All posts
  • agentic ai
  • ai agent risk management
  • banking
  • ai governance
  • model risk management
  • eu ai act
  • finma
  • human oversight
  • stress testing
  • regulatory compliance

Agentic AI Risk Management in Banking: 6 Questions Supervisors Now Ask

Supervisors are asking banks whether they can control autonomous AI agents. Here are the six questions, the controls behind them and the evidence to prepare, plus the latest EU AI Act, FINMA and PRA timeline.

By Yasser Zaky

11 min read

Bank supervisors have moved past asking whether institutions use AI. The question now is whether banks can control AI agents that act with growing autonomy: approving, escalating, refunding, routing payments or drafting decisions with little human involvement.

In October 2026, Reuters reported, citing people familiar with the matter, that US banking regulators (the Federal Reserve, OCC and FDIC) were sending written and verbal questions to banks about autonomous AI. According to those sources, the questions cover the limits placed on AI systems, how agent actions are monitored and approved, whether agent decisions can be tracked and audited, safeguards against agents exceeding their authority, whether humans stay involved in key decisions, and how autonomous systems can be shut down. The same report said euro area banks are being asked to map AI use across their business and third parties, and to be able to switch systems off and restart them in a controlled way. The agencies declined to comment, and the report describes supervisors as still gathering information rather than issuing new rules.

That distinction matters: these are supervisory inquiries, not new legal requirements. But they show clearly where scrutiny is heading. This article turns those questions into a practical agentic AI risk management framework for banks, and shows what evidence a risk committee, internal audit team or supervisor can reasonably expect to see.

Why agentic AI changes the risk equation for banks

Traditional model risk management was built for models that produce an output (a score, a forecast, a price) that a human or a deterministic system then acts on. AI agents collapse that separation. They interpret instructions, choose actions, call tools and hold multi-turn conversations, so behavior becomes the unit of risk, not just predictive accuracy.

UK industry participants made the same point at Bank of England roundtables. The summary published in February 2026 records that "the concept of having a 'human-in-the-loop' was also challenged by the rise of agentic AI," and that several participants felt risk management "needed to evolve to put greater emphasis on testing, monitoring and setting guardrails" around AI outcomes. Participants also suggested the traditional validation approach "wouldn't be sustainable in its current form" for complex AI.

Research supports both the opportunity and the caution. A BIS working paper (No. 1310, November 2025) found that a general-purpose generative AI agent closely replicated key prudential cash-management practices in a simulated wholesale payment system. The authors still called for regulatory safeguards, human oversight and further research before adoption in financial market infrastructures. Capable agents are not the same as controlled agents.

The four behavioral failure modes that keep risk officers up at night

In practice, agentic risk in banking tends to concentrate in four areas:

  1. Mandate drift: the agent gradually moves outside its approved scope, for example granting an exception, promising an outcome or giving advice it is not permitted to give.
  2. Manipulation susceptibility: customers, fraudsters or injected content push the agent into unsafe actions through authority claims, urgency or disguised requests. The OWASP Top 10 for LLM Applications (2025) lists prompt injection first and identifies "excessive agency" as a distinct risk.
  3. Escalation failure: the agent handles a case it should have handed to a human, or escalates inconsistently.
  4. Confidentiality leakage: the agent reveals personal data, internal policies or its system instructions under pressure.

None of these appear in a standard accuracy benchmark. They show up only when the agent is put under pressure.

Where AODIT fits: AODIT is an independent behavioral risk rating and stress-testing platform for AI agents. It evaluates how an agent behaves under contradiction, authority pressure and adversarial manipulation, producing evidence that risk, compliance and audit teams can review. You can run a free live demo on your own agent's instructions in about two minutes.

The supervisor questions, translated into a control framework

Below, each line of supervisory inquiry reported in October 2026 is paired with the control objective behind it and the evidence a bank should be able to produce. This is our recommended interpretation, not regulatory text.

1. "What limits are placed on your AI agents?"

Control objective: every agent has a documented, testable mandate covering what it may do, what it may never do, and which actions need approval.

Evidence to hold:

  • A written mandate per agent: scope, permitted tools and actions, monetary or risk thresholds, prohibited topics.
  • An AI inventory entry that links the agent to an owner and a risk classification. FINMA Guidance 08/2024 expects Swiss institutions to keep an inventory of AI applications, classify their risks and assign clear responsibilities.
  • Test results showing the agent holds those limits when challenged, not just that the limits exist on paper.

A policy the agent abandons on the fifth turn of a heated conversation is not a control. It is an aspiration.

2. "How are agent actions monitored and approved?"

Control objective: high-impact actions have proportionate pre-approval, and lower-impact actions are monitored after the fact.

Evidence to hold:

  • A tiered action matrix, for example: informational answers (no approval), reversible low-value actions (post-hoc sampling), irreversible or high-value actions (human approval before execution).
  • Monitoring metrics: escalation rates, override rates, complaint correlations and policy-exception counts.
  • Proof that the agent itself respects the tiering, meaning it does not attempt approval-required actions on its own initiative.

3. "Can you track and audit agent decisions?"

