OCC Bulletin Tells Bank Examiners to Rate AI Agents by What They Can Do, Q2 Says

Bank examiners have a new framework for judging the risk of AI agents, according to Q2, a vendor writing on its own blog.¹ Q2 says OCC Bulletin 2026-54, issued November 4, 2026, is a supervisory framework for examiners that rates an AI agent's risk by what the agent can do and the consequences of failure, not by how advanced it is.¹

What happened

Q2 says the framework creates no enforceable standards.¹

Sumsub reports that the framework sets out four illustrative risk tiers.² Sumsub says the top tier includes fully autonomous payment agents and AI systems that release fraud holds.² Sumsub also reports that the framework creates no new requirements for banks.²

How sure we are: this piece rests on two secondhand accounts, not on the bulletin itself. Both describe it as a supervisory framework for examiners.¹² They describe its effect on banks in different terms. Q2 says it creates no enforceable standards,¹ and Sumsub says it creates no new requirements for banks.² The two claims are consistent but not identical, so we report each as its own source gives it. The four tiers, and what sits in the top one, rest on Sumsub's report alone.²

What it means for finance leaders

If Q2's account is right, an examiner looking at an agent will weigh what it can do and what would happen if it failed, not how advanced it is.¹ By Sumsub's account, the framework's top illustrative tier includes fully autonomous payment agents and AI systems that release fraud holds.²

That raises a decision for a CFO, chief risk officer, or head of compliance. You can describe each of your agents in those terms now, before an examiner asks, or you can wait.

The strongest case for waiting

The strongest argument against preparing now comes from the reports themselves. Q2 says the framework creates no enforceable standards,¹ and Sumsub says it creates no new requirements for banks.² A bank's leaders could reasonably decide that building an inventory for a framework described this way is effort better spent elsewhere, and answer examiners' questions when they come. A reader may also weigh that one of our two sources is a vendor's blog.¹

We take that view seriously. We have no evidence that a bank that prepares fares better in an exam than one that waits, and nothing below shows it.

Our perspective: an untested hypothesis

Our working view, which we have not tested at any bank, is that a bank can prepare by answering three questions. Each question matches a pattern we have published. The patterns describe practices and give their own status. None of them studies banks or examiners.

• Which agent actions need a person? Our pattern Steps Only a Person May Take says to name the few steps no agent takes, such as granting access, deleting records, and sending words in a person's name. Agents prepare those steps exactly, and a person reads, decides, and takes the step. Its status: Running for access and administrator changes and for the exact text of messages; Early for requests to a person.³ The pattern's examples do not include payments or fraud holds. Whether those belong among the steps only a person may take is part of what our hypothesis would need to test.
• Who is told when an agent fails? Our pattern Failures Reach a Person says to send every failure to the person who can act on it, have every routine confirm that it worked, and notice when something that should have happened did not. Its status: Early for checking results and sending failures to a person; Planned for heartbeats.⁴
• What record lets an action be undone? Our pattern A Record of Every Action, with a Way to Undo It says to keep a record of every action agents take for the organization (when, who, what, where, who approved it, and how to undo it), never of what people say to them, and to "make every change you can undoable in one step." Its status: Running for recording actions; Planned for one log.⁵

We have reported similar thinking elsewhere. Singapore's government described a framework that connects agents' capabilities to safeguards, with stronger controls such as human approval for higher-risk actions.⁶ An opinion piece in Insurance Edge, which we covered, argued that AI agents touching an insurer's reserving report or month-end close need a named human owner.⁷ Those are others' views, not our findings.

What would change our mind: published accounts showing that examiners applying the bulletin weigh something other than what an agent can do and what happens if it fails, or banks reporting that these three questions do not map to the tiers.

More of our coverage is in our Finance section.

One step to take

Pick the agent in your organization that comes closest to moving money. On one page, write down three things: what it can do on its own, who hears when it fails, and how its last action would be undone. If any of the three has no answer, start there.

This is not financial or investment advice.