Compliance Guide for Financial Services

    Shadow AI Evidence for Bank Examiners: What to Show Under Current Model Risk Guidance

    An examiner-ready shadow AI evidence set pairs evidence of intent (a firm-wide inventory with a named owner, policy, training attestations) with evidence of operation (per-interaction decision records for allow, redact, block and escalate actions, exception approvals, and retention settings). As of November 2026, OCC model risk management guidance is risk-based and non-enforceable, so the depth of evidence should scale with the materiality of each use. The categories below are practitioner framing, not regulatory requirements.

    Criteria for evaluating whether controls produce durable evidence

      What changed in the model risk baseline

      Many shadow AI evidence guides still cite SR 11-7 and OCC Bulletin 2011-12 as the baseline. That is out of date for OCC-supervised banks. On April 17, 2026, the OCC, in coordination with the Federal Reserve Board and FDIC, issued revised model risk management guidance as OCC Bulletin 2026-13. It rescinds OCC Bulletin 2011-12 and OCC Bulletin 2021-19.

      The revised guidance is risk-based and tailored to a bank's size, complexity and extent of model use. It covers model development and use, validation and monitoring, and governance and controls. It also states that it sets no enforceable standards or prescriptive requirements, and that non-compliance will not result in supervisory criticism. The agencies said at release that they planned to seek input on model risk management, with particular attention to banks' use of AI, including generative and agentic AI.

      On November 4, 2026, the OCC issued Bulletin 2026-54, a supervisory framework for examiners evaluating banks' use of what it calls super intelligence. It establishes no supervisory expectations for banks. It tells examiners to tailor work to the materiality of the use case, says the subject should not be treated as a separate risk category, and positions itself as a supplement to existing model risk, governance, technology, compliance and operational-risk practices. It is not a shadow AI standard, but the direction is useful: fold shadow AI evidence into existing processes and scale depth to materiality.

      Evidence of intent versus evidence of operation

      Evidence of intent shows what the bank meant to happen. Policies, AI inventories and training attestations belong here. They matter, and the 2011 SR 11-7 text is a useful historical analogy: it called for a comprehensive inventory of models, a specific party maintaining a firm-wide inventory, defined authority to restrict model use, processes to ensure policy compliance, and internal audit assessment of the framework. Those are intent artifacts.

      Evidence of operation shows what happened at the point of use. For shadow AI, that means a record that a specific interaction or tool call occurred, which policy applied, and whether it was allowed, redacted, blocked, rate limited or escalated. A policy prohibiting unapproved AI tools does not show that any prohibition was applied to a given prompt, and an inventory does not show what a tool did with bank data.

      The distinction also affects how records are produced. Records written by the enforcement point at the time of the action differ from records reconstructed later from network, proxy or SaaS logs. Reconstruction can show that traffic existed. It usually cannot show which policy applied or what the decision was. This distinction is editorial framing from the author, not regulatory text, but it is the gap examiners are most likely to find when a bank has only documents and after-the-fact reviews.

      Question one: what AI is in use?

      Start with a firm-wide inventory that has a named owner, even if business lines keep their own. For approved tools this is a registration exercise. For unsanctioned tools, completeness depends on detection, so the inventory should record how each entry was discovered, not only that it exists.

      A workable category list for a bank environment includes browser-based AI sites, coding assistants, AI features embedded in SaaS products, direct calls to external model APIs, and MCP servers and agent tools. This is a practical framing rather than a formal taxonomy. The evidence available differs by category, and coverage of one does not imply coverage of another. Browser traffic visibility says little about API calls from a build pipeline or tool calls made by an agent.

      For each category, record whether the bank can show the tool exists, who uses it, what data it touched, and what policy applied. Where the answer to the last two is none, mark the gap explicitly. A candid gap register with owners and dates is stronger evidence of a functioning process than an inventory that implies full coverage.

      Questions two to five: policy, actions, exceptions and retention

      What was it allowed to do? The evidence is the policy version in force at the time, tied to the interaction. A current policy PDF does not show which rule applied last quarter, so policy versions need dates and change history.

      What did it do? The evidence is a decision record per interaction or tool call: the actor, the tool or model, the policy applied, and the outcome. Outcomes such as allow, redact, block, rate limit and escalate each need a record, because blocked and redacted events show that the control operated, and allowed events show what was permitted.

      Who approved exceptions? Where a sanctioned exception exists, or an action is escalated for human approval, the record should identify the approver, the policy basis, the scope, the expiry, and the specific interactions it covered. The sources do not prescribe this content, so treat it as proposed good practice. An exception without an expiry or scope is a standing bypass.

      How long is the evidence retained? Retention should be a configured setting with a known owner, not a by-product of storage defaults. Know where records reside, who can change retention or alter records, and how integrity is protected. Third-party risk and data protection reviews will ask the same questions about wherever the evidence is stored.

      Where runtime enforcement fits

      Closing the gap between intent and operation requires a control that sits where the AI use happens. Trussed is one way to do this. It enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants, with allow, redact, block, rate limit and escalate for human approval, at the level of individual interactions and tool calls. The audit-ready evidence is generated by enforcement at runtime, not reviewed after the fact. Deployment is self-managed in the customer cloud, with configurable retention and data residency.

      Trussed does not claim financial-services customers. Insurance is where it is deepest among regulated financial sectors, and its production record comes from regulated enterprises in other verticals, including 18+ months of continuous production in one regulated enterprise. Its scope is interaction and tool-call enforcement; it does not provide transaction-level governance, model lineage or end-to-end tracing of agentic workflows, and a bank's inventory and model validation processes remain separate.

      A practical next step is to take the five questions above, answer each for every shadow AI category you have identified, and mark each answer as intent evidence, operation evidence, or missing. The missing operation evidence is the work list.

      Five examiner questions, five evidence types

      What AI is in use?

      Inventory with a named owner, plus detection of unsanctioned tools.

      What was it allowed to do?

      The policy that applied to each interaction or tool call.

      What did it do?

      Decision records for allow, redact, block, rate limit and escalate.

      Who approved exceptions?

      Approver, scope, expiry and linkage to specific interactions.

      How long is it kept?

      Retention settings, residency and who can change them.

      Test your evidence against the examiner's questions

      Walk through your shadow AI categories and see which questions you can answer with records of what happened, not only documents.

      Talk to an Expert