Artificial intelligence and fintech: An overview of opportunities and risks for banking,

investments, and microfinance

Abstract- Artificial Intelligence (AI) is creating a rush of opportunities in the financial sector, but
financial organizations need to be aware of the risks inherent in the use of this technology.
Financial organizations are integrating AI in their operations: in-house, outsourced, or
ecosystem-based. The growth of AI-based fintech firms has encouraged several mergers and
acquisitions among financial service providers and wealth managers as they grapple with
volatility, uncertainty, complexity, and ambiguity. AI's unique promise of combined cost
reduction and increased differentiation makes it generally attractive across the board. However,
perhaps other than fraud detection, these benefits depend on the scale of an organization. Risk
arises from nonrepresentative data, bias inherent in representative data, choice of algorithms, and
human decisions, based on their AI interpretations (and whether humans are involved at all once
AI has been unleashed). Risk reduction requires a vigilant division of labour between AI and
humans for the foreseeable future. KEYWORDS : artificial intelligence, banking, financial
markets, fintech, microfinance
Artificial intelligence (AI) is all the rage in business and particularly in the finance sector. It is
pitched to drive economic growth through increased efficiency and productivity (Mokyr, 2018).
However, we may have reached a growth plateau due to increasing inequalities, education,
demography, and debt (Gordon, 2018), as well as pressures on our planet (Daly & Farley, 2011).
A combination of technological innovations, including AI, is changing business as it responds to
volatility, uncertainty, complexity and ambiguity in the financial sector, popularized in the
literature as a VUCA world (Millar, Groth, & Mahon, 2018). We mean financial markets (stocks,
bonds, derivatives, interest rates, and foreign exchange) and financial institutions (banks,
insurance companies, payment firms, mutual funds, and microfinance organizations) by the
financial sector. AI revolutionizes the strategy of financial actors as they are forced to become
more open and collaborative.
AI “seeks to make computers do the sorts of things that minds can do” (Boden, 2018). AI's
origins are often traced back to the middle of the 20th century about the three laws of Robotics or
the Imitation Game and Digital Computers (Turing, 1950). Indeed, the invention of the analogue
computer by an American Navy in 1938 and later a digital computer by Konrad Zuse in 1939
was also essential in AI history (Bibel, 2014). The first use of the term “Artificial Intelligence” is
attributed to the title of a workshop organized by Marvin Minsky and John McCarthy
“Dartmouth Summer Research Project on Artificial Intelligence (DSRPAI”) at Dartmouth
College in 1956 (Haenlein & Kaplan, 2019).
Commercial products became available under the umbrella terms of decision support systems,
executive information systems and expert systems. These systems were initially rule-based
systems, whereby both the knowledge and the problem-solving procedures were stored in the
form of rules. These rules were coded in the form of if-then-else statements. In other words, if a
condition is met, then the system executes a certain code or else executes another code. The main
application was for transferring expertise to aid in problem-solving (Das, 2013). The technology
came into the public eye in 1997 when IBM's Deep Blue chess program beat chess grandmaster
Garry Kasparov (Schmelzer, 2019). Although the terms AI, machine learning (ML) and deep
learning (DL) are sometimes used interchangeably, they are not the same. Figure 1 depicts the
relationship between these terms.
Rules-based systems are the earliest AI form. They do not learn over time, so their knowledge is
rigid. In this century, dramatic advances in the fields of ML, neural networks, and data mining
(propelled by advances in storage technology) have all but replaced rulesbased AI in complex
applications with an unwieldy number of rules, unless there is insufficient data for ML
(Tricentis, 2019). Certainly, even a small amount of data, modeled correctly with predictive
analytics, can lead to substantial forecasting accuracy without any use of rules- or machine-
learning-based AI (Herrmann, 2019). However, investors will want to go further and base their
decisions on the large data sets available today, which leads us to ML.
ML is a subset of AI, which uses statistical tools to learn from data and then applies algorithms
to solve problems (Oppermann, 2019). Well-known algorithms include internet search engines,
email SPAM filters, and online shopping recommendations. IBM's Watson is a ML system,
which made worldwide headlines in 2011 when it beat the best players in the US quiz show
Jeopardy.
In 2006, DL emerged as an ML subdiscipline, employing multilayer neural networks. In this
area, most advances have been made in AI. Examples include natural speech recognition,
computer vision, machine translation, bioinformatics, gaming, or self-driving cars (Husain, 2018;
WEF, 2019). When Google's AI is commonly referred to in the media, the reference is mainly to
Google's DL capabilities. Perhaps the most prominent win in the DL space was in 2016 when
Google's DeepMind AlphaGo beat the world champion of Go, Lee Sedol. AlphaGo's multi-layer
neural network was trained from 30 million Go games (Heath, 2018). Today, experts in ML/DL
are among the most sought-after roles in organizations from all industry sectors and government
(Davenport & Patil, 2012).
