Shadow AI Exception Framework for Banks: A Temporary Approval Guide
A temporary AI exception should be a bounded policy state, not a memo. Each decision records a tier, scope, permitted data, approver and expiry. Compensating controls are enforced on every interaction and tool call, and expiry reverts the use to the default action. The framework below reflects the interagency model risk guidance issued in April 2026 and is current as of November 2026.
Specify the exception terms
- Scope: named tool, named user group or business line, and permitted use case.
- Permitted data: allowed data classes, and which classes must be redacted or blocked.
- Duration: a fixed start and end date, with a maximum term that shortens as the tier rises.
- Approvers: roles required by tier, plus the role authorized to suspend.
- Compensating controls: the specific actions applied at runtime, such as redact, rate limit or escalate.
- Exit condition: what happens at expiry, and what completed review would convert the exception to a permanent approval.
What changed in the supervisory baseline
Bank business lines continue to adopt public chat assistants, coding assistants and embedded AI features faster than model risk and third-party review can complete. Risk teams need a defined path between blocking a use and approving it permanently. Two recent developments shape that path.
On 17 April 2026 the OCC, Federal Reserve Board and FDIC issued updated interagency model risk management guidance. It sets a risk-based approach and rescinds prior model risk management guidance and other issuances, so policies that cite the earlier documents should be re-based. The agencies also plan a request for information on model risk management that considers generative and agentic AI. Exception policy should be written to absorb that change.
On 11 September 2026 the agencies proposed revised third-party risk management guidance. As of November 2026 it is a proposal and not final. It would scale diligence to the assessed risk of each relationship, and it restates that using a third party does not diminish the bank's responsibility. An exception for an external AI tool therefore remains the bank's accountability.
On 4 November 2026 the OCC issued a supervisory framework for examiners on bank use of Super Intelligence, the term the OCC uses. It creates no new expectations or enforceable standards. It does list exception approval and suspension as lifecycle stages alongside risk classification, and it tells examiners to tailor their work to materiality. That is a useful signal of where examiners may look, and it supports treating an exception as a managed state with a defined end.
Build the intake and decision process
Keep intake short and structured. Each request should name the tool and provider, the business owner, the data classes involved, the intended use, and who is affected by the output. Classify risk before any exception is approved. This follows the lifecycle ordering in the OCC framework, where risk classification precedes exception approval.
The NIST AI Risk Management Framework offers a workable structure. Its four functions are govern, map, measure and manage. Applied here, which is our own mapping and not NIST's, govern covers ownership and approver roles, map covers intake and context, measure covers tiering and any testing, and manage covers enforcement and monitoring after approval.
Define three outcomes: approve, conditionally approve, or deny. Conditional approval is the temporary exception. Every decision records a rationale, the approver and an expiry. Updated model risk guidance emphasizes clear policies, roles and responsibilities, so name who can approve each tier and who can suspend an exception. Where the tool comes from an external provider, route the request into third-party risk review in parallel and scale the depth to the assessed risk.
Tier requests by data sensitivity and use-case consequence
Two axes are enough to start: how sensitive the data is, and how consequential the output is if it is wrong. The tiers below are an illustrative design for you to adapt, not a regulatory scheme. They tailor review depth to materiality, consistent with the risk-based direction of the 2026 model risk guidance and the OCC's materiality focus.
The higher the tier, the shorter the term, the narrower the scope and the more human approval is required. Outputs that feed credit decisioning, payments, fraud detection or customer communications belong in a higher consequence band than internal drafting or summarization.
- Tier 1: public or non-sensitive data, advisory output: Conditional approval is available with a short review. Typical controls are a sanctioned-user list, rate limits and redaction of identifiers.
- Tier 2: internal data, output used in business processes: Conditional approval needs the business owner and risk approval, a fixed term, data-class limits and a third-party review in progress.
- Tier 3: customer or confidential data, or consequential output: Default is deny. Any exception needs senior risk approval, a minimal scope, and human approval at the moment of specific actions.
Enforce the terms in the execution path
A ticket and an attestation record intent. They do not stop a user from pasting restricted data into a chat assistant on day 40 of an exception that was meant to cover a different use. The design hypothesis is to express each exception as machine-readable policy state: scope, permitted data classes, expiry and approver. That state is checked on every interaction and tool call.
Most exception terms map to five runtime actions. Allow applies when the request is inside scope and term. Redact removes data classes outside the permitted set. Block applies to out-of-scope tools, users or data, and is the default once the term ends. Rate limit bounds volume for a tier. Escalate for human approval pauses a specific action until a named approver responds. Suspension should take effect through the same enforcement point, so withdrawing an exception does not depend on someone remembering to act.
Some terms cannot be enforced technically, such as a business owner's commitment to review outputs. Label those as attestation-only so the exception file is honest about which controls are enforced and which are asserted.
Trussed enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants, on individual interactions and tool calls, with actions of allow, redact, block, rate limit and escalate for human approval. This is enforced at runtime, not reviewed after the fact. Trussed does not provide transaction-level governance or end-to-end tracing of agentic workflows, so scope exception terms to the interaction and tool-call level.
Evidence an examiner can test
Plan the exception file around the questions an examiner is likely to ask: who approved this, on what basis, for what scope and period, and what actually happened during that period. A complete file holds the request, the tier and rationale, the approved terms, the approver identities, the expiry date, any third-party review status, and the closure or conversion decision.
The last question is where ticket-based approvals fall short. A ticket shows that a decision was made. It does not show whether the tool was used only by the approved group, whether restricted data was redacted, or whether use stopped at expiry. Records generated by enforcement answer those questions directly, because each allow, redact, block, rate limit or escalation is recorded at the point it happened. After-the-fact attestations still have a place for terms that cannot be enforced, but they should be a minority of the evidence.
Trussed generates continuous audit-ready evidence from enforcement, mapped to NIST and NYDFS among other frameworks, with configurable retention and self-managed deployment in the customer cloud. It has run for more than 18 months in production in a regulated enterprise, and insurance, a regulated financial sector, is where it is deepest. Whatever tooling you use, set retention to match your exception and examination cycles.
A practical next step is to take your three most active shadow AI requests, assign tiers, write their terms against the checklist above, and mark each control as enforced or attested. The gap between those two columns is your remaining exposure.
Exception lifecycle as a time-bound policy state
Intake
Business line submits tool, data and use case.
Tier
Data sensitivity and use-case consequence set review depth.
Decide
Approve, conditionally approve for a fixed term, or deny.
Enforce
Terms applied to each interaction and tool call.
Expire or suspend
Use reverts to the default action, such as block.
Make each exception bounded, enforced and evidenced
Talk through how interaction-level and tool-call-level enforcement can support time-bound AI exceptions and the evidence your examiners will ask for.
Talk to an Expert