How to Retire Unapproved AI Models in Banks Under the Revised Model Risk Guidance
Retire unapproved AI models as a staged transition, not a shutdown. Map the workflows that depend on each model, triage each one into retire, remediate or onboard by risk and materiality, apply interim runtime controls (allow, redact, block, escalate for human approval), cut over to approved alternatives against defined exit criteria, and retain a record of what each model was allowed to do and did. As of November 2026, the reference point is the interagency revised guidance (SR 26-2 and OCC Bulletin 2026-13), which rescinds prior model risk management guidance.
A staged retirement approach
- 1
1. Discover and map dependencies
Build a list of unapproved models and, for each, the workflows, owners, data types and users that rely on it.
- 2
2. Triage
Assign a risk tier and a decision: retire, remediate or formally onboard.
- 3
3. Apply interim runtime controls
Constrain what each model can receive and do while it remains live.
- 4
4. Build and validate the alternative
Stand up the approved replacement or the remediated model, and compare outputs in a parallel run.
- 5
5. Cut over in phases
Shift traffic in increments with a tested rollback path until exit criteria are met.
- 6
6. Decommission and retain evidence
Revoke access, update the inventory, obtain sign-off and archive the record.
What changed in the guidance, and why scope needs a documented decision
On 17 April 2026 the OCC, the Federal Reserve Board and the FDIC issued revised interagency model risk management guidance (OCC Bulletin 2026-13 and Federal Reserve SR 26-2). It rescinds prior model risk management guidance. Programs built around SR 11-7 and OCC 2011-12 should now be mapped to the revised documents. The agencies describe the new guidance as risk-based and tailored to a bank's size, complexity and extent of model use, and say practices suited to one bank may be inappropriate for another.
Scope is the part risk leaders need to handle carefully. The Federal Reserve attachment states that its principles apply to traditional statistical and quantitative models and non-generative, non-agentic AI models. The agencies call generative and agentic AI novel and rapidly evolving and plan a request for information on model risk management that considers banks' use of AI, including those types. A third-party LLM API embedded in a live workflow therefore sits in unsettled territory. The practical response is to record your own scoping rationale for each model: whether it falls inside the stated scope, and if not, which internal standard you apply instead.
Separately, OCC Bulletin 2026-54 (November 2026) is a framework for OCC examiners. It does not set supervisory expectations for banks, and examiners are directed to tailor exams to the materiality of the use case. It is a useful signal that materiality will shape scrutiny, not a compliance standard.
Discovery and dependency mapping: methods and limits
No single method finds everything. Network egress and API traffic analysis can reveal calls to external model endpoints, including ones nobody registered, but it says little about why a call is made, and it misses models running inside vendor software or on unmanaged devices. Vendor and SaaS review catches AI features embedded in licensed products, but depends on vendor disclosure and on how current the contract and release information is. Application inventories show where teams believe models are used, but rely on self-reporting and tend to lag actual deployments. Combining the three, then confirming with business owners, gives a more complete picture than any one alone.
For each model found, record what it is actually used for in each workflow and whether that differs from its original or intended purpose. The revised guidance notes that using a model beyond its intended purpose introduces additional uncertainty and risk, so this field directly affects the triage decision. Also record who owns the workflow, what data reaches the model, and what happens downstream if it stops responding. That last item is the business impact of switching it off, and it determines how much transition time you need.
Triage: retire, remediate or onboard
The revised guidance is risk-based, and NIST AI RMF 1.0 says the risks judged highest in a given context of use call for the most urgent prioritization and the most thorough risk management. Use both to order the work. Materiality and data sensitivity should come before convenience: a model touching customer data in a significant business process outranks a low-impact drafting aid.
Retire when a approved alternative exists or the use case is not worth the control burden. Remediate when the model is acceptable in principle but lacks documentation, ownership, testing or monitoring. Onboard formally when the business need is durable and the model can meet your standards. Onboarding a third-party model should bring due diligence and validation comparable to in-house models. Industry practice for vendor models often combines example-based testing with review of the vendor's performance evidence, and a published benchmark alone is weak assurance because models can be tuned to benchmarks.
Retire, remediate or onboard decisions, risk acceptance during the transition, and exit sign-off are organizational decisions. They belong to named owners under your governance policies, not to a technical control.
What is enforced at runtime and what stays an organizational decision
Interim controls reduce exposure while a model stays live, but they do not replace judgment. Separating the two clarifies ownership and what evidence each produces. Scale the intensity of interim controls to use-case materiality.
- Enforced in the execution path: Per-interaction and per-tool-call decisions: allow the request, redact sensitive fields, block disallowed content or destinations, rate limit, or escalate for human approval. Each decision can be recorded as it happens.
- Organizational decision: Risk tiering, the retire, remediate or onboard call, scoping rationale, acceptable use during transition, exit criteria, sign-off and risk acceptance.
- Shared: Policy rules are written by risk and compliance, then enforced by the runtime layer. Exceptions routed to human approval need a named approver and a documented outcome.
Exit criteria and evidence retention
Define exit criteria before cutover starts. Typical criteria include: the replacement has passed validation for the workflow, a parallel run shows acceptable agreement with the legacy model, traffic to the old endpoint is zero for an agreed period, credentials are revoked, the inventory entry is updated, and the business owner and model risk function have signed off. These are practitioner design choices, so tailor them to materiality and keep rollback available until the last criterion is met.
The record should link each unapproved model to its dependent workflows, its scoping rationale, its triage decision, the interim controls applied, and the exit sign-off. Earlier guidance expected an inventory of models and documentation of issues and their resolution, and a record of this kind is consistent with that principle even though it is not mandated. The key point for examiners is the interval between discovery and exit: what the model was allowed to do and what it did. Review after the fact depends on reconstructing events from scattered logs, while a policy enforced at runtime produces the decision record as the interaction occurs.
Trussed enforces policy in the execution path across models, agents, tools and MCP servers, with allow, redact, block, rate limit and escalate-for-approval decisions on individual interactions and tool calls. That enforcement generates audit-ready evidence mapped to frameworks including NYDFS and NIST, with configurable retention and self-managed deployment in the customer's cloud. It does not decide whether a model should be retired, and it does not map workflow dependencies. Those remain risk-leader decisions.
Staged retirement path
Discover
Find unapproved models and the workflows that depend on them.
Triage
Decide retire, remediate or onboard by risk and materiality.
Constrain
Enforce interim policy at runtime: allow, redact, block, escalate.
Cut over
Move traffic to approved alternatives with a rollback path.
Evidence
Retain decisions, actions and sign-offs for examiners.
Plan the wind-down with enforceable controls and a defensible record
Discuss how runtime policy enforcement and audit-ready evidence can support a staged retirement of unapproved AI models.
Talk to an ExpertHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.