The use of Artificial Intelligence and Machine Learning in the banking industry is increasing. What do these techniques entail? What are their main applications and what are the risks concerned? Is the three lines of defence model still fit for purpose when using these techniques? These are the topics that will be addressed in this article.

Technology and data are playing an increasingly important role in the banking industry. While Artificial Intelligence (AI) was initially mostly used in client servicing domains of the bank, more and more applications for risk management purposes can be observed.

A common model to use within banks is the three lines of defence (3LoD) model. This model consists of a first line in the business, being responsible for managing risks, a second line risk management function in an oversight role and a third line function: internal audit. Given the expanding use of AI and machine learning (ML) within banks, the question arises whether this 3LoD model is still fit for purpose given these new developments, or if changes to the model are necessary.

This article aims to answer the question: “How can the application of Artificial Intelligence and Machine learning techniques within banks be placed in the context of the Three lines of defence model?”

This article will first address the basic concepts of AI and ML and the 3LoD model. It will then give an overview of the applications observed throughout banks and the risks and challenges of using AI and ML. After that, AI and ML are placed in the context of the 3LoD model, addressing the prerequisites to apply AI and ML in a controlled way. The article finishes with a regulatory view, the emergence of potential new market wide risks, conclusions and recommendations.

As a start, it is important to clarify the concepts of Artificial Intelligence (AI) and Machine Learning (ML), which are often interchanged. Several definitions can be found. AI is mostly viewed as intelligence demonstrated by machines, with intelligence being defined with reference to what we view intelligence as in humans (Turing 1952 cf Shieber 2004 in Aziz and Dowling 2019). Or another definition: AI refers to machines that are capable of performing tasks that, if performed by a human, would be said to require intelligence (Scherer 2016).

AI uses instances of Machine Learning as components of the larger system. These ML instances need to be organized within a structure defined by domain knowledge, and they need to be fed data that helps them complete their allotted prediction tasks (Taddy 2018). As such, ML delivers the capability to detect meaningful patterns in data, and has become a common tool for almost any task faced with the requirement of extracting meaningful information from data sets (Leo et al. 2019). ML may also be defined as a method of designing a sequence of actions to solve a problem, known as algorithms which optimise automatically through experience and with limited or no human intervention (FSB 2017). ML is limited to predicting a future that looks like the past, they are a tool for pattern recognition (Taddy 2018). According to Mullainathan and Spiess (2017), the appeal of ML is that it manages to uncover generalizable patterns. In fact, the success of ML at intelligence tasks is largely due to its ability to discover complex structure that was not specified in advance. It manages to fit complex and very flexible functional forms to the data without simply overfitting; it finds functions that work well out-of-sample (Mullainathan and Spiess 2017). So ML is a core technique of AI, learning from data, but AI often involves additional techniques and requirements (Aziz and Dowling 2019). So as Taddy (2018) states, AI is a broader concept, meaning that an AI system is able to solve complex problems that have been previously reserved for humans. It does this by breaking these problems into a bunch of simple prediction tasks, each of which can be attacked by a ‘dumb’ ML algorithm.

As Reddy (2018) states, ML comprises a broad range of analytical tools, which can be categorized into ‘supervised’ and ‘unsupervised’ learning tools. Supervised learning is an approach to ML where the historical input data is tagged with its corresponding business outcomes and the ML solution is expected to identify and learn the patterns in the input data associated with a business outcome and self-develop an algorithm based on this learning to predict a business outcome for a future instance. So supervised ML involves building a statistical model for predicting or estimating an output based on one or more inputs (e.g., predicting GDP growth based on several variables). The supervised learning approach usually operates with a classification aim (e.g. will a loan default yes or no) or based on regression, in which a quantified value is predicted (e.g. what is the probability of loan default) (Reddy 2018).

In unsupervised learning, a dataset is analysed without a dependent variable to estimate or predict. Rather, the data is analysed to show patterns and structures in a dataset (Van Liebergen 2017). So the historical input data is fed into the ML solution without any tagging of the business outcomes and the solution is expected to decipher or self-develop an algorithm for prediction based on its own interpretations of the patterns in the data without any guidance or indicators. The unsupervised learning approach usually performs via Clustering (e.g. of customers in segments for credit risk) or Association (e.g. impact of increased draw-down on credit lines prior to default) (Reddy 2018).

So the main difference between supervised and unsupervised ML is the tagging of historical data with business outcomes in supervised learning, where this is not done in unsupervised learning. ‘Reinforcement learning’ falls in between supervised and unsupervised learning. In this case, the algorithm is fed an unlabelled set of data, chooses an action for each data point, and receives feedback (perhaps from a human) that helps the algorithm learn. For instance, reinforcement learning can be used in robotics, game theory, and self-driving cars (FSB 2017).