In the following, we first present the methodology for selecting and reviewing our case studies.
Next, we critically analyze these cases and provide different perspectives on AI's opportunities
and areas of uncertainties/risk. Opportunities and uncertainties/risks are split into separate
sections. The section on opportunities is further divided into banking, investments, and
microfinance. Finally, we draw conclusions.
2. CASE STUDY METHODOLOGY
The academic domain provides insightful studies on emerging trends and technologies of AI.
However, they only provide a limited understanding for financial business practitioners in
relation to business strategies and risks involved with AI (WEF, 2019). Therefore, this research
focuses on brief practitioner case studies on the use of AI in financial businesses.
The case studies are based on reviews of media and annual reports of relevant financial
organizations and a review of the meager peer-reviewed business literature. As a hype bias is
often present in the media, this study incorporates interviews with high-growth and high-value
managers. After initial online contact via LinkedIn, a questionnaire was sent to CEOs/partners of
organizations who deploy or consider an AI strategy in their operations. The following
respondents agreed to contribute to this research:
1. Mr. Brett Matthews, Founder and CEO, My Oral Village
2. Dr. Eunika Mercier-Laurent, Chair of TC12 IFIP (AI) at International Federation for
Information Processing (IFIP)
3. Mr. Vivek Mittal, CEO at VegaMX Inc.
4. Dr. Sudeep Krishnan, Senior AVP—EXL Analytics
5. Mr. Pierre Rinaldi, COO and Co-founder of SESAMm, FinTech Company (Big Data and AI)
6. Mr. Sudhir Kulkarni, President, Digital Solutions

3. OPPORTUNITIES
AI is thriving. Its growth is driven by a set of related technology and market factors (PR
Newswire US, 2019). First, telephones and laptops' computing power is far greater than many
mainframes used by a few of us a half-century ago. Cloud computing, with its utility computing
concept added almost unlimited processing and storage capacity (WEF, 2019). Investors have
always used various kinds of technology to glean market-moving information before their
competitors and the speed of information and communication technologies has enabled those
who can afford technology to have an edge over their rivals.
Second, data has become the single most important asset (The Economist, 2017, May 6).
According to IBM (2016), 90% of the world's data was created in the previous 2 years alone
(Loechner, 2016). This velocity of information generated is bringing its own set of challenges.
For financial businesses, real-time analysis is crucial. Any delay will reduce the value of the data
and its analysis for business. For example, a large Australian bank allegedly failed to report more
than 19.5 million suspicious international funds transfer instructions over nearly 5 years (2019)
Large amounts of data become available: the number of “alternative data” sets leveraged by the
financial industry to generate investment signals is growing exponentially (currently more than
1,500 data sets are on the market). This is due to the global data growth worldwide, specifically
driven now by smart objects and open data economies. These new data sets represent a
tremendous opportunity for assets managers to generate alpha in a differentiating way. (Mr
Pierre Rinaldi, Respondent 5).
The top three trends for AI in financial markets are data intelligence, prediction (combining data
exploration and expert knowledge) and AI based simulators (including constraint propagation).
(Dr. Eunika Mercier-Laurent, Respondent 2).
The above benefits have caused AI to become a policy issue at the international level. Countries
want to be first in the field and capture the market. Many countries already have a national AI
plan for their strategic priorities (Madzou & Shukla, 2019). Global harmonization is shaped by
the USA, China, Canada, and the European Union. Other countries, such as India, are not part of
AI's harmonization because they do not have an official position on AI (Roychoudhury &
Alawadh, 2019). Nevertheless, several counties are taking an active part in driving the AI market
because of its growth.
The advances in AI and digital transformation is changing the traditional way financial markets
operated. Starting from digital personal banking to high-end intelligence in trading there are
multiple ways the high-end computing is helping the industry. (Dr Sudeep Krishnan, Respondent
4)
3.1. Banks and financial service providers
According to Dr Sudeep Krishnan, our fourth respondent, some AI deployments are changing
the way banking works.
Personal and customized banking are mining high volume of data, understanding contexts, and
requirements of bankers. Offers to customers are customized, based on their buying behavior and
specific to every individual. Fraud prevention in personal banking is studying behavioral aspects,
buying contexts and habits, using geographical locations, and automating these detections.
