Fed, OCC and FDIC examiners are asking lenders of every size how they audit, constrain and shut down autonomous AI. Most banks are not ready with an answer.
Adapted with permission from reporting by Rob Robinson, ComplexDiscovery (original article, Oct. 25, 2026).
The three federal bank regulators have started asking a question many institutions haven’t asked themselves. If an AI agent goes off script, can you stop it, and can you prove what it did?
According to a Reuters report on Oct. 13 citing people familiar with the inquiries, the Federal Reserve, the OCC and the FDIC are putting written and verbal questions to large, midsize and smaller banks. The questions cover:
- What limits are set on AI systems
- Who monitors and approves what those systems do
- Whether the bank can track and audit agent decisions
- What keeps an agent inside its authority
- Whether humans stay involved in key decisions
- How the bank would shut autonomous systems down
Sources also said the questions extend to the use of Chinese-developed technology and to AI agents in cyber defense. The agencies declined to comment. The sources described the effort as information gathering, not new expectations.
The Guidance Gap is Deliberate
Agentic AI sits outside the agencies’ newest model risk guidance by design. On April 17 the Fed, OCC and FDIC replaced SR 11-7 with revised guidance, issued by the Fed as SR 26-2. A footnote in it excludes generative and agentic AI from its scope. The guidance is explicitly nonbinding. Even so, supervisory action can still follow from unsafe or unsound practices tied to weak model risk management, and the Fed expects the letter to matter most above $30 billion in assets.
Fed Vice Chair for Supervision Michelle Bowman said in April that the revised guidance now applies narrowly to traditional models and basic AI. She said she expects other risk management and governance practices to carry generative and agentic AI.
The agencies said in April that a request for information on model risk management and AI, including agentic AI, would come “in the near future.” As of Oct. 24 it had not appeared in the Federal Register.
The Shutdown Question Hits a Known Weak Spot
In a Wolters Kluwer survey of 230 U.S. banking professionals fielded in April, 34% named model kill-switch protocols as the AI incident response area where their institution is least prepared. That measures where bankers feel least ready, not how many banks lack a kill switch. Wolters Kluwer also sells regulatory and risk software to banks.
Three older problems sharpen the risk, because an agent acting on bad data or inside a brittle system can repeat a mistake many times before anyone notices:
-
- Legacy integration. In Bank Director’s 2026 Technology Survey of banks under $100 billion, 69% ranked integrating new technology with existing systems as a top concern. In September, regulators called core providers community banks’ most material and highest-risk third-party relationships.
- Thin expertise. In Bank Director’s 2026 Risk Survey, a third of senior executives and directors said they do not understand agentic AI at all.
- Data governance. Data quality, model validation and dependence on external data ranked among the top barriers to scaling AI safely across several industry surveys.
Agent Logs Are Discoverable Evidence
The records a supervisor wants are the same records a litigant can request. They show what an agent did, on what data, and with whose approval. Under the Federal Rules of Civil Procedure, agent logs are electronically stored information, and failing to preserve them can bring court sanctions.
There is precedent for preservation orders. In May 2025, a federal magistrate judge in The New York Times v. OpenAI ordered OpenAI to preserve output logs it would otherwise have deleted. That ongoing obligation ended under an October 2025 stipulation. The case did not involve a bank, but it shows a court ordering AI logs kept despite user deletion requests and privacy objections.
Third-party Rules Are Still Taking Shape
On Sept. 11, the Fed, FDIC, OCC and NCUA proposed new principles-based third-party risk management guidance. Comments are due Nov. 16. Credit unions face an extra gap. GAO found in May 2025 that the NCUA, alone among the banking regulators, lacks authority to examine technology service providers.
Adoption is moving faster than controls. In Bank Director’s survey, 72% of banks had implemented generative AI, 30% reported agentic AI use, and 88% had drafted an AI acceptable use policy. A policy on paper is not a shutdown procedure.
What Boards Should Ask Now
-
- Do we have an inventory of every AI agent, what systems each can reach, and what it is authorized to do?
- Who can shut each one down, how fast, and when did we last test it?
- Do our retention schedules and legal hold procedures cover agent logs, including logs our vendors hold?
- Do our vendor contracts require the vendor to preserve and export those logs and to disclose where its models come from?
If an agent acted without approval tomorrow, how long would it take you to notice, and could you show an examiner or a judge exactly what happened?
Sources: ComplexDiscovery (Rob Robinson); Reuters via U.S. News & World Report; Federal Reserve SR 26-2; OCC; Federal Register; Bank Director; Wolters Kluwer; GAO. Full source list in the original article.