vector Machines. ACRN
as neural networks or SVM, fraudulent transactions can be correctly Oxford Journal of Finance and Risk Perspectives 6 (2017), 1–17. http://www.acrn-
classified. While banks can benefit greatly from AI fraud detection journals.eu/resources/jofrp0602a.pdf

8
The Impact of AI on the Banking Industry TScIT 39, July 7, 2023, Enschede, The Netherlands

[20] Dennis Kinyua. 2020. KYC, Client Onboarding: Leveraging Blockchain Technology. 10.1016/j.gltp.2021.01.006
SSRN Electronic Journal (1 2020). https://doi.org/10.2139/ssrn.3528323 [42] Nick F. Ryman-Tubb, Paul Krause, and Wolfgang Garn. 2018. How Artificial Intel-
[21] Barbara Kitchenham. 2004. Procedures for Performing Systematic Reviews. (2004). ligence and machine learning research impacts payment card fraud detection: A
https://www.inf.ufsc.br/~aldo.vw/kitchenham.pdf survey and industry benchmark. Engineering Applications of Artificial Intelligence
[22] Florian Königstorfer and Stefan Thalmann. 2020. Applications of Artificial Intelli- 76 (11 2018), 130–157. https://doi.org/10.1016/j.engappai.2018.07.008
gence in commercial banks-A research agenda for behavioral finance. Journal of [43] Surajit Sarbabidya and Tama Saha. 2020. Role of Chatbot in Customer Service: A
Behavioral and Experimental Finance 27 (2020), 100352. https://doi.org/10.1016/j. Study from the Perspectives of the Banking Industry of Bangladesh. International
jbef.2020.100352 Review of Business Research Papers 16, 1 (2020), 231–248. https://zantworldpress.
[23] Chaitrali S Kulkarni, Amruta U Bhavsar, Savita R Pingale, and Satish S Kumbhar. com/wp-content/uploads/2020/03/13.-601-Tama-Saha.pdf
2017. BANK CHAT BOT – An Intelligent Assistant System Using NLP and Machine [44] Pranali Shenvi, Neel Samant, Shubham Kumar, and Vaishali Kulkarni. 2019.
Learning. International Research Journal of Engineering and Technology 4, 5 (2017). Credit Card Fraud Detection using Deep Learning. 2019 IEEE 5th Interna-
https://www.irjet.net/archives/V4/i5/IRJET-V4I5611.pdf tional Conference for Convergence in Technology, I2CT 2019 (3 2019). https:
[24] Monica Lam. 2003. Neural network techniques for financial performance predic- //doi.org/10.1109/I2CT45611.2019.9033906
tion: integrating fundamental and technical analysis. Decision Support Systems 37, [45] Shukla Shubhendu and Jaiswal Vijay. 2013. Applicability of Artificial Intelligence
4 (2003), 567–581. https://doi.org/10.1016/S0167-9236(03)00088-5 in Different Fields of Life. 1, 1 (2013), 2347–3878. https://www.ijser.in/archives/
[25] Stefan Lessmann, Bart Baesens, Hsin-Vonn Seow, and Lyn C Thomas. 2015. v1i1/MDExMzA5MTU=.pdf
Benchmarking state-of-the-art classification algorithms for credit scoring: An [46] Troy J Strader, John J Rozycki, Thomas H Root, and Yu-Hsiang John Huang.
update of research. European Journal of Operational Research 247 (2015), 124–136. 2020. Machine Learning Stock Market Prediction Studies: Review and Research
https://doi.org/10.1016/j.ejor.2015.05.030 Directions. Journal of International Technology and Information Management 28
[26] Lei Li, Jingjia Qi, Libo Zhao, and Wei Jiang. 2019. Mobile Finance with Collabora- (2020). https://doi.org/10.58729/1941-6679.1435
tive Application of the Fintech ABCD Technologies. Lecture Notes in Computer [47] Srihari Subudhi. 2019. BANKING ON ARTIFICIAL INTELLIGENCE: OPPORTU-
Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes NITIES & CHALLENGES FOR BANKS IN INDIA. INTERNATIONAL JOURNAL
in Bioinformatics) 11792 LNCS (2019), 201–208. https://doi.org/10.1007/978-3- OF RESEARCH IN COMMERCE, ECONOMICS & MANAGEMENT (2019).
