What the bulletin states
OCC Bulletin 2026-13, Model Risk Management: Revised Guidance, is dated April 17, 2026. The OCC issued it in coordination with the Federal Reserve Board and the FDIC as updated interagency guidance. Its stated purposes are to clarify model risk management principles, set out a risk-based approach, and rescind prior guidance. The rescinded items include OCC Bulletins 1997-24 and 2011-12, along with OCC Bulletin 2021-19 and the Comptroller's Handbook booklet on model risk management, as listed in the OCC news release.
Two statements matter for GenAI. The guidance says that "Generative AI and agentic AI models are novel and rapidly evolving." It also says its principles "apply to traditional statistical and quantitative models and non-generative, non-agentic AI models." The page title calls this the GenAI exclusion, but the operative text is a statement about which model types the principles cover. As of October 2026, that is the textual basis for the exclusion, and the rest of this guide separates it from what is inferred.
The guidance also says each banking organization is ultimately responsible for adopting model risk management practices that are appropriate and effective for the risks it faces. Practices suitable for one bank may be inappropriate for another with a different risk profile or model purpose.
What the exclusion does and does not change
The exclusion changes the reference point. A bank cannot point to the bulletin's validation, monitoring and governance principles as a ready-made standard for GenAI. It also cannot treat the document as a safe harbor, because a statement that principles apply to certain model types is not a statement that other systems are unsupervised.
The following is inference, not text from the bulletin. Banks that deploy GenAI still operate under existing expectations for third-party risk management when a vendor supplies the model or tooling, for information security when systems touch sensitive data or connect to internal tools, and for consumer protection when outputs affect customers. The OCC's Spring 2026 Semiannual Risk Perspective describes banks as taking a measured approach, with GenAI and agentic AI limited to specific use cases and human-in-the-loop accountability. That description of observed practice suggests examiners may ask about approval steps and accountability. It does not say they will require them.
The practical reading is that the bank sets its own GenAI governance position and documents it. Applying model risk management principles by analogy is a reasonable choice, but it is the bank's judgment, and the record should say so. Questions at the boundary, such as whether an agent that calls tools falls inside either category, turn on the bulletin's exact wording and belong in that documented position.
Where the evidence gap appears
Traditional model validation assumes a bounded artifact: a model with defined inputs, a development record, and performance that can be tested and monitored over time. GenAI use often looks different. The risk sits in individual interactions, in what a prompt contained, what a response returned, and which tools or data sources an agent reached. Validation workflows built for traditional models may not capture that level. This is an inference from the structure of the guidance, not a statement in it.
That creates a specific evidence problem. Policies and approval memos show what a bank intended. An examiner asking about a GenAI system is likely to want to see what the system was permitted to do and what it did. Answering requires a record that links a policy decision to each interaction or tool call. Documentation and after-the-fact review can describe a control, but they cannot show that the control applied to a particular action.
The gap is also scoping. Banks need a map of the full AI footprint: models, agents, the tools they can invoke, and the servers those tools connect to. The footprint has no standard definition, so the bank defines it, and unknown use cannot be controlled or evidenced.
Where runtime policy enforcement fits
Runtime policy enforcement addresses one part of the gap: control and evidence at the level of the individual interaction and tool call. It does not replace the governance decisions above, such as scoping, ownership, and the bank's position on which expectations apply. It gives those decisions an enforcement point and a record.
Trussed is the control, provenance and evidence layer for consequential AI. It enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants. For each interaction or tool call, the action can be allowed, redacted, blocked, rate limited, or escalated for human approval. Policy is enforced at runtime, not reviewed after the fact. That enforcement generates audit-ready evidence continuously, which ties each action to the policy decision that governed it. Deployment is self-managed in the customer cloud, with configurable retention and data residency.
Trussed does not claim bank deployments. Insurance, also a regulated financial sector, is where it is deepest, and the approved general proof applies across regulated enterprise use: more than 18 months of continuous production in a regulated enterprise, 99.95% uptime on live production traffic, more than 100 models under one policy plane, and more than 30 GB of audit evidence generated per day. Those figures describe operating experience, not how a bank's examiner will respond to any particular control.
A bank evaluating any runtime layer should ask the questions the checklist implies. Can it show, per interaction or tool call, which policy applied and what happened? Where in the execution path does enforcement occur? How is the evidence retained, and which claims are specific to banking rather than general?