Implementation Guide | Financial Services

    Building a Single AI Evidence Package for Regulators Under the 2026 Interagency MRM Guidance

    A single AI evidence package is one structure, applied identically across every team, that shows the scope of AI use, the policy that applied, what was allowed, redacted, blocked or escalated, and durable records of interactions and tool calls. As of November 2026, updated interagency model risk management guidance (OCC Bulletin 2026-13) rescinds prior MRM guidance, so banks should reorganize packages around it. The most reliable way to keep the package consistent is to generate evidence from policy enforcement at runtime rather than assemble it afterward from each team's records.

    • Scope of AI use: an inventory of models, assistants, agents and tools by team, including third-party and vendor AI.
    • Applicable policy: the rules in force, with version and effective period.
    • Enforcement outcomes: for each interaction and tool call, whether it was allowed, redacted, blocked or escalated.
    • Interaction and tool-call records: durable underlying records, not summaries.
    • Change records: tool and permission changes linked to testing and approval evidence.
    • Validation, monitoring, incidents, overrides and remediation, with named owners.
    • Evidence structure: one schema across all teams, models, assistants and agents, mappable to the bank's MRM framework as reorganized around the 2026 guidance.
    • Enforcement location: policy decisions made in the execution path, not logs collected afterward.
    • Retention: configurable periods set by the bank and applied uniformly, with underlying records kept rather than only summaries.
    • Data residency and deployment: self-managed deployment in the bank's own cloud, with residency set by the bank.
    • Framework mapping: which of NYDFS, NIST and the EU AI Act are mapped, at what control level, and how mappings are kept current.
    • Declared gaps: how model lineage, dataset provenance and agentic workflow tracing will be evidenced separately.

    What changed, and why it affects the package

    OCC Bulletin 2026-13, dated April 17, 2026, announces updated model risk management guidance from the OCC, the Federal Reserve Board and the FDIC. It clarifies MRM principles, sets out a risk-based approach, and rescinds prior MRM guidance and other issuances. As of November 2026, a bank whose AI evidence is organized around earlier guidance should review that structure.

    The updated guidance covers three areas: model development and use, including testing; validation and monitoring, including conceptual soundness and outcomes analysis; and governance and controls, including clear policies, roles and responsibilities. It also states that practices appropriate for one banking organization may be inappropriate for another with a different risk profile or model purpose. A package should therefore be risk-based and specific to how the bank uses AI, not a generic template.

    Separately, an OCC supervisory framework released with news release 2026-97a offers illustrative considerations for risk-based supervisory judgment on AI use. It refers to evidence such as validation results, vendor contracts, monitoring, incidents, overrides, remediation and underlying records, rather than only management summaries. It also says tool and permission changes may be logged and linked to testing and approval evidence, and that consistency thresholds can be more useful than a single successful run. Treat these as indicators of what examiners may ask to see, not as a checklist with a binding status.

    Where evidence fragments across teams

    Lines of business, technology teams and shared platforms usually adopt AI separately. Each picks its own models, assistants, agents and tools, and each keeps its own documentation, logs and approval records. When an examiner or internal audit asks what AI was allowed to do and what it did, risk and compliance staff assemble artifacts that were never designed to fit together.

    The breakdowns are predictable. They are also a governance concern: the OCC framework lists diffuse ownership limiting effective oversight and challenge as a weakness.

    • Policy coverage gaps: Policy is written centrally but applied differently, or not at all, in some tools. The package cannot show that a rule applied to a given interaction.
    • Uneven retention: One team keeps records for a short period, another keeps them in a system the bank later retires. The package has holes that depend on who owned the tool.
    • No common record structure: Fields, timestamps and outcome labels differ by team, so records cannot be compared or mapped to an MRM framework without manual rework.
    • Summaries instead of underlying records: Teams produce management summaries because raw records are scattered or inconsistent, which is the opposite of what examiners may want to see.
    • Third-party blind spots: Vendor AI sits outside team documentation even though the bank remains responsible for outcomes.

    After-the-fact assembly versus evidence from runtime enforcement

    There are two ways to reach a consistent package. One collects and normalizes records after the fact. The other enforces policy in the execution path, so evidence is a by-product of enforcement under one policy plane. The second approach does not remove the need for governance, but it changes where consistency comes from.

    • Source of evidence: After the fact: each team's logs, documents and approvals. Runtime: the enforcement point that every interaction and tool call passes through.
    • Policy coverage: After the fact: inferred by comparing policy to records. Runtime: shown directly, because the policy decision is recorded with the interaction.
    • Structure: After the fact: normalized manually per request. Runtime: one schema from the start.
    • Timing: After the fact: assembled when the examiner asks. Runtime: generated continuously, which suits repeatable evidence over single samples.
    • Limits: After the fact: depends on what each team chose to keep. Runtime: covers only what passes through enforcement, so scope must be verified.

    Where runtime enforcement applies, and where it does not

    Trussed enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants. Enforcement acts on individual interactions and tool calls: allow, redact, block, rate limit, or escalate for human approval. Each decision produces audit-ready evidence, mapped to NYDFS, NIST and EU AI Act requirements. Trussed has run continuously for more than 18 months in a regulated enterprise, with 100+ models under one policy plane. Insurance is the regulated financial sector where Trussed is deepest.

    The approach does not cover model lineage, dataset provenance or end-to-end tracing of agentic workflows. Those require separate evidence, and the updated guidance's emphasis on development, testing and validation means the bank still needs its own validation records and documentation. Runtime records show what was allowed and what happened at the point of use. They do not explain how a model was built or trained.

    For EU AI Act purposes, the consolidated text requires a documented, maintained risk management system and effective human oversight for high-risk AI systems. Escalation for human approval and recorded outcomes can support that evidence, but whether a given system is in scope is a legal determination for the bank.

    Operating model and next steps

    Start with the inventory. List every model, assistant, agent and tool by team, including vendor AI, and assign an accountable owner to each. Then agree the common schema with model risk management, compliance and internal audit, so independent challenge is built in. A cross-functional committee spanning legal, compliance, risk and data science is one model; NYDFS guidance for insurers recommends it, though it is not a bank requirement.

    Next, choose where enforcement will sit and run a bounded pilot. Test whether a sample examiner request can be answered from underlying records without manual reassembly, and whether tool and permission changes link to approval records. Finally, document the evidence you will still collect outside runtime enforcement. A package that states its boundaries is easier to defend than one that implies completeness.

    One evidence package, one structure

    Scope

    Models, assistants, agents and tools in use, by team

    Policy

    The rules that applied to each interaction and tool call

    Outcome

    Allowed, redacted, blocked or escalated

    Record

    Durable interaction and tool-call evidence with set retention

    Review your AI evidence structure

    Walk through how policy enforcement in the execution path generates audit-ready evidence, and where separate evidence is still required.

    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.