Financial services firms have always operated on the principle that risk must be understood, governed, and ultimately owned. Agentic AI does not change that accountability, but it fundamentally changes the environment in which it must be exercised. Traditional risk governance was largely designed around mistakes occurring at human speed, with controls capable of identifying and addressing individual decisions. Agentic AI challenges that model because it can act autonomously, choose its own route to an answer, call tools, delegate tasks, and make chains of decisions that were not fully specified when the system was approved.
Financial institutions have become accustomed to governing high-volume, automated systems, from credit models to algorithmic trading and fraud detection. Agentic AI disrupts not only the scale and speed but also the breadth of possible actions. An AI agent’s output space is not fixed in advance, meaning existing controls may struggle to anticipate how it will behave. The result is a widening mismatch between machine-speed risk and human-speed governance.
How AI risk cuts across traditional financial risk silos
Financial services risk governance remains largely organized around discrete categories such as credit, market, fraud, conduct, and operational risk, but agentic AI does not respect these boundaries. A single AI system may influence credit decisions, fraud detection, pricing, customer advice, and operational processes simultaneously. If that system drifts, fails, or is compromised, the effects can propagate across multiple risk categories at machine speed.
This creates a governance gap between the risks for which an institution remains accountable and what its existing controls can address. The report identifies four ways in which AI is changing the risk environment.
Autonomy risk
An AI agent can plan and execute a multistep task, call tools, delegate to other agents, and potentially take different routes when given the same prompt. Rather than validating one model and one output, institutions may need to understand an entire chain of decisions, including what the agent did and how it arrived there. Existing model risk frameworks were not designed for this level of autonomy.
Adversarial risk
Attackers are increasingly targeting AI systems themselves. Techniques can include inputs designed specifically to manipulate a system or hidden instructions embedded in content processed by AI. Any AI system ingesting external material, such as emails, documents, or filings, therefore creates a potential attack surface.
Correlation risk
Multiple institutions deploy AI systems trained on similar data and optimized for similar objectives. Individually rational systems may respond to the same trigger in similar ways, creating the potential for correlated behavior across markets. Historical episodes such as the 1987 market crash illustrate how individually governed systems can collectively amplify instability; AI potentially extends that dynamic across more activities and at greater speed.
Detection lag
AI can execute flawed decisions rapidly and consistently, while the resulting deterioration takes weeks or months to become visible. Credit losses, harmful conduct, pricing errors, or model drift may therefore accumulate before an institution identifies the underlying problem. AI risk can be fast in execution but slow in discovery.
Strengthening AI accountability and oversight in financial institutions
In many institutions, responsibility for AI is distributed across risk, technology, digital, and business functions. Institutions moving fastest are instead identifying a single accountable owner for AI risk.
Existing senior accountability regimes provide a foundation, but institutions also need to connect individual accountability to specific AI systems and the decisions those systems make.
A senior executive should be able to identify which AI systems act on their behalf, what those systems are permitted to do, and who authorized those permissions.
Accountability for the decision, rather than simply ownership of the underlying technology, becomes increasingly important as AI systems become more autonomous.
The instinctive institutional response to a governance gap is to add reviewers, committees, and checkpoints. That approach cannot scale to an environment in which AI generates decisions and interactions continuously. Risk teams cannot manually review every AI-generated credit decision, compliance teams cannot inspect every AI-assisted customer interaction, and fraud teams cannot investigate every anomalous pattern in real time.
Instead, AI must become both a source of risk and a primary tool for managing it. Continuous AI-enabled monitoring for model drift, anomalous behavior, conduct signals, and adversarial patterns can provide the real-time controls required for a machine-speed environment. Project Glasswing, a coalition launched in April 2026, illustrates the potential: Using a frontier AI model, the initiative scanned more than 1,000 open-source projects and identified more than 6,000 high- or critical-severity vulnerabilities, with several partners reporting vulnerability detection rates more than 10 times higher than previous review processes.
This does not make human judgment less important. Rather, it concentrates human attention on a smaller number of consequential decisions where judgment and accountability matter most.
AI governance priorities for financial institutions and boards
For financial institutions, two actions should take priority.
First, institutions should map where AI is influencing risk decisions. This means moving beyond inventories of AI use cases to identify material decisions with financial, conduct, or operational consequences, including credit approvals, pricing, claims, advice, fraud detection, collections, and customer communications. The mapping must also include AI embedded in third-party vendor tools. At the decision level, institutions should understand what influences the decision, what the AI system is permitted to do, who owns the decision, and where the record of its reasoning is retained.
Second, institutions should plan for AI failure rather than assume it will not occur. Financial institutions already provision for credit losses, hold capital against market risk, and stress-test liquidity because eliminating failure entirely is unrealistic. AI governance should apply the same principle. Institutions need to understand how material AI systems behave when they fail, how quickly failure can be detected, what the maximum potential loss could be before intervention, and how affected customers would be made whole.
Boards should be able to answer practical questions about each material AI system. These questions begin with what decisions it influences, who is personally accountable, and where human sign-off is required.
Boards must then also look at which decisions have deliberately been excluded from automation; whether the system fails safely, whether its decisions can be explained after the fact, and how it performs against hostile inputs,
Finally, boards must know the correlated risks created when dependencies exist across common models, vendors, and data. They also must be able to say who can shut the system down, and how quickly.
10 questions boards should ask about any AI system
Boards can use these questions to test whether the right accountability, controls, and safeguards are in place for material AI systems.
- What decision is this AI system making or materially influencing, and what happens to a customer or the balance sheet if it gets that decision wrong?
- Who is personally accountable for the AI system as a whole and the decisions it makes — by name, not by function?
- Does it operate autonomously, or does it require human sign-off, and was that line drawn on purpose or by default?
- Which decisions have we ruled out automating entirely, regardless of performance, and is that documented anywhere?
- If it fails, or starts behaving differently than it did on the day it was approved, does it fail safely, and how long would it take us to notice?
- Can we explain, after the fact, why it made a specific decision?
- What is the largest loss it could produce before a human notices?
- When did we last try to break it, feeding it hostile inputs the way an attacker would, and what happened?
- How many of our other AI systems, and our competitors’ systems, run on the same model, vendor, or data, and what happens if they all move the same way on the same day?
- Who can turn it off, and how quickly?
These questions are increasingly relevant to supervisory expectations across Asia-Pacific. The Monetary Authority of Singapore and the Hong Kong Monetary Authority have both issued guidance directly addressing AI model governance, while Australia’s CPS 230 places board-level accountability on the operational resilience of critical operations. Japan has taken a more consultative approach through the Financial Services Agency’s AI discussion paper and subsequent public-private engagement.
Where regulatory positions remain under development, institutions still need to define, evidence, and ultimately defend their own governance.
Building AI governance for machine-speed risk in financial services
The next phase of AI adoption in financial services will depend on which institutions develop the capabilities required to govern them, and not on which institutions deploy the most capable models. It means knowing where AI is acting, what it is optimizing for, where it can fail, and what those failures could mean when decisions occur at machine speed.
While approaching AI governance primarily as a compliance exercise is likely to result in some good frameworks, those that treat it as a strategic capability can build AI-speed resilience to match AI-speed risk. In financial services, where trust underpins the business model, that distinction matters. Machine-speed risk will not wait for human-speed governance.