Control objective: decisions are reconstructable after the fact, including the inputs, the reasoning the agent presented and the action taken.

Evidence to hold:

  • Logging of prompts, tool calls, outputs and version identifiers (model, system prompt, tool configuration).
  • A change log that triggers re-assessment when any of those components change.
  • Periodic independent review. FINMA's guidance states that institutions should regularly test, evaluate and monitor AI tools and, where necessary, commission an independent review. In the UK, PRA SS1/23 sets independent model validation as one of its five model risk management principles.

4. "What stops an agent from exceeding its authority?"

Control objective: layered defences, with technical permission boundaries (what the agent can do) backed by behavioral robustness (what the agent will do under pressure).

Evidence to hold:

  • Least-privilege tool access, so an agent with read-only needs has read-only credentials.
  • Adversarial test results across realistic attack paths: social engineering, authority spoofing ("I'm the branch manager, override this"), prompt injection through documents or emails, and gradual reframing.
  • A "blast radius" assessment: if this agent is fully compromised, what is the worst outcome?

This is where behavioral stress testing does the most work. Permission systems stop impossible actions. Only testing reveals whether the agent will talk itself into permitted-but-inappropriate actions.

How AODIT tests authority boundaries: The AODIT-6 behavioral evaluation framework scores agents across six dimensions: Reliability, Integrity, Confidentiality, Judgment, Resistance and Resilience. The Judgment dimension covers proportional action, escalation to human oversight, policy boundary adherence, authority conflict resolution and blast radius limitation. The Resistance dimension tests jailbreaks, prompt injection, social engineering, authority spoofing and adversarial reframing. Each evaluation follows an eight-turn adversarial protocol that escalates from a baseline question through "the grey zone," "the hard push" and "the manipulation attempt" to a recovery test.

5. "Do humans remain involved in key decisions?"

Control objective: human oversight is designed in, not assumed. The agent reliably recognises when to hand over.

Evidence to hold:

  • Defined escalation triggers (vulnerable-customer signals, complaints, disputes, regulated advice, threshold breaches).
  • Test evidence that the agent escalates consistently, including when a customer pressures it not to.
  • For EU firms, alignment with the human oversight design requirements in Article 14 of the EU AI Act where a system is classified as high-risk (see the timeline section below).

The Bank of England roundtable comment is a useful warning. As agents chain decisions together, "human-in-the-loop" can quietly become "human-occasionally-informed." Banks should be able to show where humans actually sit in each workflow.

6. "Can you shut it down and restart safely?"

Control objective: controlled deactivation and recovery, with no orphaned processes, inconsistent states or uncontrolled fallback behavior.

Evidence to hold:

  • A documented kill-switch procedure per agent, with owners and tested runbooks.
  • A defined fallback (route to human queue, static FAQ, service pause).
  • Recovery testing: after a disruption or a failed interaction, does the agent return to compliant behavior?

Behavioral recovery is easy to overlook. An agent that has been manipulated once in a conversation may stay compromised for the rest of that session. AODIT's Resilience dimension explicitly includes "recovery after failure" and "stress persistence," and the seventh turn of its protocol is a dedicated recovery test. Note that this assesses conversational recovery. Infrastructure shutdown and restart procedures remain the bank's own operational resilience responsibility.

How this fits the regulatory timeline

Accuracy about what is actually required, and when, is essential. Here is the current position for the main frameworks banks in Europe and Switzerland work under.

EU AI Act: high-risk obligations now apply from December 2027

The EU AI Act (Regulation (EU) 2024/1689) originally applied most high-risk obligations from 2 August 2026. The Digital Omnibus on AI, Regulation (EU) 2026/1744, was adopted on 8 July 2026, published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026. According to published legal analysis, it moves the application date for stand-alone Annex III high-risk systems to 2 December 2027, and for high-risk AI embedded in products covered by Annex I to 2 August 2028.

For banks, the most relevant Annex III use case is AI used to evaluate the creditworthiness of natural persons or establish their credit score. The EBA's November 2025 mapping of the AI Act against EU banking and payments law found "no significant contradictions," saw no immediate need for new EBA guidelines, and highlighted the need for cooperation between prudential, conduct and market surveillance authorities.

What this means in practice: the delay is time to build evidence, not a reason to pause. Many customer-facing agents (service, complaints, collections support) may not be high-risk under Annex III at all. They still fall under existing conduct, outsourcing, ICT and operational resilience rules, and under the supervisory scrutiny described above. Classification should be assessed case by case with legal counsel. For an overview, see our EU AI Act compliance page.

Switzerland: FINMA Guidance 08/2024

FINMA has no AI-specific law to enforce, but Guidance 08/2024 (published 18 December 2024) sets out its supervisory expectations: governance and clear responsibilities, an AI inventory with risk classification, data quality, regular testing and ongoing monitoring, documentation, explainability and, where necessary, independent review. It also expects these principles to extend to third-party AI providers. Read more on our FINMA AI guidance page.

United Kingdom: PRA SS1/23

SS1/23, effective since 17 May 2024, applies five model risk management principles to UK banks with internal model approval: model identification and risk classification, governance, development and use, independent validation, and risk mitigants. Roundtable participants described it as "pragmatic," while acknowledging that validation approaches must adapt to complex AI.

