On Nov. 4, 2026, the Office of the Comptroller of the Currency issued a supervisory framework to help examiners evaluate risks arising from banks’ use of “Super Intelligence” – referring to technologies and systems previously encompassed by the term “artificial intelligence,” as well as the promise, potential, and rapidly advancing capabilities of those technologies. The focus is on generative AI and autonomous or “agentic” systems, specifically during normal-course activities, such as credit-decisioning, fraud detection, customer service, operations, risk management, and market activities.1 The framework signals that examiners may give closer attention to higher-consequence uses of these technologies, such as payment agents, real-time credit decisioning, and systems that release fraud holds. Examiners may also assess whether banks have controls commensurate with the risks associated with those uses.
The framework does not establish new requirements, create AI as a separate risk category, or require banks to adopt a separate AI risk appetite. In fact, the OCC explicitly instructs examiners that banks are not expected to use this framework for internal self-assessments, reinforcing that it serves as an examiner resource rather than as a mandatory compliance checklist. In that regard, the framework describes how AI may introduce, amplify, or modify existing bank risks and offers illustrative considerations for risk-based supervision. As such, it notes that examiners should exercise judgment based on the bank’s size, complexity, risk profile, and particular use cases. That said, regulated entities should consider paying close attention to its guidance in developing their own AI compliance programs.
Risk Depends on Consequences and Autonomy
The framework indicates that control intensity may increase with the potential consequences of an AI system’s failure or misuse. In assessing a use case, examiners may consider factors such as:
- The consequence and reversibility of a decision or action;
- The potential customer, financial, compliance, or operational impact;
- The system’s autonomy and the scope of its permissions;
- The sensitivity of the data involved;
- The ability to explain or trace outputs; and
- The system’s provenance, including reliance on in-house vs. third-party models or services.
In other words, technical sophistication alone does not determine risk. The OCC gives the example that a simple rules-based agent authorized to initiate transactions might pose greater risk than a sophisticated model limited to internal drafting assistance. The framework sets out four illustrative tiers, ranging from low-risk tools that are fully human-controlled with no sensitive data exposure to critical uses that are partially or fully autonomous with high consumer or financial impact. It does not prescribe a mandatory scoring formula, nor does it assign fixed controls to each tier. It indicates that higher-risk uses may warrant tighter limits on autonomy, more substantive human oversight, more rigorous testing, and stronger resilience measures.
Six Supervisory Pillars
The framework organizes its 16 components — specific risk and control evaluation topics — into six pillars:
- Purpose and Permitted Use: Addresses the system’s authorized uses and users, data and system access, boundaries and escalations, and flagging change and scope drift.
- Risk Classification: Assesses the risks with each use case, including the consequences, explainability, and reversibility of each output or action, particularly on customer, financial, compliance, or operational functions.
- Control: Considers safeguards, including human oversight, model risk management, data governance, privacy, compliance, and cybersecurity.
- Oversight: Looks to who monitors the system and who is accountable for its use, including to set the risk appetite, establish lines of accountability, and confirm appropriate reporting to leadership and the board
- Resilience: Considers whether the bank can detect, contain, recover from, and remediate failures while maintaining critical processes. The framework notes needed calibration for external agent interactions with customers, counterparties, fintechs, or agentic-commerce systems that a bank does not control. In addition, the framework flags broad permissions for agents as a cybersecurity risk.
- Assurance: Looks to evidence that controls remain effective, including through adversarial, subjective, red team, and LLM-as-a-judge testing, independent challenge, audit, documentation, and vendor-risk management. The framework notes that some consequential, rapidly changing, or highly autonomous use cases may warrant more frequent or substantive audit or independent challenge coverage.
