What did the OCC publish on November 4?
The Office of the Comptroller of the Currency published Bulletin 2026-54, titled “Super Intelligence: Supervisory Framework,” on 4 November 2026. It tells examiners how to review AI use at national banks, federal savings associations, and federal branches and agencies of foreign banks. Reuters, which saw the document before publication, reported that the OCC described it as its first guidance on how examiners should probe AI.
The bulletin is not a rule. It states that the framework “does not establish supervisory expectations for banks” and creates no enforceable standards or prescriptive requirements. Reuters noted that such guidance is not legally binding, although banks usually follow it closely. The audience is examiners, not bank compliance teams.
The framework contains sixteen components grouped into six pillars: purpose, risk, control, oversight, resilience and assurance. Its scope runs from credit decisioning and payments to fraud detection, customer service, operations, risk management and market activities. It is meant to supplement existing model risk, governance, technology, compliance and operational-risk reviews rather than replace them.
The document uses “super intelligence” (SI) instead of artificial intelligence. It defines the term by reference to 15 USC 9401(3) and Executive Order 14434, published at 91 Federal Register 63129 on 2 October 2026. Decrypt noted that the administration insists on the label. The framework’s reach still includes ordinary generative AI and autonomous agents, not only hypothetical systems that exceed human intelligence.
James M. Gallagher, senior deputy comptroller and chief national bank examiner, signed the bulletin. A footnote says the framework “was prepared with the assistance of SI” and that humans reviewed, edited and approved the content. The agency also says it will update the guide as the technology, use cases and supervisory understanding change.
How will examiners rate an AI use case?
Examiners are told to weigh three things: the potential consequences of failure, the transfer of agency from people to the system, and the degree of human oversight. The technical sophistication of the model is explicitly not one of those factors. Capability matters less than what the system is allowed to do and what happens if it errs.
The framework’s own example makes the point. A simple rules-based agent authorized to initiate transactions may carry more risk than a sophisticated model limited to internal drafting assistance. That ranking would surprise anyone who assumes better benchmarks mean safer deployments. For a supervisor, authority and reversibility matter more than eloquence.
Banks do not need a separate AI risk category or a separate AI risk appetite. The OCC says SI can introduce, amplify or modify risk inside existing categories such as operational, compliance and credit risk. Control intensity may rise with the potential consequences of an agent’s failure or misuse. Higher-consequence uses may justify tighter limits on autonomy, more direct human oversight, more independent validation and stronger resilience measures.
The framework includes an illustrative four-tier scale. Tier 1, “critical,” covers autonomous payment agents, real-time credit decisioning, fraud-hold release and market access. Tier 2, “high,” includes SI-assisted underwriting, agent-initiated customer communications, anti-money-laundering alert disposition, and code assistance for changes to internal code or data access. Tier 3, “moderate,” covers decision support with mandatory human review, such as summarizing documents for credit memos. Tier 4, “low,” covers internal productivity tools and general employee Q&A.
The tiers are scored on seven factors: decision consequence, reversibility, customer and financial impact, autonomy level, data sensitivity, explainability and provenance. A fully autonomous, customer-facing payment agent sits at Tier 1 “regardless of how well it performs in testing,” the framework states. That line rejects the idea that strong benchmark scores can offset weak controls.
Why are outside AI agents the hard part?
The framework’s operational resilience section addresses agents the bank does not control. Customer, counterparty, fintech and agentic-commerce systems that connect through bank channels can create risk even when the bank has no authority over them. The OCC suggests calibrating identity checks, provenance, delegation, authentication, monitoring, velocity limits, dispute handling, suspension and revocation to the action being requested.
That guidance lands after a difficult month for banks. On JPMorgan Chase’s third-quarter earnings call on 13 October, chief executive Jamie Dimon said the bank would not let arbitrary outside agents operate inside the company. “You don’t want rogue agents, and it could be thousands of them,” he said, according to CNBC. A JPMorgan spokeswoman confirmed that outside agents cannot act for retail customers.
JPMorgan does allow read-only access to some account data through connectors such as Plaid, so an agent can see a balance. It cannot send money or pay a bill. The distinction between viewing and acting has become the practical line in the industry. Axios reported on 22 October that Morgan Stanley, JPMorgan Chase, Bank of America, Wells Fargo and Citi were limiting third-party agent access. Reuters said it could not independently verify the specific incidents behind those restrictions.
The technical problem is authentication. Banks can see which credentials were used to sign in. They often cannot tell whether a human directed the session, whether a customer-authorized agent ran it, or whether an agent interpreted a vague instruction in a way the customer never approved. A password proves access, not intent or scope.
Liability is the second problem. Dimon said customers using outside agents should understand that “there’s a liability transfer.” The OCC’s third-party section asks banks to clarify contract terms for model-related errors and losses, including indemnification, caps and who bears customer harm. It also asks banks to require notice when a vendor adds AI features or changes a model behind an existing product.
Competition adds pressure. JPMorgan has piloted an internal tool called Smart Cash that predicts cash flow and moves funds to improve yield, CNBC reported. Consumer agents such as Meta’s Muse and OpenAI’s Dots have pushed banks to decide how much account control they will delegate. The OCC guidance does not settle that question. It gives examiners a checklist for judging how each bank answers it.
What does the framework require banks to do?
Strictly, nothing new. The OCC says banks are not expected to use the document to assess their own AI programs, and the framework creates no additional legal obligations. But the questions examiners are told to ask reveal what good practice looks like. Banks that ignore those questions may still face supervisory criticism, even without a formal rule.
The first expectation is a documented purpose. A bank may describe what a system is designed to do, what it is not designed to do, who may invoke it, which data and systems it can reach, what actions it can take, and when a human must review or suspend it. The framework treats activity outside the approved purpose as a control failure even if the output looks reasonable.
