17
By Tobias Adrian, Financial Counsellor and Director of the Monetary and Capital Markets Department, International Monetary Fund (IMF)
Artificial intelligence (AI) is rapidly becoming part of the financial system. Banks are using it to assess borrowers, detect fraud and monitor risks. Investors are deploying it to analyze markets and execute trades. Supervisors are increasingly relying on AI-powered tools to process vast quantities of data and identify emerging vulnerabilities.
The promise is enormous. AI can improve productivity, lower costs, expand access to financial services and strengthen risk management. It can help financial institutions process information at a scale and speed previously unimaginable.
Yet the implications go beyond efficiency. AI is not simply another technology layered onto existing financial processes. It can change how decisions are made, how quickly they are made and how similar those decisions become across institutions. As a result, AI is becoming a financial-stability issue.
The key question for policymakers is no longer whether individual firms can use AI safely. It is whether the financial system as a whole will remain resilient when many firms rely on similar models, depend on the same technology providers and react to the same signals at increasing speed. The answer will hinge not only on innovation, but also on governance, oversight and international cooperation.
Why AI changes the financial-stability conversation
Financial systems have always evolved alongside technology. But AI differs from previous innovations because it compresses both time and distance. Decisions that once took hours, days or weeks can now occur in seconds, while vast amounts of information can be processed instantaneously across institutions and markets.
This acceleration brings benefits. Markets may become more efficient. Credit decisions may become faster. Supervisors may be able to detect risks earlier. But it also creates new channels through which disruptions can spread.
As AI becomes more widely embedded across financial institutions, risks can become correlated in ways that are difficult to detect. Individual firms may appear well managed, yet the system may become vulnerable because many firms are making similar decisions at the same time. The lesson from previous episodes of financial instability is that system-wide risks often emerge not from the failure of a single institution but from collective behavior. AI has the potential to reshape that collective behavior.
Three areas illustrate both the promise and the challenge: lending, trading and risk management.
Three areas illustrate both the promise and the challenge: lending, trading and risk management.
Smarter lending—but more synchronized credit cycles?
Lending is one of the most consequential applications of AI in finance. AI systems can help loan officers and credit analysts evaluate borrowers more efficiently. They can analyze financial statements, process large volumes of information, identify patterns that traditional credit models may overlook and incorporate alternative data sources into risk assessments.
The benefits could be substantial. Borrowers may receive decisions more quickly. Institutions may reduce operating costs. Credit could become more accessible for households and small businesses that have historically lacked extensive credit histories. Over time, AI may transform lending from a largely human-driven process into one in which intelligent systems increasingly support—and in some cases automate—multiple stages of credit assessment.
But financial-stability concerns arise when multiple lenders begin relying on similar models and data.
One challenge is explainability. Some advanced AI models operate as “black boxes”, making it difficult for managers, boards, supervisors or even developers to understand why a particular lending decision was made. If institutions cannot explain the drivers of their credit decisions, they may struggle to identify concentrations of risk or emerging vulnerabilities.
This concern is likely to deepen as lending becomes more agentic. In “agentic lending”, AI systems do not merely support human credit officers; they increasingly initiate, adapt and execute parts of the lending process themselves. Such decisions may evolve in ways that are difficult for firms and supervisors to understand, reinforcing concerns about opacity and explainability.
Greater concern may emerge during economic downturns. Under favorable conditions, AI models may appear highly effective because they are trained on abundant recent data. The real test comes when conditions deteriorate unexpectedly. Models may be less effective at recognizing rare events, severe recessions or structural economic shifts that are underrepresented in historical data.
If many institutions use comparable AI systems and those systems detect rising risks at the same time, lending behavior could become more synchronized. Credit standards could tighten simultaneously, amplifying downturns and reducing financing precisely when households and businesses need it most. Financial cycles have always involved swings between optimism and caution. AI has the potential to amplify those swings if common models generate common responses.
The goal for policymakers should not be to prevent innovation in lending. Rather, it is to ensure that institutions understand the limitations of their models, maintain strong governance and test systems rigorously under adverse conditions.
Smarter trading—but more synchronized markets?
AI is also reshaping trading and investment. Investment firms increasingly use machine learning (ML) and generative AI (GenAI) to analyze earnings calls, regulatory filings, news reports, market sentiment and a growing range of alternative data sources. These capabilities allow investors to process information faster and develop trading strategies more efficiently.
Under normal market conditions, the effects are largely positive. AI-powered trading can improve liquidity, reduce transaction costs and accelerate the incorporation of information into asset prices. Yet speed can become a source of vulnerability during periods of stress. Financial markets are already highly automated in many segments. AI may extend sophisticated automated trading into a broader set of markets, including less liquid asset classes that historically have been dominated by slower-moving investors.
This raises important questions about market functioning. During episodes of market turbulence, investors often react to similar information. If many AI-driven trading systems respond similarly to new information, they may generate increasingly similar trading strategies. Such herding behavior is not new. Investors have always chased trends and crowded into popular trades. But AI changes the speed and scale at which these dynamics can occur.
Research suggests that some AI-based investment strategies rebalance positions more rapidly than traditional approaches. If many systems respond simultaneously to the same signal, market movements could be amplified: price declines could accelerate; liquidity could evaporate more quickly; and volatility could increase.
These dynamics could become more complex as trading becomes more agentic. In “agentic trading”, generative-AI systems do not simply analyze information or execute predefined strategies; they can adapt strategies, infer the behavior of other market participants and respond dynamically to changing conditions. If many such systems were to interact in the marketplace, they could create system-wide risks through tacit collusion, correlated behavior and strategic complementarities—whereby one system’s decision to buy, sell or withdraw liquidity makes it more attractive for others to do the same. The result could be new forms of systemic risk as many sophisticated AI systems reach similar conclusions at nearly the same moment.