Machine learning algorithms also reduce the cost of dealing with these financial frauds and
speedy detection. (Dr Sudeep Krishnan, Respondent 4)
Some financial service providers are in the lead for building their technology platforms. Just as
the experience of clouds shows that a network of small servers can perform as well, it is expected
that a network of small specialized financial service operators may have more shared data than a
big tech firm, especially if we include data on their bilateral and multilateral transactions within
the network. The network of small fintech firms is also getting formalized into a large corporate
entity through mergers and acquisitions.
Some approaches involve outsourcing, as demonstrated by Amazon and Alibaba. According to
Mr Vivek Mittal (Respondent 3), a key trend is that AI is changing the way the back-office
functions. It will be increasingly outsourced to ensure it is focused on continuous improvement.
Around the world, FIS people are working side-by-side with our clients and partners on projects
that take advantage of the power of emerging technologies such as AI, natural language
processing, robotics process automation, and blockchain (FIS, 2019).
The emergence of ecosystems, mergers and acquisitions have created behemoth enterprises.
Fiserv Inc, a technology provider to banks known for its core banking systems, now has a market
capitalization of $78 billion (November 25, 2019). It reached this position thanks to several
acquisitions over the last three decades. In 2019, Fiserv had absorbed First Data, a card payment
processing firm, for $22 billion. As a second example for 2019, another payment processor,
WorldPay, was acquired by Fidelity National Information Services (FIS), a technology provider
to the financial services sector, for $43 billion. FIS has a market capitalization of $84 billion
(November 25, 2019). A third transaction was Global Payments, which bought TSYS for $22
billion. In their annual report, Global Payment included First Data, WorldPay, and major banks
as its competitors (Global Payments, 2019). Concerning their credit card networks, most of these
firms are acquirers, that is, they are intervening at the merchant level instead of banks that sell
the card to the retail client.
 All these large firms provide services to small banks, and it is in their interest to ensure that
banks survive. Nevertheless, many small banks who are clients of these firms are concerned
about merged entities slowing down their innovation process, causing competitive disadvantages
versus large banks and nimble fintech operators (Crosman, 2019). Indeed, mergers could take
many years for dissimilar systems to become integrated with their legacy systems (Schaus,
2019). Conversely, specialized fintech companies, such as Stripe and Square, are focused and
differentiated (Adams, 2019).
Established organizations fear losing customers to merged banks or banks developing/acquiring
their own financial services. They consider big tech firms (including GAFA) to be competitors
(Fiserv, 2019). Their use of AI is especially in the field of fraud prevention. Figure 2
provides some examples of how financial economy actors are competing by using AI.
We conclude that acquisition strategies cause complexity and ambiguity on the acquirer's
offering. Smaller, specialized fintech may have an advantage. Indeed, according to Sudhir
Kulkarni, AI is being adopted by very specialized firms.
In the three large segments of the financial markets,
there are specific areas that are finding massive adoption of AI/ML
Insurance Value Chain: Underwriting, Risk Management and Marketing
Retail Banking and Alternative Financial Services Value Chain: Market Research, NPD,
Marketing and Sales
Capital Market Value Chain: Investor and Investor Agent Engagement (Mr Sudhir Kulkarni,
Respondent 6). In the next section with look at the investments sector
3.2 | The investments sector
Computers now run the stock market: the role humans play in stock trading has diminished
rapidly over the last five decades. Today, the floor traders shouting at each other in the pits have
been replaced by electronically executed trades. Not only are they faster, but they are also
instantly recorded and added onto audit trails. High-frequency trading provides returns to the
fastest, especially when non-discriminating latency can be guaranteed across an ecosystem
network.
Risk management and trading with processing capabilities to handle structured and unstructured
data help customers, and trading institutes improve their algorithms and handle both frequency
and quality of trading. Knowledge of stock performance, real-time data processing, and self-
learning capabilities improve the overall outcome. This fact is evident from high-end trading
now happening from even cell phones. (Dr Sudeep Krishnan, Respondent 4).
Many security analysts have been replaced by algorithms and passive investment funds that
match the stock market's return. Passive funds tracked by Morning Star are worth more than
those run by humans (The Economist, 2019b, October 5). Within an investment firm such as
BlackRock (a global leader), higher fees come from active funds, but clients move out from
active funds and move into passive funds managed by robots (Tokic, 2018). Some industry
experts argue that ML might reverse this trend only temporarily as insights from ML are copied
by other fund managers as they develop ML capabilities. Then it may become even more
challenging to trade stocks and bonds that outperform their benchmarks. Therefore, ML might
reinforce the trend toward passive investment (Pozen & Ruane, 2019).
At the end of the 20th century, we already saw the appearance of quantitative hedge funds that
scan the market for data. However, only human managers purely run a quarter of such funds. Far
more important are those using ML. Like Bridgewater and many of BlackRock's funds, some
quantitative funds use algorithms to perform data analysis but call on humans to select trades.
