Talk to us
Research6 min read

What FFIEC examiners are now asking about your AI agents

The examination question has changed from 'do you have an AI policy' to 'show me what this agent did and who approved it.' Here is what examiners ask and the records that answer them.

Written by

Key takeaways

  • The bank examination question about AI is changing. It is no longer only “do you have an AI policy” — a review of an agent workflow is likely to ask “show me what this agent did, who approved it, and how you stopped it.”
  • SR 26-2 covers conventional statistical and quantitative models and excludes generative and agentic AI from its scope. The controls below are Moring’s suggested practices for agent workflows, not requirements imposed by SR 26-2.
  • Supervisory review of automated systems follows an established accountability pattern: attribution, approval, evidence, and control. Banks that already operate this way for payments and trading have a head start on agents.
  • The gap is almost always between policy and evidence. Documents establish intent; per-action records establish execution. Reviews ask for the second.
  • Preparation is a records exercise, not a document-writing exercise: inventory, identity, one-query reconstruction, and a tested stop control.

The one-sentence summary

A supervisory review is likely to treat AI agents as actors that must be attributable, approved, stoppable, and recorded — and the banks that answer from records rather than reconstructions are the ones that prepared.

From “do you have a policy” to “show me the record”

For years, the AI question in a bank examination ended at the policy document: does the bank have an AI governance framework, who owns model risk, are third-party AI vendors in the register. That question is answered by documents, and most banks have the documents.

The 2026 model risk guidance moved one part of that conversation. SR 26-2, issued by the Federal Reserve, OCC and FDIC on 17 April 2026, supersedes SR 11-7 and SR 21-8 and covers traditional statistical and quantitative models, including non-generative AI. It does not set enforceable standards.

Its scope excludes generative and agentic AI, and footnote 3 directs banks to their own risk management and governance practices for tools and systems the guidance does not cover. So an agent is not examined against SR 26-2.

It is examined against whatever the bank itself has decided governs it — which is why the bank’s own record is the thing that gets tested.

SR 26-2 excludes generative and agentic AI. The controls below are Moring’s suggested practices for agent workflows, not requirements imposed by the guidance.

The questions that follow in a review of an agent workflow have a consistent shape:

  • Attribution: which agent took this action, and on whose behalf? If the bank’s answer starts with “the service account,” the answer has already failed.
  • Approval: who approved this agent’s deployment, and who approves its consequential actions — the payments, the data exports, the external communications? What did the approver see?
  • Policy: which version of the policy judged this action, and how do you know?
  • Stop: how is this agent stopped if it misbehaves, who may stop it, and when was that last tested? May the person who stopped it also restart it?
  • Evidence: show me the record for one completed action, end to end.

None of these are answered by the policy document. All of them are answered by records — if the records exist.

The pattern examiners are applying is not new

Banks have been through this arc for automated trading, payment initiation, and straight-through processing. Distributed actors, delegation, high-consequence actions — and a supervisory demand for attributable actors, central rules, approvals on the consequential actions, and records that reconstruct the chain. Agents are a new class of actor, not a new class of accountability.

The practical consequence: banks that operate attributable, approved, recorded action chains for their existing automated systems can extend the same machinery to agents and answer AI examination questions from the same records. Banks that governed those systems by policy document alone will find agents amplify the gap, because agents act more often and in more directions than any earlier automation.

The five records that answer the exam

Preparation is a records exercise. Five artefacts address the questions in the standard set:

  1. The agent inventory — every production agent, its named owner, its workflow, and its approval status. Not a spreadsheet of pilots: the register of actors.
  2. Identity records — each agent’s own credential, scoped and short-lived, with the on-behalf-of chain recorded when it acts for a user. Shared service accounts are the first control gap most reviews surface.
  3. The per-action record — request, policy decision and version, approvals with what the approver saw, tool calls, outcome. Correlated and tamper-evident, retrievable as one query.
  4. The stop-drill record — when the kill switch was last tested, what stopping blocked, how restart was authorised, and by whom. A stop control that has never been tested is a hypothesis; the drill record is what turns it into a control.
  5. The exception log — every policy override and exception, who granted it, and what it covered.

Banks holding these five records can answer review questions during the session. Banks without them have to assemble the answers afterwards, and an unrehearsed reconstruction is what makes a program look ungoverned.

