How AI is shaping the future of payment system oversight

In recent years, artificial intelligence (AI) has begun to revolutionise the financial sector. One compelling area of AI’s great potential impact is the oversight of national payment systems as critical infrastructures for the smooth functioning of financial markets. Ensuring the resilience of such systems is essential to maintaining financial stability, since their failures can trigger systemic risks that affect the broader economy. And central banks have been early adopters of AI techniques for supervisory purposes (see chart 1).¹

The use of AI for payment system oversight involves a central bank activity, where the objectives of safety and efficiency are promoted by monitoring existing and planned payment, clearing, settlement and related arrangements, and assessing them against objectives and, where necessary, inducing change.²  

What can AI do for payment system oversight?

Advanced machine learning (ML) algorithms, natural language processing and predictive analytics can significantly bolster the oversight of payment systems by enabling more effective data analysis, real-time monitoring and proactive risk assessment, ultimately leading to improved decision-making and regulatory compliance. Let’s break down some of the ways AI can be applied in this context.

Real-time risk monitoring. AI-driven machine learning algorithms can sift through massive volumes of financial data, identifying patterns, anomalies, or potential threats in real time. This allows overseers and system operators to spot liquidity risks or operational issues that could disrupt market stability before they escalate. These tools can automatically flag unusual transactions, helping prevent potential defaults or market collapses.³

Specifically, when integrated with large-scale payment systems or clearing houses, AI-based models can monitor risks by analysing real-time payment flows, transaction data and liquidity positions across the network of participating institutions. AI-driven algorithms can process and analyse real-time transaction records, payment histories and market conditions, to predict where liquidity pressures might emerge. For instance, if a clearing house detects unusual delays or deviations in a bank’s settlement times, AI tools can flag this as a sign of liquidity strain. This would trigger alerts for overseers, allowing them to investigate and potentially intervene before the situation worsens.

In addition to liquidity issues, AI can also monitor the operational integrity of the systems themselves. AI models can detect early signs of technical malfunctions, transaction processing delays, or communication bottlenecks within these systems. By spotting these disruptions early, overseer and system operators can respond quickly, mitigating risks before they materialise into more serious failures.

Stress-testing with predictive analytics. Traditional stress-testing has limitations in its ability to model extreme events or the complex interconnections between institutions. Stress tests typically rely on historical data and predetermined risk scenarios to evaluate potential vulnerabilities, and are therefore unable to account for the evolving interconnections within the financial system. Specifically, they struggle to model extreme, low-probability events (such as a sudden global liquidity crunch or the collapse of a critical institution) and the interdependencies between financial entities, which can exacerbate a local problem into a broader systemic crisis.

AI-based predictive analytics can overcome many of these limitations by leveraging machine learning and big data to simulate more sophisticated stress-testing scenarios, allowing overseers and system operators to anticipate systemic risks. Examples of these techniques are:

• Dynamic scenario generation: AI models, particularly those using reinforcement learning⁴ and neural networks,⁵ can generate a broader range of stress scenarios by learning from vast datasets that include not just historical events but also real-time market data, macroeconomic indicators and behavioural patterns of market participants. These models can simulate extreme, rare events that would not typically be considered in traditional stress tests. For example, they can simulate how sudden geopolitical events or unprecedented market conditions could trigger liquidity shortages across multiple sectors.
• Modelling interdependencies: AI-driven stress-testing models, such as graph-based models,⁶ can map out and analyse complex interdependencies between financial institutions. These models use network theory to visualise how a shock to one entity might spread to others, causing cascading failures. For instance, an AI model could simulate how the default of a major bank could impact the liquidity of its counterparties, eventually leading to a broader liquidity freeze across the system.
• Handling multivariate risks: AI models can simulate multivariate risk scenarios, combining different types of shocks (eg, a market crash coupled with a cyber attack or a sudden change in fiscal policy). These models use multivariate regression and deep learning techniques⁷ to account for correlations between variables that would typically be ignored by traditional methods.

Cybersecurity enhancement. As financial institutions face ever-growing cyber threats, AI-powered tools are essential in identifying vulnerabilities and preventing cyber attacks. AI-powered anomaly-detection systems can identify unusual patterns or behaviours that may signal a potential cyber attack. AI systems, specifically those using ML, can analyse vast volumes of transaction data and system logs in real time to detect deviations from normal behaviour, and can automate responses to detected threats, providing faster and more efficient mitigation efforts than human-based processes. For example, ML models can simulate different cyber attack scenarios to understand how attackers might exploit certain vulnerabilities in the future. These insights can help prioritise cyber security resources by identifying which systems or processes are at the highest risk of attack, allowing financial market infrastructures to focus their efforts on the most critical areas.