This is because ML-driven algorithmic investing often identifies factors that humans have not.
But the human analysts still need to understand and interpret what ML has identified to be new
explanatory factors (Pozen & Ruane, 2019). Many quant funds, such as Two Sigma and
Renaissance Technologies, as well as a few of BlackRock's funds, are pushing automation even
further by using ML and DL to enable the machines to pick which stocks to buy and sell. There
is likely to be more wide-spread use of natural language processing, as indicates Mr Pierre
Rinaldi:
More widespread use of modern natural language processing technologies: NLP has been in use
for tens of years in hedge funds, but until now, most systems in production relied on manually
crafted rules. With teams getting involved by themselves in annotation campaigns and leveraging
Deep Learning techniques, large firms are now starting to use more advanced systems. (Mr
Pierre Rinaldi, Respondent 5).
Since the financial crisis, economists have been trying to understand systemic risk (Acemoglu,
Ozdaglar, & Tahbaz-Salehi, 2015; Anderson, Paddrik, & Wang, 2019; Chen, Iyengar, &
Moallemi, 2013; Löffler & Raupach, 2018). Network models, big data analysis and text mining,
have been used to understand risk (Gai & Kapadia, 2019; Gang, Xiangrui, Yi, Alsaadi, &
Herrera-Viedma, 2019; Yun, Jeong, & Park, 2019). They realize that there are complex network
effects between the different markets. Humans are too slow to gather data and compute the
different impacts using network-based models. Therefore, machines may be better at both mining
big data as well as analyzing it:
The volume of data that are accessible to investment managers is rising sharply from credit card
to geolocation data and many other types. Such large volumes of data need advanced AI
capabilities to utilize their information content fully. The added benefit of using AI would be a
refined approach to risk management of positions across many dimensions instead of boiling
down risk into a few numbers as human traders tend to. (Mr Vivek Mittal, Respondent 3).
Big data can be analyzed to form correlations between markets and sentiments. Sentiments are
nowadays being discovered by analyzing tweets, social network messages, news reports, and
laws (Gang et al., 2019). The quality and regularity of data remain an issue, and skilled staff can
devise AI systems and interpret the results. There is a lack of “availability of trained manpower
to run the algorithms and software built on top of the algorithms” (Mr Sudhir Kulakarni,
Respondent 6).
 The result of all these developments is that the stock market is now extremely efficient. The new
robo-markets result in much lower costs. Passive funds charge 0.03–0.09% of assets under
management each year. Active managers often charge 20 times as much (The Economist, 2019b,
October 5). The lower cost of executing a trade means that new information about a company is
instantly reflected in its price. Leading consumer brokerage sites expect that retail trading fees
will be zero since brokers like RobinHood take commissions from the High-Frequency traders to
route the trades and sell the consumer data (The Economist, 2016, March 10). The zero fees to
retail customers add to the liquidity of the shares. More liquidity means a lower spread between
the price a trader can buy a share and the price he can sell one. Table 1 provides five examples of
startups that are offering fringe services to users on an automated basis. The five largest startups
have been taken from a list of 14 (MEDICI, 2019). Four out of the five are in the 2019 Forbes
Fintech 50 startup list (https:// www.forbes.com/fintech/2019/#3f73ea772b4c). As costs equalize
to zero, competition will be based on better service provision to clients: customization. The zero-
cost transactions had to be emulated by giants like Charles Schwab and TD Ameritrade, who
have now decided to merge to enhance their client base (The Economist, 2019a, November 30).
With more clients, the routing commissions could be better negotiated.
Perhaps, one of the lessons from the above discussion is that human jobs may be progressively
be replaced by technology.
3.3 | Microfinance The most significant impact investments are in microfinance, where socially
responsible investment needs to be balanced with risk (Ashta, 2020). Indeed, Pierre Rinaldi
indicates ESG as a major opportunity area for using AI.
ESG: responsible investments have increased the need for AI techniques, as these specific
factors can be derived from new types of data sets (such as public news or social media). The
incredible trend around ESG investments has created new opportunities for AI in finance. (Mr
Pierre Rinaldi, Respondent 5).
