Implementation Guide for Bank CIOs

    Enterprise AI Operating Model for Banks: A CIO Implementation Guide Under SR 26-2

    A bank AI operating model document should read as a control specification. It should cover scope and AI inventory, roles and decision rights, policy taxonomy, a runtime enforcement layer, evidence and retention, and alignment to model risk management and the NIST AI RMF. For each requirement it should state where the control is enforced and what record proves it operated. As of November 2026, SR 26-2 replaces SR 11-7 as the interagency model risk guidance, so the document should be built against the current text.

    • Current AI inventory with owner, risk level, third-party designation, underlying model and version, and last risk assessment date
    • Approval and validation records, including the independent reviewer and any exceptions granted
    • The policy set in force at the time of each period under review, with change history
    • Execution records showing what the AI was permitted to do, what it did, and which control acted, including allow, redact, block, rate limit and escalation outcomes
    • Human approval records and results of tests of the approval backstop
    • Version and change records for models, prompts, configurations and providers, retained under the stated retention and residency terms

    Start from the current baseline: SR 26-2

    The Federal Reserve, OCC and FDIC issued Revised Guidance on Model Risk Management, which supersedes and replaces SR 11-7 and SR 21-8. As of November 2026, SR 26-2 is the current interagency model risk guidance. The agencies updated it based on supervisory experience and industry feedback since SR 11-7, to clarify principles and to emphasize a risk-based approach tailored to a bank's model risk profile and the size and complexity of its operations.

    Two points matter for the document. First, SR 26-2 states that every component of the model lifecycle benefits from governance supported by clear policies and effective controls, and that policies should define risk management expectations and a framework for assessing the magnitude of model risk. Second, its stated scope covers traditional statistical and quantitative models and non-generative, non-agentic AI models. Your document should therefore say explicitly how generative and agentic systems are treated, with the model risk function and counsel signing off on that position. The OCC's supervisory framework for bank use of Super Intelligence, which presents sixteen components with illustrative considerations, is a useful cross-reference for agents. It describes considerations for risk-based supervisory judgment rather than prescriptive requirements, so treat it as a lens for what examiners may ask.

    The sections the document should contain

    Organize the document so each section answers one question: what must the bank be able to prove? Mapping each section to a NIST AI RMF function (GOVERN, MAP, MEASURE, MANAGE) gives examiners and model risk staff a neutral cross-reference. Keep each section short and specific, and write requirements so they can be tested.

    • 1. Scope and AI inventory (MAP): Models, agents, tools, coding assistants and third-party services in use, with owner, risk level and date of last risk assessment.
    • 2. Roles and decision rights (GOVERN): Who approves, who validates or monitors, who owns exceptions, and who grants audit access.
    • 3. Policy taxonomy (GOVERN): Acceptable use, restricted uses, data handling, tool and agent permissions, human approval, coding assistant use.
    • 4. Runtime enforcement layer (MANAGE): Which policies are enforced in the execution path, and the actions available: allow, redact, block, rate limit, escalate.
    • 5. Evidence and retention (MEASURE): Evidence schema, retention, residency, access and independence of review.
    • 6. Model risk and change control (MEASURE, MANAGE): Alignment to SR 26-2, risk tiering, and version control across models, prompts, configurations and providers.

    Inventory, roles and the policy taxonomy

    For the inventory, the NIST GenAI profile offers a practical field list: acceptable use policies and exceptions, restricted uses, business or model owners, change management and monitoring plans, dependencies on other systems, in-house or third-party designation, risk level, incident response plans, human oversight roles, underlying foundation models with versions and access modes, and the date of last risk assessment. Adopting these fields makes third-party dependencies visible and gives the risk tier something concrete to attach to.

    On roles, SR 26-2 calls for clear roles and responsibilities with well-defined incentives between reporting lines such as development and validation. The OCC framework adds that effective challenge may depend on independence from both the group that develops or owns a system and the group that validates or monitors it. Name those groups in the document. Also name who grants and provides audit access to system documentation and data lineage, since a role left unassigned becomes a finding.

    The policy taxonomy should be written as statements a control can evaluate, not as principles. "Customer account data must not leave approved boundaries" can be enforced and evidenced. "AI must be used responsibly" cannot.

    Evidence and retention as operating model components

    The OCC framework describes an assurance component that provides evidence that controls continue to operate effectively, and notes that records may support reconstruction of material outputs or actions even when a model cannot give a rule-based rationale. The implication is that the document must specify what execution records are captured, not only what policies exist. Define an evidence schema with three parts: what the AI was permitted to do, what it did, and which control acted. Extend version and change control beyond model weights to cover prompts, configuration, fine-tuning, data and provider, since the OCC framework names all of these.

    Retention, residency and deployment belong in the same section. Specify retention periods against the bank's existing records management schedule, state where evidence may be stored, and say who can read it. If evidence contains sensitive content, a self-managed deployment in the bank's own cloud may simplify third-party risk review, because the evidence store stays within boundaries the bank already governs. Cross-reference the bank's third-party risk and NYDFS obligations here rather than restating them, and have the relevant owners confirm the mapping.

    Trussed is one example of tooling in this area. It enforces policy in the execution path across models, agents, tools, MCP servers and coding assistants, with actions to allow, redact, block, rate limit or escalate for human approval, at the level of individual interactions and tool calls. Enforcement generates audit-ready evidence mapped to NYDFS and NIST, and deployment is self-managed in the customer cloud with configurable retention and data residency. Trussed's deepest regulated experience is in insurance, and it does not claim bank deployments. The operating model should be written so that it holds regardless of the tooling chosen.

    From policy to provable control

    Define

    Inventory, roles and a policy taxonomy written as testable requirements.

    Enforce

    Allow, redact, block, rate limit or escalate in the execution path.

    Record

    Capture what was permitted, what happened and which control acted.

    Demonstrate

    Produce an evidence set examiners and model risk teams can reconstruct from.

    Turn the document into demonstrable control

    See how policy enforced at runtime produces audit-ready evidence for individual interactions and tool calls, and how it maps to your operating model sections.

    See How Evidence Is Produced
    Next step

    How does your AI governance program compare?

    See where your program has gaps in less than 2 minutes.