Challenges and risks of AI

Despite the tremendous potential of AI, its integration into financial market infrastructures comes with several challenges and risks. AI systems rely on high-quality and comprehensive datasets to deliver accurate insights. However, in some jurisdictions, financial data may be fragmented, incomplete, or not readily accessible in real time, limiting AI’s effectiveness. Furthermore, issues such as data privacy and sharing between institutions can pose significant barriers. In emerging markets, the lack of standardised data and regulatory infrastructure can make it challenging to implement AI-based oversight.

AI models, particularly deep learning models, can be opaque or ‘black boxes’, meaning it is often difficult to understand how they arrive at certain decisions. This lack of transparency can be problematic in regulatory environments where accountability is crucial. Overseers and financial institutions need to balance the benefits of AI-driven insights with the need for clear and explainable decision-making.

AI models are also prone to replicating biases present in their training data, which could lead to unequal treatment of market participants or flawed risk assessments. Addressing these biases is essential to ensure that AI does not inadvertently reinforce existing inequalities in the financial system.

Finally, while AI can enhance cyber security, it is vulnerable to manipulation. Hackers can target AI systems with adversarial attacks. The Financial Stability Board (FSB) has warned about the growing reliance on AI systems for critical financial functions, emphasising the need for robust defences against potential cybersecurity risks.⁸

The returns can outweigh the risks

Central banks are directly affected by AI’s impact in their role as stewards of financial stability. The potential of AI to improve their payment system oversight capacity is enormous. Central banks can leverage AI tools to bolster the security of their payment systems, and we can anticipate a significant rise in the adoption of these technologies in the coming years.

However, there are clear challenges that need to be addressed to ensure that AI is effectively integrated into payment system oversight. With the right policies in place, it can be a powerful ally in maintaining financial stability and mitigating systemic risks in the increasingly complex global financial landscape.

Notes

1. See Araujo, D, S Doerr, L Gambacorta and B Tissot (2024), Artificial intelligence in central banking, BIS Bulletin, N.84, Bank for International Settlements, January 23.

2.  See Central bank oversight of payment and settlement systems, report by the Committee on Payment and Settlement Systems, Bank for International Settlements, May 2005.

3. For a review of recent attempts at developing AI tools for transactions monitoring in payment systems, see Desai, A, A Kosse and J Sharples (2024), Finding a needle in a haystack: a machine learning framework for anomaly detection in payment systems, BIS Working Papers No 1188, Bank for International Settlements, May 4.

4. Reinforcement learning is a type of machine learning where an agent learns by interacting with an environment. The agent makes decisions to maximise cumulative rewards over time, learning from the consequences of its actions. Through trial and error, the agent adjusts its strategies based on the feedback it receives (rewards or penalties), gradually improving its performance. This approach is widely used in areas such as robotics, game playing and autonomous systems.

5. Neural networks are computational models inspired by the human brain, composed of interconnected nodes (neurons) arranged in layers. Each neuron processes input data and passes the result to the next layer. These networks are designed to recognise patterns and make decisions or predictions by learning from large amounts of data. Neural networks are the foundation of deep learning and are applied in areas such as image recognition, natural language processing and speech recognition.

6. Graph-based models are a class of machine learning models that use graph structures to represent relationships between data points. In these models, nodes represent entities and edges capture the connections or interactions between them. These models are especially useful for analysing complex, interconnected systems such as social networks, biological systems, or financial transactions, enabling tasks like node classification, link prediction and clustering.

7. Deep learning techniques refer to advanced machine learning methods that use artificial neural networks with multiple layers (deep neural networks) to model and understand complex patterns in data. These techniques are particularly effective for tasks like image recognition, natural language processing, and speech analysis, where large amounts of data and intricate patterns are involved.

8. See Remarks on Artificial Intelligence in Finance, by Nellie Liang, US under secretary for domestic finance, and chair of the Financial Stability Board standing committee on assessment of vulnerabilities, at the OECD – FSB Roundtable on Artificial Intelligence in Finance, Paris, May 22, 2024.

This article represents the views of the author and does not necessarily represent those of his employers.