The OCC will judge bank AI by what it can do, not how clever it is
The OCC’s first guidance for examiners on bank AI ranks systems by the consequence of a mistake, how much authority they hold and how much a person still oversees. A simple agent that can move money can outrank a frontier model that only drafts. Banks will need to show, for each system, what it may reach and do, and who can stop it.
What happened
On 4 November the Office of the Comptroller of the Currency issued OCC Bulletin 2026-54, a supervisory framework for how examiners should review banks’ use of AI, which Reuters described as the regulator’s first such guidance. The OCC now calls the technology “super intelligence,” a term it says covers everything previously called artificial intelligence, following an executive order in October. The bulletin is explicit that it “does not establish supervisory expectations for banks” and that banks “are not expected to use this document.” It is a guide for examiners, signed by the Chief National Bank Examiner. Examiners are told not to treat AI as a separate risk, to scale their work to a bank’s size and complexity, and to focus on the consequences of AI-driven mistakes rather than the sophistication of the model. In the OCC’s own example, a simple agent authorized to initiate transactions may be riskier than a sophisticated model used internally.
The framework itself carries the detail that matters for operators. It says a bank may document, before deployment, what a system or agent is designed to do and not do, who or what may invoke it, the data and systems it may reach, the actions it could take and when it needs human review or suspension. That purpose statement becomes “a basis for identifying scope drift, unauthorized use, and control failures.” Use cases are classified by consequence, transfer of agency and the degree of human oversight, in four illustrative tiers. Tier 1, Critical, covers direct action that is hard to reverse, such as an autonomous payment agent, real-time credit decisions or releasing fraud holds. Tier 4, Low, covers internal productivity tools and general employee questions. A fully autonomous payment agent sits at Tier 1 “regardless of how well it performs in testing.”
The framework separates human-in-the-loop, where a person approves before anything is used, from human-on-the-loop, where a person monitors with the authority and practical ability to intervene. It warns that bulk approval can turn review into “a nominal click” and contribute to alert fatigue. It names kill switches, circuit breakers and least-privilege permissions among the ways to contain a misbehaving agent, and suggests testing whether the business process can continue through trained staff or a narrower fallback agent while the main system is stopped. Agents from providers are covered too: a bank may define their authority, identity, tool use and escalation limits. American Banker and Forbes both read the document as a playbook banks should prepare for.
My view
This is the right question, and it moves the work from the model team to the people who own the process.
Don’t be misled by the line saying the framework sets no expectations for banks. I’ve worked inside the Federal Reserve Bank of New York and inside the banks it supervises. An examiner guide tells you what examiners will ask about. A bank that waits for a formal rule will answer these questions during an exam, under time pressure, instead of before one.
For two years most AI governance in banks has been organized around the model: which one, how it was validated, what it scored on a benchmark. That made sense when AI drafted text a person then read. It makes much less sense for an agent that can initiate a payment, release a hold or send a customer a message. The risk in those cases isn’t in the weights. It’s in the permissions.
Authority is the unit of risk
The OCC’s framing makes authority the thing to classify. The payment agent example says it plainly: a strong test score doesn’t lower the tier. Three questions follow for every system: what can it reach, what can it do without asking, and who can stop it. The first is about data entitlements. The second is about delegated actions and their limits. The third is about a working off switch and a person with the standing to use it.
Banks have long answered the first question for their human staff, through entitlement reviews and access recertification. It’s harder for an agent, because the agent often runs with someone else’s access: a developer’s service account, a shared integration credential or the entitlements of whoever started it. When that happens, the agent’s purpose and its reach are different things, and the OCC has just told examiners to treat the gap between them as a sign of scope drift and control failure.
Records decide whether you can show it
The second shift is evidential. If examiners ask what a system did and why, the answer has to exist before they ask. A policy that says agents are reviewed isn’t evidence. A record of each request, the data it touched, the action it took and the person who approved it is. Consequence-based supervision is only as good as the bank’s ability to reconstruct a decision after the fact, and that is mostly a data problem, not a model problem.
The guidance also takes a sensible line on human review. Bulk approvals and alert fatigue turn oversight into theater. A reviewer who sees two hundred agent actions an hour isn’t overseeing anything. Calibrating review to consequence and reversibility, as the OCC suggests, means some actions need no human at all and others need a person before they happen.
Why it matters
For the industry. The OCC supervises national banks, but the logic travels. Any firm putting agents into production will be asked the same three questions by a regulator, an auditor or its own board. Expect inventories that list authority, not just models; agents with their own identities and limits rather than borrowed ones; and pause plans that are tested rather than written down.
How it connects to our work. AI Governance Evidence sets out the request records and release evidence an examiner can inspect: what was asked, what was reached, what was done and who approved. Agent Onboarding covers the other half. Each agent gets a named owner, a stated purpose, its own bounded identity and a revocation path, so the inventory the OCC expects is a byproduct of how agents are admitted, not a spreadsheet built for the exam. In practice that means:
- Classify by authority. Tier each use by what it can do alone and how reversible that is, not by the model behind it.
- Give agents their own access. Purpose, data permissions and action limits are set per agent and checked at the data, not inherited from a person.
- Keep the record as you go. Each consequential action carries its request, inputs, output and approver.
- Test the stop. Pausing an agent should be rehearsed, with the manual fallback known before it’s needed.
Sources
- OCC Bulletin 2026-54: Super Intelligence: Supervisory Framework, Office of the Comptroller of the Currency, 4 November 2026
- Supervisory Framework for Bank Use of Super Intelligence (PDF), Office of the Comptroller of the Currency, 4 November 2026
- US bank regulator issues first-ever AI guidance for examiners, Reuters (republished by 93.3 The Drive), 4 November 2026
- OCC: Banks need strong controls when delegating to AI agents, American Banker, 5 November 2026
- OCC gives examiners a playbook for bank use of “super intelligence”, Forbes, 5 November 2026
Linked sources are the record. Check them for the full reporting.