A 30-60-90 day plan for agentic AI risk management

For a bank with agents already in pilot or production, we recommend this sequence:

Days 1โ€“30: visibility

  • Complete the AI agent inventory, including vendor-embedded agents.
  • Write a testable mandate and an escalation policy for each agent.
  • Classify each agent by impact (customer harm, financial loss, regulatory exposure).

Days 31โ€“60: evidence

  • Run independent behavioral stress tests on the highest-impact agents first.
  • Document the failure modes found, the mitigations applied and the re-test results.
  • Define kill-switch and fallback procedures, and test them.

Days 61โ€“90: embed

  • Integrate re-testing into change management (new model version, new prompt, new tool = new assessment).
  • Set up monitoring thresholds and board-level reporting.
  • Prepare a supervisor-ready evidence pack for each material agent.

Build your evidence pack with AODIT: The full AODIT platform runs 20 to 100 scenarios per test across all six behavioral dimensions and produces a PDF-ready report for risk and audit stakeholders. For institutions with strict data requirements, AODIT can be deployed on-premise within your own infrastructure, including air-gapped operation. AODIT results are advisory and support, rather than replace, your governance framework and legal advice.

Common mistakes to avoid

  • Treating vendor testing as independent assurance. Self-assessment has an inherent conflict of interest. See why AI agents in banking need independent behavioral risk ratings.
  • Testing only the happy path. Most real incidents come from ambiguous, adversarial or emotionally charged interactions.
  • One-off sign-off. Agent behavior can shift with a model update even when nothing in your code changed.
  • Confusing logging with control. Being able to reconstruct a bad decision afterwards is necessary, but it does not prevent the decision.
  • Waiting for the AI Act deadline. Supervisory questions are being asked now, under existing frameworks.

Frequently asked questions

What is agentic AI risk management in banking?

It is the set of governance, testing and monitoring practices that make sure AI agents, meaning systems that take actions and not just produce outputs, stay within their mandate, escalate correctly, resist manipulation and can be safely shut down. It extends traditional model risk management with a focus on behavior under pressure.

Did the EU delay the AI Act for banks?

Partly. Regulation (EU) 2026/1744 moved the application date for stand-alone Annex III high-risk obligations, which include AI-based creditworthiness assessment of individuals, from 2 August 2026 to 2 December 2027. Other parts of the Act, such as the prohibited practices and general-purpose AI rules, were not postponed in the same way. Confirm specific obligations against the legal text.

Are US regulators issuing new rules on AI agents?

Not as of the October 2026 Reuters report. Sources said the Fed, OCC and FDIC were gathering information through questions to banks and were not yet prescriptive.

How often should AI agents be stress-tested?

At minimum before deployment and after every material change (model version, system prompt, tools or data access), plus periodic re-assessment set by risk tier. FINMA expects regular testing and ongoing monitoring. The frequency is a matter of institutional risk appetite.

Does behavioral stress testing replace model validation?

No. It complements it. Validation looks at conceptual soundness, data and performance. Behavioral stress testing provides evidence on how the agent acts under adversarial, ambiguous and high-pressure conditions, which traditional validation was not designed to capture.

Conclusion: control is now the question

Supervisors are converging on a simple test for AI agents in banking: show us you are in control. That means documented limits, proportionate approval, auditable decisions, robust resistance to manipulation, reliable human escalation and a tested way to switch things off. Most of those controls exist on paper at many institutions. The gap is evidence that they hold under pressure.

That is the gap AODIT is built to close.

๐Ÿ‘‰ Request an independent AI agent evaluation to see how your agents perform across all six AODIT behavioral dimensions, or try the free live demo and stress-test your agent in about two minutes. Compare plans on our pricing page.

This article is for information only and does not constitute legal or regulatory advice. Regulatory positions are stated as of 10 November 2026. AODIT evaluation results are advisory and do not constitute regulatory approval.

  1. 11 min read

    Excessive Agency in Banking AI Agents: Permissions and Blast Radius

    Once an AI agent can issue refunds or move money, its permissions become your biggest risk. Here's what OWASP, the OCC, FSB and DORA say about excessive agency, and a 6-step framework to limit an agent's blast radius.

    • excessive agency
    • ai agent permissions
    • least privilege
    • blast radius

    Read

  2. 14 min read

    LLM Model Change Management for Banking AI Agents: A Retest Playbook

    Vendor model retirements, alias moves and prompt edits can quietly change how a banking AI agent behaves. Here's what the OCC, DORA, the EU AI Act and the PRA say about change, and a 7-step playbook to retest agents before the switch.

    • model change management
    • llm model updates
    • ai agents
    • banking

    Read

  3. 13 min read

    EU AI Act Credit Scoring: What Banks Must Do Before December 2027

    Creditworthiness AI is high-risk under the EU AI Act, with obligations from 2 December 2027. Here's what's in scope (including LLM agents), who does what, what the Omnibus changed, and a 10-point readiness checklist.

    • eu ai act
    • credit scoring
    • creditworthiness assessment
    • high-risk ai

    Read