030-30949-7 [48] Altyeb Altaher Taha and Sharaf Jameel Malebary. 2020. An Intelligent Approach to
[27] Mahinda Mailagaha Kumbure, Christoph Lohrmann, Pasi Luukka, and Jari Porras. Credit Card Fraud Detection Using an Optimized Light Gradient Boosting Machine.
2022. Machine learning techniques and data for stock market forecasting: A IEEE Access 8 (2020), 25579–25587. https://doi.org/10.1109/ACCESS.2020.2971354
literature review. Expert Systems With Applications 197 (2022), 116659. https: [49] Fabian Teichmann, Sonia Boticiu, and Bruno S. Sergi. 2023. RegTech – Potential
//doi.org/10.1016/j.eswa.2022.116659 benefits and challenges for businesses. Technology in Society 72 (2 2023). https:
[28] Anil B Malali and S Gopalakrishnan. 2020. Application of Artificial Intelli- //doi.org/10.1016/j.techsoc.2022.102150
gence and Its Powered Technologies in the Indian Banking and Financial In- [50] Harshada Thok. 2022. REGTECH: MARKET FORECASTS, EMERGING TRENDS
dustry: An Overview. IOSR Journal Of Humanities And Social Science (IOSR-JHSS & REGULATORY IMPACT 2022-2026. https://www.juniperresearch.com/
25, 6 (2020), 55–60. https://www.iosrjournals.org/iosr-jhss/papers/Vol.%2025% researchstore/fintech-payments/regtech-market-size-report
20Issue4/Series-6/I2504065560.pdf [51] Ashok Kumar Tiwari and Deepak Saxena. 2021. Application of Artificial In-
[29] DR J Mccarthy. 1958. PROGRAMS WITH COMMON SENSE. (11 1958). https: telligence in Indian Banks. 2021 International Conference on Computational
//stacks.stanford.edu/file/druid:yt623dt2417/yt623dt2417.pdf Performance Evaluation, ComPE 2021 (2021), 545–548. https://doi.org/10.1109/
[30] Siddhant Meshram, Namit Naik, Megha Vr, Tanmay More, and Shubhangi Kharche. ComPE53109.2021.9751981
2021. Conversational AI: Chatbots. 2021 International Conference on Intelligent [52] Carl Christian Trönnberg and Sven Hemlin. 2014. Lending decision making in
Technologies, CONIT 2021 (6 2021). https://doi.org/10.1109/CONIT51480.2021. banks: A critical incident study of loan officers. European Management Journal
9498508 32, 2 (4 2014), 362–372. https://doi.org/10.1016/j.emj.2013.03.003
[31] Musaab Mohammad Alhaddad. 2018. Artificial Intelligence in Banking Industry: [53] Diederick Van, Thiel Advicerobo, and W Fred Van Raaij. 2019. Artificial Intelligent
A Review on Fraud Detection, Credit Management, and Document Processing. Credit Risk Prediction: An Empirical Study of Analytical Artificial Intelligence
https://researchberg.com/index.php/rrst/article/view/37/ai_finance Tools for Credit Risk Prediction in a Digital Era. Journal of Accounting and Finance
[32] Navleen Kaur, Supriya Lamba Sahdev, Monika Sharma, and Laraibe Siddiqui. 2020. 19, 8 (2019). https://doi.org/10.33423/jaf.v19i8.2622
Banking 4.0: ‘The Influence of Artificial Intelligence on the Banking Industry & [54] Bart van Liebergen. 2017. Machine learning: a revolution in risk management
How AI Is Changing the Face of Modern Day Banks’. https://papers.ssrn.com/ and compliance? https://www.capco.com/en/Capco-Institute/Journal-45-
sol3/papers.cfm?abstract_id=3661469 Transformation/Machine-learning-a-revolution-in-risk-management-and-
[33] Isaac Kofi Nti, Adebayo Felix Adekoya, and Benjamin Asubam Weyori. 2019. compliance
A systematic review of fundamental and technical analysis of stock market [55] Diederick van Thiel and W. Fred van Raaij. 2019. Artificial Intelligent Credit
