Skip to main content
LLM Squared

For bank and credit union risk and technology owners

Agent authority review

Who approved the agent, what may it do, and who can stop it?

Prepare the authority record before an agent launch, a wider permission or a vendor change.

When an agent needs a review

A drafting assistant produces an output. An agent may also call tools, retrieve records or take steps in a workflow. The review begins with the actions your institution is considering, rather than a broad promise about AI safety.

Use this review when a launch is near, a vendor adds agent functions or a team asks for wider access. Include security agents that identify or respond to threats. Keep each action tied to the person who may approve it and the systems it may reach.

What state examiners review

The OCC framework uses Tier 1 for its most critical examples. The CSBS worksheet uses Tier 3 for its highest risk tier.

OCC Bulletin 2026-54 gives examiners a tool for review. It sets no supervisory expectations for banks. Checked November 10, 2026.

CSBS procedure GE-8 covers agents that act with limited human direction. It asks about allowed actions, human checkpoints, logging, reversal and the ability to restrict or halt the system.

The framework is discretionary and creates no new legal obligations. Each state decides its use. National banks follow OCC oversight; credit unions follow NCUA oversight. Use the questions to prepare records, rather than predict an exam result.

Core Examiner Guide, page 7 and Examiner Work Program, GE-8. Checked October 16, 2026.

What federal regulators are asking

Reuters reported on October 13, 2026 that the Federal Reserve, OCC and FDIC were asking banks about AI agents. Topics included action limits, approvals, audit trails, human involvement and shutdown. The report also covered model origin and agents used in cyber defense.

The agencies declined to comment. Reuters described information gathering, rather than new rules. Treat the report as news, separate from published agency guidance.

Reuters, carried by Investing.com, October 13, 2026. Source checked October 16, 2026.

What the review puts on file

  • Actions and limits: Which tasks are allowed, which systems or records are reachable and where permission ends.
  • Approval and monitoring: Who authorizes each action, what needs another approval and who reviews exceptions.
  • Audit trail: Which records show the instruction, action, time, outcome and human response.
  • Human checkpoints: Where a person must act before the next step and what happens when that person is absent.
  • Shutdown and reversal: Who can stop the agent, remove access and assess whether an action can be reversed.
  • Model origin: Who built the model, where that is established and what remains unknown about providers or locations.

We read the agreed source versions and compare those descriptions. Missing evidence stays open with an accountable role and a question to answer. A statement that a control exists is distinct from evidence that the control was tested.

Records and review steps

Bring the agent description, tool permissions, vendor disclosures and current approval rules. Include any records of access removal, shutdown or reversal that your team may share. Keep confidential material out of the initial inquiry; agree its transfer route first.

  1. Set the action, use and source boundary in writing.
  2. Compare current permissions with the stated business purpose and approvals.
  3. Assign missing evidence and retain returned records with their versions.
  4. Prepare proposed conditions and identify changes that would reopen the review.

Your team keeps the six-column summary, supporting authority record, open actions and decision line. Timing is two to four weeks after agreed inputs and reviewer availability. Sessions, correction rounds and acceptance milestones belong in the written scope.

Who signs off: The institution's authority owner and risk reviewer. The institution records the operating decision and its conditions.

Inspect the agent example

The illustrative row names an action and target system. The authorizer, expiry, revocation and shutdown evidence remain open. Read the five GE-8 items below that row before drawing a conclusion.

Open the agent authority row. The principals coauthored Know Your Agent: The Governance Standard for AI in Financial Services.

Questions about agent evidence

Does an off-switch statement demonstrate a tested shutdown?

No. The statement describes a control. A test record must show what was tested and its result in an authorized setting.

What if the agent uses several model providers?

Keep the known providers and locations separate. Assign missing details to an owner rather than filling unknown fields with assumptions.

Can this review authorize a production agent?

The institution makes that decision. This review prepares the record; runtime testing or enforcement needs a separate authorized scope.

Book an agent review call Review scope and method limits.