What the preparation actually looks like

The sequence that closes the gap before the exam request letter arrives:

  1. Inventory the agents. Start with IAM exports and workflow owners; the size of the estate sets how long the pass takes. Every production agent gets an owner or gets stopped.
  2. Give each agent its own identity. Shared credentials are the fastest finding to write and the fastest to fix.
  3. Wire the approvals and the record. Approval gates on irreversible action classes, enforced where the action executes; one correlated record per action.
  4. Drill the stop control. Staging first, then production in a quiet window, with the drill record kept. The drill takes about 45 minutes on one agent.
  5. Rehearse the reconstruction. Pick a completed action and reconstruct the chain end to end in one query. If it takes a manual join across log stores, that is the finding — fix it while it is cheap.

The same five steps are what a review follow-up is likely to ask about, which is why banks that run them once run them on a cadence afterwards.

Where this fits with the rest of the regulatory picture

Agent governance does not sit in one regulation. In US banking, SR 26-2 covers the conventional models in and around the workflow but excludes the agents themselves; DORA and the EU AI Act bind EU financial entities to overlapping operational and high-risk obligations, CPS 230 does the same for Australian financial services, and the seven frameworks map onto four recurring controls. Build the controls once; the framework mappings become documentation.

For the scope question in detail, the dedicated SR 26-2 page explains what the guidance covers, what it leaves to the bank and the footnote that puts agents outside it; the readiness checklist scores all four control domains in 36 checks; and what Moring does shows the same records produced as a product.

If you need to brief the board before the exam request letter arrives, the AI governance board briefing covers the fiduciary framing and the five questions directors should ask management. The insurance equivalent — state audits under the NAIC model — is covered in how to prepare for a NAIC AI model audit.

Frequently asked questions

What do FFIEC examiners ask about AI agents?

A supervisory review of an agent workflow is likely to ask about specific agents rather than about the existence of an AI policy: what actions has this agent taken, who approved its deployment and its consequential actions, what record exists of each action, how is it stopped if it misbehaves, and how was the stop tested. The questions follow the structure regulators apply to other automated systems — attribution, approval, evidence and control — applied to a new class of actor. They are not a published FFIEC checklist for agents, and the guidance below is Moring's suggested preparation.

What evidence do banks need for AI agent governance?

Evidence that controls ran, not that they exist: per-action records showing which agent acted, on whose behalf, under which policy version, with which approvals, and what the outcome was — correlated and tamper-evident. Policy documents establish intent; records establish execution. The preparation gap is almost always between the two.

How should a bank prepare for an AI examination?

Before the exam: inventory every production agent with a named owner; verify each has its own identity, scoped credentials and an approved policy; test stop controls and keep the drill record; and confirm one query reconstructs a completed action end to end — request, decisions, approvals, tool calls, outcome. Banks that can reconstruct one workflow quickly answer review questions in the session; banks that cannot may need to assemble the evidence afterwards. The effort depends on the size of the estate and how the records are held.

How does SR 26-2 relate to AI agents?

SR 26-2 (17 April 2026), which supersedes SR 11-7 and SR 21-8, covers traditional statistical and quantitative models, including non-generative and non-agentic AI. SR 26-2 excludes generative and agentic AI. The controls described here are Moring's suggested practices for agent workflows, not requirements imposed by SR 26-2. For agents, the guidance directs banks to their own risk management and governance practices, and it states that it does not set enforceable standards.

What happens if a bank cannot evidence AI agent controls?

That is a preparation gap, and the practical consequences are worth planning for: unresolved questions lengthen review, agent programs may be paused while answers are assembled, and reconstruction can consume staff time if records were never correlated. The cheapest moment to close the gap is before the request letter arrives.

Sources

Federal Reserve: SR 26-2, Revised Guidance on Model Risk Management (17 April 2026), including the scope language and footnote 3 on generative and agentic AI. federalreserve.gov/sr2602a1.pdf. Checked 27 October 2026.

Federal Reserve: SR 26-2 cover letter, superseding SR 11-7 and SR 21-8 and stating that the guidance does not set enforceable standards. federalreserve.gov/SR2602.htm. Checked 27 October 2026.

Architecture & Governance

Governing Autonomous AI Agents in Production

Moring provides the runtime control layer, identity verification, human-in-the-loop approvals, and audit records needed to deploy production-grade AI agents safely.

FFIEC AI examinationbank AI agentsAI agent examination evidencefederal reserve ai governance examinationbank exam preparationAI governance

About the author

Balaji Nagaraj Kumar

Balaji Nagaraj KumarChief AI Officer, Moring AI, Moring AISweden

Leads platform and delivery engineering at Moring AI, turning each engagement into a substrate the customer’s own team operates at handoff. His background spans developer-platform leadership at Volvo Cars and engineering at CERN. He writes about governance for agents that take actions in production systems.

LinkedIn profile