In discussions about AI, the concept of deep learning or neural networks is also mentioned often. In deep learning, multiple layers of algorithms are stacked to mimic neurons in the layered learning process of the human brain. Each of the algorithms is equipped to lift a certain feature from the data. This so-called representation or abstraction is then fed to the following algorithm, which again lifts out another aspect of the data. The stacking of representation-learning algorithms allows deep-learning approaches to be fed with all kinds of data, including low-quality, unstructured data; the ability of the algorithms to create relevant abstractions of the data allows the system as a whole to perform a relevant analysis. Crucially, these layers of features are not designed by human engineers, but learned from the data using a general-purpose learning procedure. They are also called ‘hidden layers’ (Van Liebergen 2017). Deep learning can be both supervised and unsupervised forms of learning, depending on the purpose for which it is applied. Deep learning techniques are complex, they are often perceived as a black box. It is not always clear how inputs have been recombined to create a predicted output (Aziz and Dowling 2019). This has obvious implications for use in risk management, the presence of a black box in decision making has its own challenges and can be a risk in itself.

Also, the concepts of predictive versus prescriptiveAI are relevant. Predictive AI is about understanding and predicting the future, so about using statistical models and forecast techniques to understand the future to predict what could happen. Prescriptive AI uses optimization and simulation algorithms to advice on possible outcomes and to instigate what action to take.

Other concepts within AI are speech recognition and Natural Language Processing (NLP). This is the ability to understand and generate human speech the way humans do by, for instance extracting meaning from text or generating text that is readable, stylistically natural and grammatically correct (Deloitte 2018).

One could wonder in which way AI and ML are different from more traditional statistical modelling techniques. Statistical modelling gives insight in correlation, derives patterns in the data using mathematics. It is a formalization of relationships between variables in the form of mathematical equations.The main difference compared to AI/ML is that the ML model trains itself using algorithms, it can learn from data without relying on rule based programming (Srivastava 2015). ML requires almost no human intervention because it is about enabling a computer to learn on its own from a large set of data without any set instructions from a programmer. It explores the various observations and creates definite algorithms that are self-sufficient enough to learn from data as well as make predictions (Mittal 2018).

In the 3LoD Defence model (IIA 2013):

Regardless of how the 3LoD model is implemented, senior management and governing bodies should clearly communicate the expectation that information should be shared and activities coordinated among each of the groups responsible for managing the organization’s risks and controls (IIA 2013).

The 3LoD model has also received criticism. The core concern according to Davies and Zhivitskaya (2018) is that the existence of three separate groups who are supposed to ensure proper conduct towards risks has led to a false sense of security. If several people are in charge, no one really is. Different criticism addresses that the 3LoD model could downplay the importance of strong risk management in the business areas themselves: “not enough emphasis is placed on the first line of defence which is management” or that it could lead to an excessively bureaucratic, costly, and demotivating approach to risk management. The Financial Stability Institute (2017) also mentions weaknesses in the 3LoD model. The responsibility for risk in the first line conflicts with their primary task which is generating sufficient revenues and profit, which requires risk-taking. So there are misaligned incentives here. In other cases, second line functions may not be sufficiently independent, or lack sufficient skills and expertise to effectively challenge practices and controls in the first line (Arndorfer and Minto 2015). Lim et al. (2017) state that whilst the 3LoD model has formally spread the responsibility for risk management across different organisational lines, a real impact on the hierarchy within the organisation is not observed enough yet: often, traders are perceived as more valuable to the organisation than risk and compliance personnel (Lim et al. 2017). Supporters of 3LoD argue that, while these criticisms may have been valid in the past, the system has been made stronger since the Global Financial Crisis (Davies and Zhivitskaya 2018). When placing the use of AI and ML into the context of the 3LoD model, the criticism should be kept in mind.

To get a better insight in the risks associated with using AI and ML, this section addresses some use cases of AI and ML within banks throughout all of the 3LoD functions. These are depicted in figure 2 as well.

As depicted in figure 3, there are quite a few risks that need to be addressed when using AI and ML techniques.

As shows from the use cases mentioned above, AI and ML can be used within each of the 3LoD, or throughout multiple lines. It appears that the techniques are most used within the first line, or in use cases where first and second line are both involved.

If used purely in the first line, the 3LoD model can be applied as designed. In this case, it is important to safeguard that sufficient knowledge of the techniques and its use is also present in second and third line functions, to ensure compliance, to identify and manage risks, to challenge the first line on replicability of decisions and validity of the model and to perform audits effectively. As mentioned above, the scarcity of resources with the required skills and knowledge can be an issue (FSB 2017).