predictions. Artificial Intelligence Review 2019 53:4 53, 4 (8 2019), 3007–3057. Risk Prediction: An Empirical Study of Analytical Artificial Intelligence Tools for
https://doi.org/10.1007/s10462-019-09754-z Credit Risk Prediction in a Digital Era. https://articlegateway.com/index.php/
[34] Marli Damião Abade Nunes. 2021. Automated Trading Systems VS Manual JAF/article/view/2622/2493
Trading in Forex Exchange Market. (2021). https://run.unl.pt/bitstream/10362/ [56] Alice Saldanha Villar and Nawaz Khan. 2021. Robotic process automation in
119886/1/TEGI0514.pdf banking industry: a case study on Deutsche Bank. Journal of Banking and Financial
[35] S.S. Osipov and N.V. Ulimova. 2017. Advantages and Disadvantages of AI. , Technology 2021 5:1 5, 1 (5 2021), 71–86. https://doi.org/10.1007/s42786-021-00030-
77–78 pages. https://paper.researchbib.com/view/paper/116409 9
[36] Debajyoti Pati and Lesa N. Lorusso. 2018. How to Write a Systematic Review of [57] Jane Walshe and Tom Cropper. 2018. Should you be banking
the Literature. Health Environments Research and Design Journal 11, 1 (1 2018), on RegTech? Journal of Securities Operations & Custody (2 2018).
15–30. https://doi.org/10.1177/1937586717747384 https://www.ingentaconnect.com/contentone/hsp/jsoc/2018/00000010/
[37] Kanchan Patil and Mugdha S Kulkarni. 2019. Artificial Intelligence in Financial 00000002/art00010?crawler=true&mimetype=application/pdf
Services: Customer Chatbot Advisor Adoption. International Journal of Innova- [58] Claes Wohlin. 2014. Guidelines for snowballing in systematic literature studies and
tive Technology and Exploring Engineering (2019). https://doi.org/10.35940/ijitee. a replication in software engineering. ACM International Conference Proceeding
A4928.119119 Series (2014). https://doi.org/10.1145/2601248.2601268
[38] Dino Pedreschi, Fosca Giannotti, Riccardo Guidotti, Anna Monreale, Salvatore [59] Yu Xiao and Maria Watson. 2019. Guidance on Conducting a Systematic Literature
Ruggieri, and Franco Turini. 2019. Meaningful Explanations of Black Box AI Review. Journal of Planning Education and Research 39, 1 (3 2019), 93–112. https:
Decision Systems. AAAI Conference on Artificial Intelligence 1213 (2019), 9780– //doi.org/10.1177/0739456X17723971
9784. https://doi.org/10.1609/aaai.v33i01.33019780 [60] Yingzi Xu, Chih Hui Shieh, Patrick van Esch, and I. Ling Ling. 2020. AI
[39] Lisa Quest, Anthony Charrie, and Subas Roy. 2018. THE RISKS AND customer service: Task complexity, problem-solving ability, and usage inten-
BENEFITS OF USING AI TO DETECT CRIME. Risk Journal 8 (2018). tion. Australasian Marketing Journal (AMJ) 28, 4 (11 2020), 189–199. https:
https://www.oliverwyman.com/content/dam/oliver-wyman/v2/publications/ //doi.org/10.1016/J.AUSMJ.2020.03.005
2018/december/Risks_and_benefits_of_AI-RiskJournal2018.PDF
[40] Pradheepan Raghavan and Neamat El Gayar. 2019. Fraud Detection using Machine
Learning and Deep Learning. Proceedings of 2019 International Conference on
Computational Intelligence and Knowledge Economy, ICCIKE 2019 (12 2019), 334–
339. https://doi.org/10.1109/ICCIKE47802.2019.9004231
[41] Asha Rb and Suresh Kumar. 2021. Credit card fraud detection using artificial
neural network. Global Transitions Proceedings 2 (2021), 35–41. https://doi.org/