The most defining features of the Framework are as follows:

• It is an examiner tool, not a rule. The OCC states that the Framework does not establish supervisory expectations for banks, does not set enforceable standards or prescriptive requirements, and is not intended to be used by banks to assess their own AI risk management.
• AI is not a separate risk category. Examiners should not treat AI as a distinct risk category, and banks are not required to define a separate AI risk appetite. AI instead may introduce, amplify, or modify risk in existing categories such as operational, compliance, and credit risk.
• Consequences, not sophistication, drive scrutiny. Control intensity should scale with the potential consequences of an AI system’s failure or misuse, the transfer of agency from people to the system, and the degree of human oversight, rather than the technical sophistication of the underlying model.

Although addressed to examiners, the Framework is the clearest available statement of what OCC examination teams will look for when reviewing generative and agentic AI. Institutions should expect it to shape examiner requests, scoping, and conversations with management.

Background

Banks have used machine learning for years in fraud detection, anti-money laundering monitoring, and credit underwriting. The arrival of generative AI, which produces text, code, and other content, and agentic AI, which can make decisions and take actions with limited human involvement, has sharply accelerated adoption and tested the limits of the supervisory tools regulators have traditionally relied on.

For fifteen years, the principal supervisory tool was the interagency model risk management guidance issued in 2011 as OCC Bulletin 2011-12 and Federal Reserve SR 11-7. That guidance, supplemented by “Model Risk Management” booklet in the Comptroller’s Handbook, became the de facto framework for overseeing AI and machine learning. On April 17, 2026, the OCC, the Federal Reserve, and the FDIC replaced the model risk management guidance with revised interagency guidance, issued by the OCC as Bulletin 2026-13 (and as SR 26-2 and FIL-15-2026).

OCC Bulletin 2026-13 is shorter and expressly risk-based. It narrows the definition of “model” to complex quantitative methods grounded in statistical, economic, or financial theory, excluding spreadsheet arithmetic and deterministic rules-based software. It ties the rigor of development, validation, and monitoring to a model’s materiality, and states that it is expected to be most relevant to institutions with more than $30 billion in assets. The Bulletin also states that it creates no enforceable standards and that non-compliance will not result in supervisory criticism, although unsafe practices or violations of law arising from poor model risk management can still lead to supervisory action.

Most significantly, Bulletin 2026-13 placed generative and agentic AI outside its scope, describing these technologies as “novel and rapidly evolving.” That left a visible gap, which the new Framework is designed to fill. Bulletin 2026-13 governs traditional statistical models and non-generative AI, while the Framework is “more applicable to generative and agentic SI.” The Framework also draws on other recent developments, including the 2023 interagency third-party risk guidance (for which revisions were proposed on September 15, 2026), the AI Lexicon and Financial Services AI Risk Management Framework that Treasury released in February 2026, and Executive Order 14434 (October 2, 2026), which introduces the “super intelligence” terminology.

Key Provisions of the Framework

The Framework is built around sixteen components, which the OCC groups into six “pillars”: Purpose, Risk, Control, Oversight, Resilience, and Assurance. We discuss each of the six pillars below, summarizing the components of each and what they are likely to mean in practice. The Framework also includes an illustrative risk-tiering method and a glossary of terms, and the OCC describes it as a “living document” that will change as the technology and supervisory understanding mature.

Purpose: What Is the System Allowed to Do?

The Framework starts from a simple premise: a bank should be able to say, in concrete and testable terms, the purpose of each AI tool. In practice, that means documenting what a system is designed to do and not do, who may use it, which data and systems it can reach, and when a person must step in or the tool must be paused. The OCC suggests that a tool operating outside its approved purpose may be treated as a control failure even if the outcome looks reasonable. A tool approved for internal research, for example, may need fresh review before it touches customers.

Risk: What Happens If AI Gets Something Wrong?

The Framework asks examiners to focus on consequences. Relevant factors include whether a tool merely informs a human decision or acts on its own, whether a mistake can be undone, the potential harm to customers or the bank, how much autonomy the tool has, and the sensitivity of its data. The OCC illustrates the point by noting that a simple rules-based agent that can initiate transactions may pose more risk than a sophisticated model that only helps draft internal documents. An illustrative four-tier scale runs from Tier 1 (Critical), such as autonomous payment agents and real-time credit decisioning, to Tier 4 (Low), such as general productivity tools.

Control: What Keeps the System Within Bounds?

The Control pillar sets forth the safeguards to prevent unauthorized or unsafe outcomes:

• Meaningful human oversight. The Framework distinguishes a person who approves each output before it is used (“human-in-the-loop”) from one who monitors an operating system and can intervene (“human-on-the-loop”). Regardless of the classification, the reviewer must have the expertise, authority, and time to perform the task. The OCC warns that high volumes and bulk approvals can turn review into a formality.
• Adapted model risk management practices. Familiar model risk management principles can still be applied, but they must flex for systems that change frequently, give different answers to the same question, and resist simple explanation. A single successful test run, for instance, proves less for a generative AI system than repeated sampling.
• Data governance and privacy. Banks should understand where training and operating data come from and whether they are representative. Privacy concerns extend beyond conventional breaches to a model memorizing or inferring sensitive information, prompts and outputs becoming stores of customer data, and vendors using bank data to improve products sold to others.
• Compliance and consumer protection. AI has the potential to make a bank’s legal obligations harder to meet. The Framework flags BSA/AML and sanctions screening, consumer-facing uses such as servicing and collections, and the risk that a chatbot “hallucinates” an inaccurate fee, rate, or account status. Explanations must also be sufficient to satisfy legal requirements such as adverse action notices.
• Cybersecurity. AI introduces new avenues of attack, including manipulative inputs (“prompt injection”), corrupted training data, and agents with overly broad credentials. The OCC suggests treating AI agents’ identities and permissions with the same care as privileged human users.