For a number of applications, such as credit risk modelling and approval, transaction monitoring or fraud detection, both the first and the second line are involved. Here it gets more difficult to apply the 3LoD model. Depending on the nature of the involvement of the second line function, e.g. whether they are developing AIML tools themselves, there should be an independent function involved that provides independent validation and challenge. So applying the 3LoD model without any adjustments does not seem wise in this case. When zooming in on the second line risk management function, this function “facilitates and monitors the implementation of effective risk management practices by operational management and assists risk owners in defining the target risk exposure and reporting adequate risk-related information throughout the organization” (IIA 2013). So if the risk management function is operationally involved in e.g. developing a model using AI and ML techniques, or the AI and ML model is developed for purely second line purposes such as in model risk management or stress testing, an alternative solution is warranted. In this case, as a minimum, independent oversight, challenge, validation and assurance should be safeguarded by a separate function performing this second line role. In addition, the internal audit function must also be involved. No use cases have been found in purely third line functions but if AI and ML techniques were to be used, external assurance surrounding the use of AI and ML is warranted.

A potential better way of ensuring a controlled deployment of AI and ML techniques, which is at the same time in line with the principles of the 3LoD model is to assign specific roles (Burt et al. 2018):

This could be performed by an independent function, or if the size of the bank is insufficient, by data scientists who are not associated with the specific model or project at hand.

Together with the business owners, a group of data owners and data scientists comprise the first line of defence. The validators comprise the second line of defence, together with the governance personnel. The third line function could be performed by independent internal auditors, provided that they have the expertise needed. This set up is necessary to safeguard an effective challenge throughout the model lifecycle by multiple parties, separate from the model developers. In assigning these specific roles, the principles of the 3LoD model are safeguarded.

Some other points are relevant when thinking about AI and ML in the context of the 3LoD model and controlled application. All ML projects should start by clearly documenting initial objectives and underlying assumptions, which should also include major desired and undesired outcomes. This should be circulated and challenged by all stakeholders. Data scientists, for example, might be best positioned to describe key desired outcomes, while legal personnel might describe specific undesired outcomes that could give rise to legal liability. “Such outcomes, including clear boundaries for appropriate use cases, should be made obvious from the outset of any ML project. Additionally, expected consumers of the model — from individuals to systems that employ its recommendations – should be clearly specified as well” (Burt et al. 2018).

The materiality of the model that is deployed should be taken into account in all three lines (Burt et al. 2018). This means that the intensity and frequency of involvement of second and third line functions, or validators and governance personnel, should be based on the impact that the model has within the banks or towards its clients.

How ‘black box’ the AI technique is, is often a result of choices made by developers of the model. Predictive accuracy and explainability are frequently subject to a trade-off; higher levels of accuracy may be achieved, but at the cost of decreased levels of explainability. This trade off should be documented from the start, and challenged by other functions. “Any decrease in explainability should always be the result of a conscious decision, rather than the result of a reflexive desire to maximize accuracy. All such decisions, including the design, theory, and logic underlying the models, should be documented as well” (Burt et al. 2018). Note that using Deep learning techniques requires even more specific knowledge throughout the 3LoD.

When viewing the significant amount of risks in using AI and ML as described above, and the challenges when it comes to applying the 3LoD model, a sound governance surrounding the use of AI and ML is essential. The risks concerned need to be properly identified, assessed, controlled and monitored. This also means clearly defining the roles and responsibilities for the functions involved, be it in the first, second or third line of defence. “Any uncertainty in the governance structure in the use of AI and ML might increase the risks to financial institutions” (FSB 2017). Given the challenge to view all risk integrally, a dedicated oversight function of all AI and ML use throughout the bank is required, especially for larger banks. A sound framework is necessary to create, deploy and maintain AI and ML techniques in a controlled way and to manage the risks involved properly. It is also important to develop policies and processes for the use of AI and ML, ensuring that the deployment of these techniques fit the strategy and risk appetite of the bank. “Any uncertainty in the governance structure could substantially increase the costs for allocating losses, including the possible costs of litigation” (FSB 2017). As part of sound governance, a sound model risk management framework is also necessary and it should be updated or adjusted for AI/ML models. Given all of the risks mentioned above and the self-learning nature of AI/ML models, extra attention is warranted. As Asermely (2019) describes it: “The dynamic nature of machine learning models means they require more frequent performance monitoring, constant data review and benchmarking, better contextual model inventory understanding, and well thought out and actionable contingency plans”. Given increasing volumes and complexity of data, increasing use of AI/ML and the growing complexity of AI/ML, sound governance will also be increasingly important towards the future (Asermely 2019).