Ten years ago, when we spoke of advanced technologies for the microfinance industry, we were
essentially talking about management information systems (MIS), online lending and the use of
mobile banking (Ashta, 2011). More recently, microfinance work focused on information
systems for operational excellence (Ashta, Barnett, Dayson, & Supka, 2015). Little work
considered AI deployment in dynamic information systems, except for a paper on a Distributed
Credit Bureau System (Krishnan & Pal, 2015), which predicted changes based on dynamically
changing algorithms:
The proposed system had used the architecture of decentralized and peer to peer control between
collection agencies of microfinance institutes. There would still be latencies in terms of central
agencies approving the funds. A possible opportunity of blockchains' emerging technology can
improve the borrowing process and reliability in peer-to-peer systems. Trust is a crucial factor in
such scenarios, and blockchain technologies would be a perfect solution to improve efficiencies.
Another opportunity is to develop tools for better credit scoring and faster decision making. High
volume data analysis would help improve credit score check algorithms and help individual
context be considered in such instances. Intelligent automation is another opportunity being
leveraged. Back office operations, Intelligent and automated customer support like chatbots, IVA
platforms support these initiatives. (Dr Sudeep Krishnan, Respondent 4).
More recently, online crowdlending's development addressed a major problem in lending to the
poor: the high transaction costs relative to the small loan size. Over the past decade, mobile
banking has shown that it can reduce the transaction costs of going to the borrower or the
borrower traveling to the bank and the risk of being robbed on the way. Therefore, mobile
payments have gone up. The second major problem of lending to the poor, information
asymmetry owing to the unavailability of information on the borrower, was addressed through
credit scoring models that bankers have been using for several decades (Bumacov, Ashta, &
Singh, 2017). Recently, microfinance institutions have used credit scoring to increase their loan
officers' productivity and expand financial inclusion (Bumacov, Ashta, & Singh, 2014).
However, giving loans still requires an assessment of risks for new borrowers. A spate of new
credit scoring platforms is now using online data points collected from social networks to
determine if the potential borrower is safe or risky.
In Europe, they use data from Facebook and LinkedIn. Based on this data, platforms such as
Kreditech or ID Finance can rate borrowers as safe or risky, and based on this formation, they
can offer microloans (Ashta, 2018). An interesting recent example of using payment behavior is
the Swedish operator iZettle. Rather than let the public buy iZettle for $1.1 billion, Paypal
acquired it for $2.2 billion. The reason is that iZettle had the technology to analyze the payment
behavior on its card reader to enable credit rating for giving cash advances to small businesses.
The use of social data for credit scoring in the developing world was probably pioneered by
Lenddo, a company that operated in the Philippines and Columbia and has since expanded
overseas. In Africa, Credit Suisse is trying to capture similar data from mobile telephones, using
biometrics and facial recognition for identity checks as well as behavioral characteristics
captured on their telephone including calling histories and financial transactions.
One Fintech operator in the mobile banking space is MyBucks, a Luxembourg based company
listed on the Frankfurt Stock Exchange and operating primarily in African countries. Mobile
banking reduces cash transactions, and the digital transaction itself becomes data for future
analysis. This data analysis helps the microfinance institution understand its customers better, by
region or by type of client, and lower risk. Moreover, microfinance institutions are also
experimenting with facial recognition allowing the lender to see if the same face has multiple
loans or accounts, something humans cannot do. MyBucks uses AI for credit scoring, fraud
prevention, chatbot, collection algorithm, loan uptake prediction, and client churn prevention. AI
is also allowing lending to customers directly on Facebook or WhatsApp.
“An exciting development has been the Text-based Virtual Assistant that we call TESS. TESS
has moved us into the leading space of WhatsApp and Facebook lending.” Dave van Niekerk,
(MyBucks, 2019)
ML/DL technologies enable easier and accelerated access to loans as well as insurance.
Technology accelerates the lending processes by making better decisions about potential
customers, automatically calculating affordability and likelihood of default, even for customers
with no formal credit history and no bank account. The customer installs an App on his/her
mobile phone, and the lender can use the information on the customer's phone and browser to
estimate risk and disburse funds within minutes if approved. There is no human intervention or
paperwork required, and therefore, operational costs are extremely low. Only applications that
are not automatically approved require human involvement. To encourage people to record their
mobile expenditure, in a pilot program at a microfinance firm, Swadhar, Fidelity Information
Services (FIS) partnered with Citibank. They provided free telephones to women along with a
money managing app named Saathi. The women were taught to record all incomes and expenses.
One year into its pilot, Saathi has successfully helped its users better understand their financial
behaviour, make them conscious and accountable for their expenses, and help users improve
their knowledge of basic financial tools. (FIS, 2019)
Simultaneously, while helping the women, the app allows collecting information that can be
analyzed to provide a host of services. As information becomes increasingly valuable, the
smartphones' donations and subsidized internet access may be treated as loss-leaders that permit
other income to be earned.
