A new playbook for bank examiners
The U.S. Office of the Comptroller of the Currency has issued its first dedicated supervisory framework for how its bank examiners should evaluate banks’ use of artificial intelligence, labeled in the document as “Super Intelligence” or “SI” . The move is not a ban, a licensing regime or a new standalone AI rulebook. It is more operational than that: a guide for the people who inspect national banks, federal savings associations and federal branches or agencies of foreign banking organizations .
That distinction matters. For several years, financial firms have heard general warnings that AI can amplify model, cyber, operational, compliance and consumer-protection risks. The OCC’s November 4, 2026 action shifts the conversation from broad caution to examination practice. Reuters reported that the regulator described the document as its first-ever guidance for examiners probing banks’ AI use, and that the core message is to monitor AI without treating it as a risk category entirely apart from the rest of banking .
The OCC’s own bulletin says the framework is intended to support examiners evaluating banks’ SI use in normal-course business activities, while making clear that it does not create new supervisory expectations, enforceable standards or prescriptive requirements for banks . In regulatory reality, however, an examiner guide can still matter deeply. Banks may not be legally required to adopt every line of a nonbinding document, but when examination teams use a common framework, governance committees, model-risk teams and vendors tend to align their evidence accordingly.
Risk follows consequences, not buzzwords
The most important idea in the new framework is proportionality. Examiners are told to tailor their review to the materiality of the use case, the bank’s size and complexity, and the consequences, autonomy and complexity of the SI system involved . That means a glamorous large language model used only for internal drafting may receive less supervisory concern than a simpler automated agent with authority to move money, alter a customer record or interact with markets.
The OCC’s release states that examiners should not treat SI as a separate and distinct risk category and that banks should not be required to define a separate SI risk appetite . Instead, AI can introduce, amplify or modify risks already familiar to supervisors: operational risk, compliance risk, credit risk, privacy risk, third-party risk and cybersecurity risk . This is a technology-neutral posture, but not a hands-off one.
The framework’s executive summary gives concrete examples of bank AI use cases: credit decisioning, payments, fraud detection, customer service, operations, risk management and market activities . It also emphasizes that SI and agentic SI can improve speed, consistency, cyber defense, fraud detection, financial-crime monitoring and operational efficiency when used with adequate controls . The message is not “do not use AI.” It is closer to “show the examiner what the system may do, what it may not do, who is accountable, and what happens when it fails.”
That is a meaningful shift for procurement and product design. A bank considering an AI vendor can no longer rely only on accuracy metrics, a polished model card or a cybersecurity attestation. It will need to explain permitted use, data access, tool permissions, escalation points, testing results, monitoring thresholds, incident handling and auditability. Compliance has entered the chat with administrator rights.
Six pillars, sixteen components
The OCC organizes the framework around six pillars: purpose, risk, control, oversight, resilience and assurance . Under those pillars sit sixteen components covering defined purpose, risk classification, human oversight, governance, model-risk extensions, data governance, privacy, compliance, third-party risk, operational resilience, cybersecurity, performance monitoring, testing, documentation, internal audit and reporting .
The first component is deceptively simple: define the purpose and permitted use of the system. The framework says a bank may document what an SI system or agent is designed to do, what it is not designed to do, who or what may invoke it, which data and systems it may access, what actions it can take, and when human review or suspension is required . In other words, “we use AI for customer support” is not enough. The exam question becomes: support for which products, using which data, with authority to say what, under what controls?
The second component is risk classification. The OCC points to factors such as decision consequence, reversibility, customer and financial impact, autonomy, data sensitivity, explainability, provenance and interdependencies . This is especially important for agentic systems. The risk is not simply that an AI answer is wrong. It is that a wrong answer can become an action, and that the action may be difficult to reverse.
Human oversight receives similar treatment. The framework distinguishes human-in-the-loop controls, where a person reviews or approves before action, from human-on-the-loop controls, where a person actively monitors and can intervene . The OCC cautions that oversight must be meaningful: reviewers need expertise, authority, information and enough capacity to actually challenge or stop the system . A nominal approval click in a high-volume queue will not look like robust governance if the human cannot understand or halt the agent’s behavior.
The agentic risk problem
The most forward-looking parts of the document concern agentic AI. The OCC defines agentic SI as systems capable of independently making decisions, interacting with their environment and optimizing processes without direct human intervention . In banking, that is not a theoretical issue. The more an AI system can retrieve data, call tools, initiate workflows, update records, draft communications or trigger transactions, the more it resembles an operational actor rather than an analytical tool.
The framework’s illustrative risk-tiering section makes the point plainly. A low-risk internal productivity tool with no customer or financial data exposure can sit in a lower tier, while a fully autonomous customer-facing payment agent can fall into the highest tier regardless of how well it performs in testing . That formulation will shape vendor conversations. A bank buying an “agent” will have to ask not only whether the model is accurate, but whether its permissions are narrow, observable, revocable and tested under failure conditions.
Operational resilience is also reframed for AI. The OCC notes that an agentic system could remain available while acting incorrectly, requiring banks to detect, contain, recover from and remediate harmful actions while continuing the underlying business process . Traditional uptime metrics do not solve that problem. A system can be “up” and still be causing damage at machine speed.
The framework points to tools such as monitoring, observability, deterministic rules, circuit breakers, rate limits, least-privilege permissions and kill switches to contain misfires . For banks, this means AI resilience is not only about disaster recovery. It is about proving that an autonomous or semi-autonomous process can be paused, rolled back, investigated and corrected before customers or markets absorb the full harm.
Vendors face clearer but tougher diligence
For technology vendors selling models, copilots or agents into banking, the OCC framework is a double-edged development. It may lengthen procurement because banks will need more evidence about provenance, data rights, model changes, prompt management, logging, testing, incident notification and liability. The framework specifically notes that banks often obtain SI capabilities from third parties and that such reliance can introduce model, operational, data, legal, concentration and resilience risks beyond conventional service-level and information-security reviews .
But standardization can also help vendors. A repeatable supervisory vocabulary gives the market a clearer checklist. Providers that can demonstrate permission boundaries, monitoring hooks, version transparency, human-override workflows, audit trails and contractual clarity will be easier for banks to defend during examinations.
The vendor section is particularly direct about change management. Existing vendors may add SI features or change models connected to products already in use, and banks may require notice, transparency, reassessment or retesting when those changes occur . That is a warning against “silent AI upgrades.” In a regulated bank stack, a feature update can become a supervisory event if it changes what a system can do or how it behaves.
Why this milestone matters
The OCC says it intends to update the framework regularly as SI evolves and as stakeholder feedback develops . That makes the guide a living supervisory document rather than a final settlement of AI risk. It also means banks should expect examination questions to become more specific over time.
The bigger significance is institutional. AI oversight in banking is moving from principle to routine. Examiners now have a common map for asking how a system is used, what it can affect, who owns it, how it is tested, how it is monitored, which vendors support it, how incidents are escalated and how decisions can be reconstructed.
For banks, the practical next step is not to create a separate AI empire. The OCC has explicitly avoided that. The better response is to connect AI inventories to existing risk, compliance, technology, audit and third-party governance. The new exam question is not whether a model is impressive. It is whether the bank can prove that the model, the agent and the humans around it are controlled well enough for the consequences they can create.
Comments
Be the first to comment.