According to the Financial Stability Board (2017), because AI and ML applications are relatively new, there are no known dedicated international standards in this area yet. Apart from papers on this topic published by regulatory authorities in Germany, France, Luxembourg, The Netherlands and Singapore, no European or international standards were published. Although calls to regulate AI and ML are heard more often, the current regulatory framework is not designed with the use of such tools in mind. Some regulatory practices may need to be revised for the benefits of AI and ML techniques to be fully harnessed. “In this regard, combining AI and ML with human judgment and other available analytical tools and methods may be more effective, particularly to facilitate causal analysis” (FSB 2017). DNB (Van der Burgt 2019) states “Given the inherent interconnectivity of the financial system, the rise of AI has a strong international dimension. An adequate policy response will require close international cooperation and clear minimum standards and guidelines for the sector to adhere to. Regulatory arbitrage in the area of AI could have dangerous consequences and should be prevented where possible.”

DNB recently published a set of general principles for the use of AI in the financial sector (Van der Burgt 2019). The principles are divided over six key aspects of responsible use of AI, namely soundness, accountability, fairness, ethics, skills and transparency.

“The Basel Committee on Banking Supervision (BCBS) notes that a sound development process should be consistent with the firm’s internal policies and procedures and deliver a product that not only meets the goals of the users, but is also consistent with the risk appetite and behavioural expectations of the firm. In order to support new model choices, firms should be able to demonstrate developmental evidence of theoretical construction; behavioural characteristics and key assumptions; types and use of input data; numerical analysis routines and specified mathematical calculations; and code writing language and protocols (to replicate the model). Finally, it notes that firms should establish checks and balances at each stage of the development process” (FSB 2017).

Many of the use cases described in this article could result in improvements in risk management, compliance, and systemic risk monitoring, while potentially reducing regulatory burdens. AI and ML can continue to be a useful tool for financial institutions by implementing so called “RegTech”, aiming to facilitate regulatory compliance more efficiently and effectively than existing capabilities. The same goes for supervisors via “SupTech”, which is the use of AI and ML by public sector regulators and supervisors. The objective of “SupTech” is to enhance efficiency and effectiveness of supervision and surveillance (FSB 2017).

From a market wide perspective, there are also potential new and/or systemic risks to take into account when using AI and ML techniques. If a similar type of AI and ML is used without appropriately ‘training’ it or introducing feedback, reliance on such systems may introduce new risks. For example, if AI and ML models are used in stress testing without sufficiently long and diverse time series or sufficient feedback from actual stress events, there is a risk that users may not spot institution-specific and systemic risks in time. These risks may be pronounced especially if AI and ML are used without a full understanding of the underlying methods and limitations. “Tools that mitigate tail risks could be especially beneficial for the overall system” (FSB 2017).

A more hypothetical issue is that models used by different banks might converge on similar optimums for trading causing systemic risk as well (Aziz and Dowling 2019). “Greater interconnectedness in the financial system may help to share risks and act as a shock absorber up to a point. Yet if a critical segment of financial institutions rely on the same data sources and algorithmic strategies, then under certain market conditions a shock to those data sources could affect that segment as if it were a single node and thus could spread the impact of extreme shocks. The same goes for several financial institutions adopting a new strategy exploiting a widely-adopted algorithmic strategy. As a result, collective adoption of AI and ML tools may introduce new risks” (FSB 2017).

“AI and ML may affect the type and degree of concentration in financial markets in certain circumstances. For instance, the emergence of a relatively small number of advanced third-party providers in AI and ML could increase concentration of some functions in the financial system” (FSB 2017). DNB states that “Given the increasing importance of tech giants in providing AI-related services and infrastructure, the concept of systemic importance may also need to be extended to include these companies at some point” (Van der Burgt 2019). The role of Big-Tech companies requires attention here. “Many BigTech firms also offer specific tools using artificial intelligence and machine learning to corporate clients, including financial institutions. The activity of BigTech firms as both suppliers to, and competitors with financial institutions raises a number of potential conflicts of interest, at the same time that their dominant market power in some markets is coming under greater scrutiny” (Frost et al. 2019).

“The lack of interpretability or ‘auditability’ of AI and ML methods has the potential to contribute to macro-level risk if not appropriately audited. Many of the models that result from the use of AI or ML techniques are difficult or impossible to interpret”. Auditing of models may require skills and expertise that may not be present sufficiently at the moment. “The lack of interpretability may be overlooked in various situations, including, for example, if the model’s performance exceeds that of more interpretable models. Yet the lack of interpretability will make it even more difficult to determine potential effects beyond the firms’ balance sheet, for example during a systemic shock. Notably, many AI and ML developed models are being ‘trained’ in a period of low volatility. As such, the models may not suggest optimal actions in a significant economic downturn or in a financial crisis, or the models may not suggest appropriate management of long-term risks” (FSB 2017).