The framework emphasizes that human oversight should be substantive. It also highlights both the importance of both human-in-the-loop (i.e., pre-execution oversight) and human-on-the loop (i.e., active monitoring with authority and practical ability to intervene) oversight. A reviewer may need the expertise, information, capacity, and authority to reject, modify, pause, or halt an action. High-volume approvals or monitoring arrangements that leave too little time to intervene may undermine the effectiveness of oversight. The OCC’s framework also highlights risks that may warrant specific attention in AI deployments, including hallucinations in customer-facing communications, prompt-injection and other cyberattacks, model or data changes made by vendors, use of sensitive information in prompts and logs, and systems that remain operational while taking inappropriate actions. For agentic systems, the framework points to controls such as least-privilege permissions, rate limits, circuit breakers, kill switches, tested fallback processes, and customer-remediation procedures. The OCC includes a glossary of technical terms (such as agentic systems, hallucinations, prompt injection, and guardrails) at the end of the framework to help align bank examiners, compliance officers, and software engineers on terminology during an inspection.
Existing Obligations and Risk Frameworks Remain Relevant
The OCC states that AI does not change banks’ underlying legal obligations. The framework identifies, among other areas, potential compliance considerations for transaction monitoring and sanctions screening; consumer-facing communications and servicing; and representations about fees, rates, account status, or product terms. The framework also states that existing model risk management principles described in OCC Bulletin 2026-13 apply to non-generative and non-agentic AI models. The new framework is aimed more squarely at generative and agentic systems, whose changing behavior, non-determinism, and varied outputs may call for adapted testing, monitoring, and validation approaches.
Relationships With Third Parties
Where a third party provides or hosts AI capabilities, the OCC’s framework highlights the need to consider provider access, model and service changes, incident notification, data use, contractual responsibility, and concentration risk. The framework states that outsourcing development or hosting does not generally shift the bank’s accountability for managing the risks of its use. In fact, this scrutiny extends to “nested” or fourth-party AI arrangements. Many software vendors now build their products on top of underlying AI engines supplied by other technology companies. As a result, a bank might unknowingly rely on a fourth-party AI provider with whom it has no direct contract, visibility, or control. Even when an AI tool is embedded inside a vendor’s product, the bank remains fully accountable to regulators if that system malfunctions, leaks data, or harms customers.
Practical Considerations for Banks
Although the framework is supervisory guidance rather than a new set of binding requirements, OCC-supervised banks should consider whether their existing controls address its themes. Such institutions may wish to:
- Maintain an inventory of AI use cases, owners, providers, permissions, deployment status, monitoring, and continuity arrangements.
- Consider overall leadership structure for setting risk appetite, identifying accountable executives and use case owners, and establishing reporting for incidents.
- Document purpose and boundaries in testable terms, including what the system may access or do, what requires approval, and when it must abstain or escalate.
- Assess risk by use case, including consequence, reversibility, autonomy, data sensitivity, and potential exposure across shared providers or infrastructure.
- Make human review meaningful by setting clear review triggers and verifying that reviewers have adequate authority, information, and capacity to act.
- Test and monitor throughout the lifecycle, including for adversarial inputs, hallucinations, drift, upstream model changes, and control effectiveness.
- Review vendor arrangements for transparency, change notice, data use, incident reporting, audit or access rights, allocation of responsibility, and exit options.
- Plan for incorrect operation, not just outages, with tested containment, fallback, recovery, and customer-remediation procedures.
- Preserve evidence sufficient to reconstruct material outputs and actions, including relevant inputs, sources, tool calls, approvals, overrides, and system changes.
The OCC describes the framework as a “living” document that may change as the technology, use cases, and supervisory understanding evolve. Banks may therefore wish to treat it as a current indication of supervisory focus, not a static checklist. State-regulated institutions should also consider guidance applicable from their own regulators, as the Conference of State Bank Supervisors issued a separate AI examination framework in September.
1 The OCC Bulletin is formally titled “Super Intelligence: Supervisory Framework,” and it adopts the federal statutory term “Super Intelligence” (SI) rather than “AI” throughout its text.