Of course, many markets are still dependent on cash, which is important for the poor who
become micro-entrepreneurs and enter the financial marketplace. ATMs are still selling well in
such markets but are themselves using technology and sensors to detect problems and repair the
machines (Castor, 2019). Voice-enabled machines have been around for many years, catering to
the visually impaired.
The next challenge may be looking at about a billion illiterate people, many of whom do not
even have numeric literacy and rely only on oral intelligence. Illiteracy leads to high transaction
costs as distrust among this population rises of anything they do not understand. Financial
numeracy requires the ability to recognize numbers, input these numbers, place value on large
numbers, do simple calculations and have an idea of time. People who do not have numeric
literacy do not have the capacity or the incentives to use digital wallets, and they need “Oral
intelligence Management” (Matthews, 2019). One solution for this would be developing and
combining data analysis on user experience with assistive technology through smartphones.
While many of these poor people use noncash items to store value, research shows that they can
use money (notes and coins) as visual aids to numeracy. Therefore, smartphones could use
photographs of the local currency to designate prices along with voice communication.
“For years, no one in microfinance would believe that this ‘oral’ population even needed a
different interface from the sort you find on PayTM or M-Pesa or JazzCash. However, I have
made substantial progress in recent years in shaking the complacent assumption that a person
with a phone in their pocket can read a financial statement. How can we achieve financial
inclusion when we use Indo-Arabic notation, not images of cash, as a (highly abstract) code for
money when we communicate with poor people?” (Mr Brett Matthews, Respondent 1)
4. UNCERTAINTIES AND RISKS
 So far, we have looked at optimistic case studies from the financial sector that show efficiency
will increase, leading to more efficient financial markets. However, some pessimists think that
AI creates problems, pragmatists who feel that we can usher in controls to take care of problems
and doubters who think that general AI will not happen (Makridakis, 2017). However, most
experts in the field are not doubters and consider a 50% probability that super-intelligence or
even singularity is only three decades away (Makridakis, 2017; Turchin, 2019).
As we know, the market works to increase inequalities by making first-movers and the most
productive firms bigger and thus creating monopolies. In systems thinking terms, this effect has
become known as “success to the successful” and forms the basis of many antitrust laws
(Meadows, 2008). The first fear is that the big tech players (such as GAFA in the USA or BAT
in China) will dominate the market and push out incumbents like banks and small fintech firms
(Ashta & Biot-Paquerot, 2018). Big tech firms have a culture of experimenting and offering
freemium products to attract attention and data. It is easy for them to offer financial services like
Alipay or ApplePay. This concentration of economic power can also be observed in the big deals
taking place over the last decade, such as the aforementioned example of PayPal's acquisition of
iZettle. An increase in monopoly power means customers will finally have less choice and pay
more. On the other hand, if the speed of technological change keeps on increasing, today's
market powers may not exist tomorrow.
A more technology-oriented fear is that the algorithms used may not be suitable for financial
markets, a problem Mr Vivek Mittal (our third respondent) describes as “overfitting patterns”
: AI methods have had difficulty gaining traction on the investment side due to the high chances
of overfitting patterns. Unlike image recognition problems, financial markets have far greater
noise to signal ratios and are not stationary systems that can cause complex algorithms to latch
onto training set patterns. Usual methods such as splitting into training and test sets are
inadequate for complex time series patterns as present in financial markets. Highly volatile
correlation relationships between markets further compound these problems. (Mr Vivek Mittal,
Respondent 3)
This risk of data fitting is echoed by Sudhir Kulkarni, who considers a lack of “availability of
structured training data sets to get the algorithms to work reliably in a dynamically changing
ecosystem”. (Mr Sudhir Kulkarni, Respondent 6).This concern may be mitigated by a
regularization method for DL from limited data sets, referred to as dropout (Hinton, Srivastava,
Krizhevsky, Sutskever, & Salakhutdinov, 2012).
Pierre Rinaldi talks about a trend toward standardization and industrialization of the ML process
to reduce errors.
Several hedge funds have used machine learning on time series in the past few years to generate
trading signals with classical machine learning models or deep learning (multi-layer neural
networks). While these approaches have helped them achieve great results, the research process
is usually strongly isolated from software engineering activities. Therefore, funds are often
struggling with nonstandard code that is hard to maintain and hard to use in production. Hedge
funds and asset managers that are strongly involved in this domain are quickly finding internal
tools to optimize this process and allow their teams to concentrate on modeling. Sometimes, the
greatest advances in AI do not rely on the most complex model but rather on the easiest use the
system. (Pierre Rinaldi, Respondent 5).