Less liquid markets could be particularly vulnerable. While mechanisms such as circuit breakers have helped stabilize some equity markets during periods of extreme volatility, similar protections may be less effective in markets in which trading is naturally thinner.
The challenge for regulators will be to understand these emerging dynamics before stress arrives. Monitoring AI-driven strategies, mapping correlations across firms and incorporating AI behavior into stress-testing exercises will become increasingly important.
Financial institutions may also need stronger safeguards, including clear escalation procedures, kill switches and human-oversight mechanisms that allow automated systems to be slowed or interrupted when market conditions become unstable.
Shared technology—but greater concentration risk?
Some of the most important AI-related risks may not arise from the algorithms themselves but from the infrastructure that supports them. Modern AI systems rely heavily on cloud computing, specialized data providers and advanced model developers. Many financial institutions depend on a relatively small number of providers with the scale and expertise needed to support sophisticated AI applications. These dependencies can remain largely invisible at the level of individual firms.
A single institution may view its technology arrangements as prudent and diversified. Yet when many institutions rely on the same providers, concentration risk can emerge across the financial system. A disruption at a critical cloud provider, AI platform or data service—whether caused by operational failure, cyberattack, geopolitical tension or technical malfunction—could affect many institutions simultaneously.
This type of systemic dependency is not unique to AI, but AI may deepen it.
Recognizing the growing importance of these risks, several authorities have begun expanding operational-resilience frameworks to include critical third-party service providers. The objective is not to eliminate reliance on external providers but to ensure that essential services remain resilient even during disruptions. Policymakers increasingly need a system-wide perspective. Understanding how institutions are connected through common providers may be as important as understanding their direct financial linkages.
Better risk management—but new governance challenges?
AI is also transforming how risks are managed and how financial systems are supervised.
Financial institutions are using AI to identify anomalous transactions, strengthen compliance functions, improve model validation, detect operational weaknesses and enhance cybersecurity monitoring.
For supervisors and central banks, AI-powered supervisory technology (SupTech) offers considerable promise. Machine-learning tools can help authorities monitor markets, analyze large datasets, identify emerging vulnerabilities and allocate supervisory resources more efficiently. As financial systems become more complex, these capabilities may strengthen oversight and improve early-warning systems.
However, greater reliance on AI also introduces governance challenges. Supervisors must avoid becoming overly dependent on automated outputs. Models can perform well under normal conditions yet fail in unexpected ways during stress. Strong human judgment remains essential. Several authorities have embraced an important principle: AI should augment supervisory judgment, not replace it.
Maintaining that balance will require investments in technical expertise, effective governance frameworks and mechanisms that ensure transparency and accountability.
Cyber threats at machine speed
Perhaps the most rapidly evolving risk is cybersecurity. Generative AI is changing both sides of the cyber-defense equation. Defenders can use AI to identify vulnerabilities, monitor networks, detect suspicious activity and respond more rapidly to incidents. Cybercriminals can employ many of the same technologies to automate attacks, develop convincing phishing campaigns and exploit vulnerabilities more quickly. The result is a race between attackers and defenders that is increasingly measured in minutes rather than months.
As the interval between identifying a vulnerability and exploiting it continues to shrink, institutions have less time to respond. Speed itself is becoming a core component of resilience. This evolution has important implications for financial stability.
Cyber incidents are no longer simply operational events affecting individual firms. Large-scale disruptions can undermine confidence, interrupt critical financial services and create broader economic consequences. As AI expands the capabilities of malicious actors, cyber resilience is becoming a macro-financial issue.
For policymakers, priorities include strengthening cyber-resilience expectations, conducting system-wide exercises that incorporate AI-enabled attack scenarios and improving information sharing among institutions and authorities.
Investing in defensive AI will become increasingly important. But technology alone will not be sufficient. Governance, preparedness and coordination will remain critical.
Harnessing AI without sacrificing stability
The debate about AI in finance is often framed as a choice between promoting innovation and managing risk. In reality, those objectives are complementary. Financial systems are most successful when innovation occurs within frameworks that preserve trust, resilience and stability.
The path forward rests on four principles.
The path forward rests on four principles:
Governance matters. Institutions need clear accountability for AI-driven decisions, robust model-risk management and strong oversight by senior management and boards.
Transparency matters. Policymakers and supervisors need better visibility into where AI is being used, how it is affecting decision-making and where systemic dependencies may be forming.
Resilience matters. Financial institutions must prepare for operational disruptions, model failures and cyber incidents in a world in which risks can spread at unprecedented speed.
Cooperation matters. AI-related risks do not stop at national borders. Because financial markets and technology providers operate globally, international collaboration will be essential for managing common vulnerabilities and sharing best practices.
The impact of AI on financial stability is not predetermined. AI could make the financial system more efficient, inclusive and resilient by improving capital allocation, supervision and risk management. But these outcomes will not emerge automatically.
Whether these benefits are realized will depend on the choices made today by financial institutions, technology providers, supervisors and policymakers. If those choices are made wisely, AI can support both innovation and financial stability. Otherwise, future episodes of financial instability may unfold with greater speed, stronger correlations and fewer opportunities for intervention.
There is no doubt that AI will transform finance. The real test is whether it strengthens financial stability—or undermines it.
ABOUT THE AUTHOR
Tobias Adrian is the Financial Counsellor and Director of the Monetary and Capital Markets Department at the International Monetary Fund (IMF). He leads the IMF’s work on financial-sector surveillance, monetary and macroprudential policy, digital money, financial regulation, capital markets, debt management, bank resolution and climate finance.