Artificial Intelligence (AI) refers to machines that are capable of performing tasks that, if performed by a human, would be said to require intelligence. AI uses instances of Machine Learning (ML) as components of a larger system. ML is able to detect meaningful patterns in data. The main difference when comparing AIML techniques with more traditional statistical modelling techniques is that the AI/ML model trains itself using algorithms, so it can learn from data without relying on rule based programming or instructions from a human programmer.

Among the most used AI and ML techniques within banks are credit risk modelling- and approval, transaction monitoring regarding Know Your Customer and Anti Money Laundering and fraud detection, which are usually jointly developed by first and second line functions. Frequently observed use cases in the first line are client servicing solutions and market risk monitoring- and portfolio management. The techniques are used to a lesser extent for pure second line risk management purposes until now, while no use cases have been observed for third line functions. It is expected that applications in the risk management and internal audit domain will increase in the years to come.

There are obvious benefits to using AI and ML techniques, they may enhance machine-based processing of various operations in financial institutions, thus increasing revenues and reducing costs. It is expected that the time for data analysis and risk management will decrease, e.g. by earlier and more accurate estimation of risk, making risk management more efficient and less costly. The ability of ML techniques to analyse volumes of data without being constrained by assumptions of distribution is significant. Also, meeting regulatory requirements could become more efficient by automating repetitive reporting tasks and by the increased ability to organize, retrieve and cluster non-conventional data such as documents.

There are also numerous risks and challenges to address. Modelling issues and data issues can occur when insufficient suitable data is available, or when hackers maliciously manipulate big data. Also, the model outcomes have not been tested through a financial cycle yet. There are risks regarding consumer protection and privacy as well as reputational risks stemming from ethical issues. Sufficient specialized and skilled staff is needed within banks and there are numerous risks regarding transparency and auditability.

This article aimed to answer the question: “How can the application of Artificial Intelligence and Machine learning techniques within banks be placed in the context of the Three lines of defence model?”

When AI and ML are placed in the context of the 3LoD model, there are quite some prerequisites to apply AI and ML in a controlled way. If the second line risk management function is involved in the operational development of the model, independent oversight, challenge, validation and assurance should be safeguarded by a separate function performing the second line role. In addition, the internal audit function must be involved. Ensuring the proper functioning of the 3LoD model could also be done by assigning specific roles within each AI/ML project, that safeguard the controlled deployment of AI and ML techniques. Data owners and data scientists comprise the first line of defence, together with the business owner. The second line role could then be comprised of validators and other governance personnel that review and approve the work from a technical and a compliance perspective, respectively. Other prerequisites are a sound governance surrounding the use of AI and ML, clearly defined roles and responsibilities, a dedicated oversight function, a sound model risk management framework, a sound framework for managing all of the risks and policies and processes for the use of AI and ML, ensuring that the deployment of these techniques fit the strategy and risk appetite of the bank.

Collective adoption of AI and ML tools may introduce new systemic risks. If e.g. a critical segment of financial institutions rely on the same data sources and algorithmic strategies, under certain market conditions a shock could affect this entire segment and thus spread the impact of the shock throughout multiple financial institutions. Without sufficiently long and diverse time series or feedback from actual stress events, it is possible that tail risks are not spotted in time. The current regulatory framework does not sufficiently address the field of AI and ML and therefore needs to be revised and updated. This is perceived necessary to address all new risks at hand, as well as the challenges presented regarding the application of the three lines of defence model. In this effort, regulators might leverage on the existing regulation for e.g. credit risk modelling. Risk managers should follow the developments in this field closely, to be able to assess the (new) risks within individual institutions and for the financial system as a whole. Also, sufficiently skilled resources should be available within the internal and external audit community, as to ensure the proper auditing of the techniques deployed by banks.

Taking into account the risks, the application of AI and ML could be expanded in the area of market risk, liquidity risk, model risk management, stress testing and in the third line. Also, the use of AI and MI to manage tail risk could be further investigated. Another area to monitor and possibly further investigate is the role of BigTech companies and their duality in being suppliers of AI and ML technology as well as competitors of banks. Given the expanding use of AI and ML techniques, new issues and risks will undoubtedly emerge and may warrant further research. It is key that existing governance is strengthened and adjusted following these new issues and risks.