Another technical fear of market operators is to ask who takes the accountability for an error in
nonrepresentative data? Data needs to be correct for it to be meaningful. The industry term for
this is “garbage in [bad data], garbage out [wrong conclusions].” Care needs to be taken to
ensure the data captured is accurate and data integrity is maintained— especially as the data
increases in volume and variety. Data scientists need to select representative data sets, massage
the data and integrate existing knowledge to prevent bias from creeping in (Pozen & Ruane,
2019). For example, US banks rejected a disproportionate number of loan applications from
black and Hispanic Americans, which is an example of bias (Yale, 2018, May 7). ML learns
from past data—this means it perpetuates what we want to change. Can a manager possibly
monitor or control the overload of information (for a human) when it is delegated to a machine
(Frind, 2019)? It is easy to delegate decision-making to a machine, but when things go wrong,
opening up the machine's “black box” and making sense of it is rather complicated (Pierre
Rinaldi, our fifth respondent). Mr Sudhir Kulkarni (our sixth respondent) terms this as a lack of
transparency of the algorithm. Indeed, in credit scoring, it is often believed that credit agents
should not know of the factors included in the algorithm so that they do not manipulate the data
of the potential borrowers. This lack of knowledge, termed the black box, makes users
uncomfortable. Fox et al. (2019, September 17) explore the impact of AI onboard directors'
responsibility and conclude that directors cannot use AI as a substitute for making their
independent assessments. However, a number of jurisdictions are considering the ethical and
liability implications of the use of AI in the boardroom.
A labour market perspective focuses on the future of work, predicting winners and losers. Some
fears are related to Schumpeter's (1942) creative destruction of jobs almost a century ago. A
recent book predicted blue- and white-collar workers could be displaced by automation
(Brynjolfsson & McAfee, 2014). A second study estimated that the top end of the workforce
(managerial, professional, and technical-specialist jobs) and service jobs at the bottom end had
gained advantages. At the same time, the middle classes lost out (Autor, 2015). It indicated that
although automation replaces workers, it also increases their marginal productivity and therefore,
they tend to gain from automation. For example, ATMs' advent enabled bank tellers to refocus
on relationship banking, which is harder to achieve for Neo-banks (Castor, 2019).
Contrary to the common belief that unemployment will be a result of advances in AI and
technology, these emerging technologies create the need for better-skilled humans to support the
technology. However, obsolete and repetitive jobs are going to go away. (Dr Sudeep Krishnan,
Respondent 4)
As AI moves into a more creative learning phase (i.e., it becomes more human), a win-win type
of labour division between AI and the workforce may not be sustainable. Recent statistics from a
McKinsey report (Manyika et al., 2017) indicate that about half of workers' activities could be
automated. It is estimated that 15% of the global workforce could lose jobs and that new jobs
will need different skills. Technical staff using AI who remain at the forefront of learning will
always be required by incumbents and challengers. However, automation will put pressure on
wages of other staff roles in advanced countries. It has been suggested that replacing managers
by machines will increasingly take over thinking tasks, and only feeling tasks, such as
communication, will remain the domain of humans (Huang, Rust, & Maksimovic, 2019). Many
banks are reducing their workforce in developed countries, but it is difficult to ascertain whether
fintech startups have compensated for this job destruction. Bankers are in a dilemma: if they do
not automate, they will fall behind their competition (Dicamillo, 2019). There may be a need to
move toward a shorter work-week or use other work-sharing forms (Ashta, 2021).
While the press is focusing on the gig economy and the plight of unsecured low paid workers,
the financial system as a whole may also explode as institutions become leaner and more
specialized, creating a need to work together in dynamic partnerships, with a strategic risk of
falling out (Ashta, 2009; WEF, 2019). Some partners' scale and agility create insecurity for other
partners as each is looking out for better partnerships. This point was also mentioned by our third
respondent, Mr Vivek Mittal.
Science fiction writers (Fesnak, 2019) have expressed sociopsychological concerns. One fear is
the threat of totalitarian governments in an Orwellian-type of the state that can monitor the
actions, thoughts, and emotions of its citizens. Another could be that corporations can use all the
data to force you to consume what they decide. A third could be that humans become slaves to
robots or that they become weaker than robots. The current thinking is that decision-making
could be a delegated task, an integrated task or a hybrid task with different relations between
man and machine (Shrestha, Ben-Menahem, & von Krogh, 2019).
There are also privacy concerns about the risks of pirating concentrated bid data by a few
organizations. If the data is shared among them to increase their collaborative strength, cyber-
attacks' risk increases with such networks' weakest link (Vartanian, 2019).
Loss of privacy—when every human action is being used for decision making, the obvious
concern is related to privacy of data and how it is being used for a specific context. ‘Dr Sudeep
Krishnan, Respondent 4).