Human oversight must be real. The OCC distinguishes human-in-the-loop review, which approves an action before it executes, from human-on-the-loop monitoring, where a person can intervene. It warns that approval can become “a nominal click” when volumes are high. It also asks banks to consider how many agents, accounts or alerts one person can credibly monitor.
Testing needs to go beyond a single successful run. The framework lists adversarial testing, red teaming, production sampling and hallucination measurement. It also cautions that using one model to judge another requires attention to the judge’s validation, independence, position bias, verbosity bias and correlated failure modes. That is a notable warning from a regulator about a common industry shortcut.
The framework also covers documentation, audit and vendor risk. Banks may need to reconstruct inputs, outputs, tool calls, approvals and overrides for a material decision. Internal audit should be independent of both the team that built the system and the team that validates it. For vendors, the OCC asks banks to define the authority and permitted tool use of provider agents and to monitor what those agents do.
This fills a gap. The interagency model risk guidance updated in April 2026 left generative and agentic AI outside its scope, according to Ballard Spahr. The OCC framework focuses on exactly those systems. State regulators moved earlier: the Conference of State Bank Supervisors released a discretionary AI framework on 16 September 2026 for state-chartered banks and licensed nonbanks. Each state decides how much of it to adopt, and Ballard Spahr notes that state regulators supervise a substantial majority of US banks by number.
What are the biggest gaps?
The most obvious gap is enforceability. The framework is a supervisory resource, not a rule. It sets no deadline and requires no filing, and it gives customers no new route to challenge an AI-driven decision. A bank can comply with the letter of existing law and still run an agent the OCC considers poorly controlled.
The tiers are illustrative, not mandatory. The OCC tells examiners to apply judgment and to scale reviews to a bank’s size, complexity and risk profile, including at community banks. Two examiners could score the same use case differently without violating the guidance. That flexibility is deliberate, but it makes outcomes harder to predict.
Agent authorization remains unsettled. There is no common standard for how a customer grants an AI agent limited permission to act on a bank account, how a bank verifies that permission, or how a customer revokes it. Read-only connections through open-banking providers are a workaround, not a solution for payments. The OCC lists the controls banks should consider but does not define a protocol that builders can implement.
Some risks resist measurement. The framework acknowledges that generative systems are non-deterministic, that explainability tools do not reveal a model’s reasoning, and that small prompt or model changes can alter behavior between reviews. Those properties complicate the testing, monitoring and audit evidence the framework asks for.
Vendor concentration is another weak point. Many banks buy the same foundation models and cloud infrastructure, so a shared failure could hit several institutions at once. The OCC asks examiners to consider aggregate exposure and provider concentration, but the guidance offers no thresholds or caps.
The naming may also cause confusion. Calling conventional AI “super intelligence” raises the temperature of a document that is mostly about inventories, permissions, testing and logs. Readers who take the label literally may overestimate what the guidance covers. Readers who dismiss it may miss the operational detail.
What should banks do before an AI exam?
The framework works as a preparation list even though it is not binding. Start with an inventory that records each use case, its owner, provider, permissions, deployment status, monitoring and failover arrangements. The OCC says the inventory may be updated when those characteristics change, so a static spreadsheet will not stay useful for long.
Write purpose statements in testable terms. “Can summarize documents” is vague. “Can read approved credit files, cannot send email, cannot change limits, escalates exceptions to a named reviewer” is testable. The framework prefers the second style because it gives examiners something concrete to check against logs and permissions.
Map each use case to a tier and document the reasoning. A payment agent that can move funds without review belongs at the top, even if its benchmark scores are strong. A drafting tool that never touches customer data belongs lower. Recording why a use case landed in a tier helps an examiner follow the decision and challenge it.
Test the controls, not just the model. The OCC lists circuit breakers, rate limits, least-privilege permissions and kill switches as containment measures for an agent that remains available while acting incorrectly. Banks should be able to show those controls working, not merely describe them in policy.
Prepare for vendor questions. Contracts may need notice periods for model changes, clarity on data use, and allocation of liability for errors. The OCC’s third-party section treats those as part of the risk review, not as procurement details to settle after launch.
Keep evidence. Logs of prompts, tool calls, approvals and overrides are what allow a bank to reconstruct a decision later. Without them, a bank may struggle to answer an examiner’s most basic question: what did the system do, and who approved it?
What happens next?
The OCC says it intends to update the framework regularly. The agency’s Spring 2026 Semiannual Risk Perspective said the federal banking agencies planned to issue a request for information on model risk management and AI, including generative and agentic systems, according to Ballard Spahr. That request has not yet produced a rule.
For banks, the near-term test is whether examiners begin asking for AI inventories, purpose statements, tiering worksheets and agent permission logs during routine reviews. The framework gives examiners a common vocabulary for those requests. That may matter more than its nonbinding status.
For builders, the document is a design checklist. Scoped permissions, kill switches, rate limits, audit trails, vendor change notices and human escalation paths all appear in the text. Products that cannot demonstrate those controls will be harder for banks to adopt, especially where agents can move money.
For customers, the practical question is how an assistant connects to a bank account. A read-only open-banking link is different from handing over a username and password. The OCC framework does not regulate consumer agents directly, but its emphasis on delegation and revocation signals where supervisory attention is heading.
The larger question is whether consequence-based supervision works for fast-moving technology. The OCC has chosen to judge agents by what they can do rather than how clever they are. That is a defensible principle. Whether it produces consistent exams, safer agent designs or clearer recourse for customers will depend on how examiners apply it, and on whether other regulators adopt the same approach.