Compliance Guide | Financial Services

Bank Examiner GenAI Evidence Checklist Under the 2026 Model Risk Guidance

Examiner scrutiny of GenAI use tends to center on eight evidence categories: inventory and use-case scoping, policy and acceptable use, access and data handling, control operation, human oversight and escalation, third-party dependencies, incident and exception records, and retention. The first two and parts of the sixth can be met with documents. Control operation, escalation records and exception records require evidence generated by controls running in production. This checklist is a readiness aid, not a regulatory requirement. Updated model risk management guidance issued on April 17, 2026 is risk-based and does not set prescriptive requirements.

    What changed and how to use this checklist

    On April 17, 2026, the OCC issued updated model risk management guidance in coordination with the Federal Reserve Board and FDIC. As of October 2026, it rescinds OCC Bulletin 2011-12 and OCC Bulletin 2021-19. Programs that cite only the earlier OCC bulletin, or refer to SR 11-7 by name, should confirm the current guidance with their primary regulator and update internal references.

    The guidance states that practices should be risk-based, tailored and commensurate with an institution's size, complexity and extent of model use. It is expected to be most relevant to banking organizations with over $30 billion in total assets, and it may also apply to smaller banks with significant model risk exposure. It covers model development and use, validation and monitoring, governance and controls, and vendor and third-party products. It does not set enforceable standards, and non-compliance will not result in supervisory criticism.

    That shapes how to read the checklist below. The eight categories are a practitioner's organization of what a risk team should be able to show for GenAI. They are not a regulator's list. For tools and systems the guidance does not directly address, the Federal Reserve's page says an institution's own risk management and governance practices should guide governance and controls. A defensible evidence set is therefore one you can tie to your own stated practices. Each item below notes whether the artifact is documentary or runtime-generated.

    Separating documentary evidence from runtime-generated evidence

    Most institutions already hold the documentary set: inventories, policies, validation reports and vendor files. These show intent and design. They do not show what a GenAI system did on a given day with a given prompt, which is the question an examiner can ask once a control is said to exist.

    Runtime-generated evidence answers that question. For a single interaction, the useful fields are the policy applied, the decision (allow, redact, block or escalate), the timestamp, the actor, any tool call made and the retention treatment of the record. These fields are a practical working set, not a prescribed list. Different examiners and institutions may ask for more.

    One inference is worth stating plainly. The guidance frames model risk through validation, monitoring, and governance and controls, and evidence of control operation most plausibly maps to monitoring and controls. That mapping is our reading, not regulator text. It also means runtime evidence supports validation and governance and does not replace either. Validation reports and governance programs remain distinct deliverables.

    The common failure pattern is reconstruction. When records are scattered across application logs, vendor consoles and ticketing systems, teams assemble the story after an examiner asks. Reconstructed evidence is slow to produce and harder to defend than records written at the moment of the decision.

    Where runtime policy enforcement can generate the missing evidence

    Trussed enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants. For each interaction or tool call it can allow, redact, block, rate limit or escalate for human approval. Because the decision is enforced at runtime, not reviewed after the fact, the evidence is produced by the same step that made the decision. Trussed generates continuous audit-ready evidence from enforcement, mapped to frameworks including NAIC, NYDFS, HIPAA, the EU AI Act and NIST.

    This applies most directly to categories 3, 4, 5 and 7. Escalation for human approval creates the oversight record in category 5, and blocked or redacted events populate category 7. Self-managed deployment in the customer cloud, with configurable retention and data residency, bears on category 8, because the institution decides where evidence is stored and for how long.

    Scope matters. Trussed's evidence covers individual interactions and tool calls. It does not provide model lineage, dataset provenance or end-to-end tracing of agentic workflows, and it does not replace model validation or governance programs.

    Trussed's production depth is in insurance, a regulated financial sector, alongside pharma, healthcare and education. Proof points include 18+ months of continuous production in a regulated enterprise, 99.95% uptime on live production traffic, 100+ models under one policy plane, and 30+ GB of audit evidence generated per day. Evaluate any approach, including this one, against your own institution's examination context.

    Running a gap assessment against the checklist

    Treat the checklist as a working table, one row per GenAI use in your inventory. The goal is to find categories where you can describe a control but cannot produce a record of it operating.

    • Confirm the current reference point: Check your size against the $30 billion relevance indicator and confirm that internal documents cite the current guidance rather than rescinded bulletins.
    • Label each artifact: For all eight categories, mark the evidence as documentary or runtime-generated, and note where it lives.
    • Test one interaction end to end: Pick a real GenAI interaction and try to produce the policy applied, decision, timestamp, actor and tool call within a working day.
    • Trace one escalation: Select a case where human approval was required and confirm you can show the approver, outcome and link to the interaction.
    • Check storage and retention: Confirm where evidence is held, who can modify it, how long it is kept and whether that meets your residency and retention requirements.
    • Split vendor from internal evidence: For each third-party model or tool, record what the vendor provides and what your own controls must generate.

    GenAI evidence by source type

    Documentary

    Inventory, policies, validation reports, vendor due diligence

    Runtime-generated

    Policy applied, decision, timestamp, actor, tool call per interaction

    Shared

    Retention and residency settings backed by stored records

    Test your GenAI evidence set against the checklist

    Review which categories you can satisfy with documents and which need evidence produced in production, and see how enforcement in the execution path generates it.

    Talk to an Expert