5. CONCLUSION
Our overview shows that AI innovations analyze big data to help cut costs, reduce risk, and
increase customization, leading to economic growth through an increase in aggregate demand
and investments. The top three opportunities for the future may well be “Risk Management,
Customer Targeting and Customer Engagement” (Mr Sudhir Kulkarni, Respondent 6).
Our review has shown that cost-cutting creates major moves in the financial marketplace, forcing
firms to merge to build bigger customer platforms. This customer acquisition strategy has been
found in financial service providers and financial markets. However, the more prominent
platforms may only have a short-term advantage as customers look to specialized expertise from
focused fintechs.
 End-users of AI do not accept one solution for a broad range of situations. They want solutions
to consider deeper context, domain expertise and differentiation. (Dr Sudeep Krishnan,
Respondent 2).
Microfinance markets in developing countries are too dispersed to show meaningful results for
the moment. Nevertheless, considerable experimentation in credit scoring techniques using social
data could be helpful for mobile lending. There is still a long way to go.
Many traditional financial organizations are worried about AI grappling with where it is not
working well. The field of AI needs to understand what still needs to be researched and indeed
productized to put financial institutions at ease:
1. Maturity of autoML tools—This involves data preparation, model discovery, model building
and training, serving, and postproduction management and monitoring. It makes it easier for
companies to build their solutions on top of prebuilt, proprietary domain data sets, such as image
recognition.
2. Training data—Unbiased training data remains a major ML/DL pain point. The “Black Lives
Matter” movement is a recent prime example.
3. Explicability—Opening the “black box” to explain and audit why a particular prediction led to
an operational change in an organization. New techniques and products are emerging to address
this issue.
4. Model management, performance, and monitoring—As organizations increasingly deploy AI,
its quality, performance, and understanding of business KPIs (Key Performance Indicators)
become quite significant.
5. Privacy—Federated learning, for example, is an AI architecture for preserving end-user
privacy and security. It allows AI to learn from user data in a privacy-sensitive way.
The power of the ML/DL subdisciplines of AI fundamentally deals with managerial blind spots:
unknown knowns and unknown unknowns. Figure 3 depicts this key conclusion by mapping AI
versus other technologies.
We attempt a concluding response to the question of “humans in the loop.” Intuition and data-
based logic interplay in human brains. Kahneman won the Nobel Prize for this conclusion.
However, today we have algorithms only equivalent to reptile-brain animals, according to AI
guru Yoshua Benigo. The good news is that DL learns by itself what weight should be given to
the data. There are algorithms for a broader range of applications such as approving loans,
recognizing pictures/photos or audio clips, translating languages, preventative maintenance, and
self-driving cars. However, these are context-specific “toy problems” or “narrow AI,” which is
not the same as abstracting and transferring learning into a different context. In other words,
learning how to learn must be the next level of AI to be accepted in the Finance sector and
beyond.
Despite that, AI remained evolving around toy problems in the previous decade. For example, in
the early 2000s, trees were the hot topic of research, focusing on extending tree capabilities. In
the 2010s, there has been a focus on DL with multi-layer neural networks. The next decade
might herald a different focus, but what it might be is hard to say. Perhaps, it may be geometric
or topological methods because the number of papers in this area is growing substantially. What
is going to be the next big thing might be just another better technique for today's toy problems.
Moreover, that ethos would not accelerate AI's deduction and reasoning capabilities, based on
what was learned before in a different context.
We have also seen that we need to keep bias away from the AI factory by expanding the
diversity of people working on AI to include nontechnical professional backgrounds. Companies
developing AI should consider hiring creatives, writers, linguists, and sociologists. In particular,
unsupervised AI is unique in that it keeps learning and therefore changing after it leaves the AI
factory. The problem is that there is a lack of testing AI products throughout their development
cycle to detect potential harms they may do to human society, ethically, or emotionally, once
they hit the market. Bias testing would help account for the blind spots created by the lack of
diversity present in engineering terms of who build AI. It is encouraging that scholarly
publications in the field of AI ethics have exploded in the late 2010s.
Educational institutions need to prepare students for seizing opportunities and moving between
them in a fast-changing world, by teaching students agility, resilience, adaptability, and career
preparedness, leading to a continuous learning and innovation culture.
Financial institutions benefit greatly from retaining employees. They will need to provide
budgets for ensuring that their employees remain competitive. According to Mr Vivek Mittal
(Respondent 3), the most efficient banks will transform their jobs where necessary.
Indeed, according to a respondent, “the greatest fear is that AI may be used alone, without
human intervention.” (Dr Eunika Laurent-Mercier, Respondent 2).
ACKNOWLEDGMENT
Our thanks to the interviewees