Oversight: Who Is Watching and Who Is Accountable?

The Framework expects clear lines of accountability including named executives and use-case owners, a complete inventory of AI uses, and defined approval points for deployment, material changes, and retirement. As in other third-party arrangements, accountability stays with the bank even when a vendor builds or hosts the tool. Board and management reporting should cover material use cases, incidents, customer harm, and remediation in a form that allows timely challenge, and ongoing performance monitoring should detect when a system’s behavior drifts away from what was tested.

Resilience: Can the Bank Contain and Recover from Failures?

A distinctive insight of the Framework is that an AI system can remain fully operational while doing the wrong thing. Resilience planning therefore must address not just outages but misbehavior. The OCC points to practical brakes such as transaction limits, narrowly scoped permissions, and emergency shutdown (“kill switch”) capabilities; tested fallback processes so the business can keep running; and a dedicated workstream to identify and make whole any customers affected by an AI error. It also flags risk from outside AI agents, such as customer or fintech bots, interacting with bank channels.

Assurance: Is There Evidence the Controls Keep Working?

The final pillar asks banks to demonstrate, not merely assert, that their controls are effective. Testing should look for unknown failure modes, not just confirm expected results, using techniques such as adversarial testing, red teaming, expert review, and measurement of hallucination rates. Documentation should allow a third party to reconstruct a material AI decision. Internal audit should be independent of both the teams that build AI tools and those that validate them. The OCC also places vendor oversight here, highlighting four points for contracts: limits on what a provider’s AI agents may do inside bank systems, agreed definitions and notice periods for AI incidents, clear allocation of liability for errors and customer harm, and advance notice when a vendor adds AI features to products the bank is currently using.

Implications for Financial Institutions

The “not an expectation” framing has limits. Bulletin 2026-54 states that banks are not expected to use the Framework, and Bulletin 2026-13 likewise states that non-compliance with the guidance will not result in supervisory criticism. In practice, however, examiners will use the Framework to scope reviews and request documents. Institutions that cannot speak to purpose boundaries, risk tiers, human oversight design, and vendor controls for their material use cases should expect harder conversations.

Reconcile the risk appetite messaging. The Framework states that banks are not required to define a separate AI risk appetite, yet the governance component notes that boards and management “may establish an SI risk appetite.” Institutions can reasonably integrate AI into existing enterprise risk appetite statements, but they should be able to show how autonomy, permitted actions, and customer impact are reflected in such statements.

Generative and agentic AI now sit outside formal model risk management guidance. Because Bulletin 2026-13 excludes these systems, the Framework is the primary OCC reference point for them. Institutions should decide deliberately which generative and agentic AI systems remain within their model risk management programs, and which are governed through parallel controls.

Vendor contracts are a pressure point. Many institutions obtain AI through embedded vendor features rather than in-house development. The Framework’s focus on change notice, data-use limits, incident definitions, and liability allocation will put existing contracts under examiner scrutiny, particularly while the proposed third-party guidance remains pending.

Proportionality matters, especially for community banks. The OCC has instructed examiners to tailor reviews to a bank’s size, complexity, and risk profile, and to the nature, consequences, autonomy, and complexity of its AI use cases. Smaller institutions using low-risk internal tools should not expect the control intensity appropriate for an autonomous payment agent.

Other regulators are moving too. State regulators are issuing their own AI governance guidance (for example, the Connecticut Department of Banking in September 2026), and Treasury's FS AI RMF provides a common industry reference. Mapping internal programs to the Framework, the FS AI RMF, and applicable state guidance will reduce duplication.

Takeaways and Next Steps

Timing. The Framework became effective as an examiner resource on November 4, 2026. The OCC intends to update it regularly, including in response to stakeholder feedback.

• Inventory and tier. Confirm that your AI inventory captures generative and agentic use cases, including vendor-embedded features, and assign each a consequence-based risk tier.
• Document purpose boundaries. Prepare testable purpose and permitted-use statements for material use cases, with escalation and suspension triggers.
• Stress-test human oversight. Evaluate reviewer competency, authority, workload, and override logging for human-in-the-loop and human-on-the-loop controls.
• Close model risk management gaps. Determine how generative and agentic AI will be validated and monitored now that they fall outside Bulletin 2026-13.
• Revisit vendor contracts. Address AI change notice, data-use restrictions, incident definitions and timelines, audit rights, and liability allocation.
• Test resilience and remediation. Confirm kill-switch, fallback, and customer remediation procedures for higher-tier use cases.
• Brief the board. Update board and committee reporting on material AI use cases, incidents, and assurance results.
• Monitor pending items. Track the planned interagency request for information on model risk management and AI, the proposed third-party risk management guidance, and future Framework updates, and consider submitting comments.

For questions regarding the Framework or assistance in evaluating or implementing changes based on the Framework, please contact any member of the King & Spalding team.