How to Restrict Azure OpenAI to Approved Models Under Current Model Risk Management Guidance
Keep one authoritative approved model inventory, then enforce it twice: at provisioning, to limit which deployments can exist and who can create them, and at request time, to validate which model each call actually reaches. Retain the allow or block decision, the inventory record and the policy version as evidence. As of November 2026, this work should reference the updated interagency model risk management guidance (OCC Bulletin 2026-13, issued April 17, 2026), which rescinds prior guidance.
Layer two: enforce the list on each request
- 1
1. Intercept the call
Capture the request before it reaches the model endpoint.
- 2
2. Resolve the target
Determine the actual model and version behind the requested deployment.
- 3
3. Evaluate against the inventory
Check approval status, application, use case and risk tier.
- 4
4. Decide
Allow, block, or escalate for human approval, with a recorded reason.
- 5
5. Write the evidence
Store the decision with the inventory record and policy version evaluated.
- 6
6. Route exceptions
Send unapproved model requests to the named approver and document the outcome.
Evidence to retain for model risk management and examiners
- Timestamp, calling application and identity
- Requested deployment, resolved model and version
- Inventory record and approval status at decision time
- Decision (allow, block, escalate), reason and policy version
- Exception request, approver and outcome where applicable
- Retention period and storage location, set by bank policy
What changed in the supervisory baseline
On April 17, 2026, the OCC, the Federal Reserve Board and the FDIC issued updated interagency model risk management guidance, announced in OCC Bulletin 2026-13. Its stated aims are to clarify model risk management principles, set out a risk-based approach, and rescind prior model risk management guidance and other issuances. Bank policies and control narratives that cite earlier guidance, including SR 11-7, should be reviewed and updated to reference the current issuance. This is the status as of November 2026.
Three points in the guidance matter for an approved model list. It describes a comprehensive model inventory as common industry practice, with enough information to understand model risk at individual and aggregate levels. It says adequate documentation supports effective model risk management. It also says ongoing monitoring may include assessing model issues before and after a model is approved for use. The OCC bulletin further notes that generative and agentic AI models are novel and rapidly evolving, so the approved list should be expected to change.
The guidance is risk-based. Practices suitable for one bank may be unsuitable for another, and it is generally expected to be most useful for models supporting significant business lines, operations, services and functions. Allowlist strictness and evidence depth should be justified against your risk profile and the purpose of each model.
Define the approved model inventory first
Both enforcement layers should read from the same record. If provisioning rules and request-time rules are maintained separately, they drift, and you cannot show an examiner which list was in force on a given date.
Treat approval as a go/no-go gate. A model or version enters the list only after documented approval, and removal follows the bank's risk tolerance. Reassessment is also reasonable after fine-tuning or adapting a third-party model to a new domain, since that changes the risk the approval covered.
Candidate fields are a design suggestion, not a regulatory list: model and version, approval status, approver, approved use cases and applications, risk tier, approval date and next review date. A risk tier lets you scale control depth to materiality instead of applying one standard to every model. Record each change to the list as a versioned entry, so the inventory itself becomes documentation of what was permitted and when.
Layer one: constrain what gets provisioned
The provisioning layer controls which model deployments can exist in your subscriptions and which identities can create or change them. Native Azure mechanisms such as Azure Policy, role-based access control and network controls belong here. The goal is to keep the set of reachable deployments small and traceable to the inventory.
Apply least privilege. Few identities should be able to create or modify deployments, and application identities should reach only the deployments their approved use case requires. Derive policy definitions from the inventory rather than writing them by hand, so that approving or retiring a model changes one record and the provisioning rules follow.
This layer answers one question: what was allowed to be deployed. It does not by itself answer which model a specific application call reached. A deployment that is permitted today can be repointed, an application can hold a valid credential to a deployment it was never approved to use, and a call can be sent to a deployment that is permitted but unsuitable for that application's use case. These gaps are why the request layer is needed.
- Restrict creation: Limit who can create or modify deployments and which models they can reference.
- Scope access: Grant each application identity only the deployments its approved use case needs.
- Constrain network paths: Reduce the routes by which applications can reach model endpoints.
- Derive from the inventory: Generate policy definitions from the approved record to avoid drift.
Tradeoffs, maintenance and where Trussed fits
Request-time enforcement adds a component to the call path. Plan for its latency, availability and failure behavior before it carries production traffic, and decide per risk tier whether failure means block or escalate. Maintenance is the other cost. Because generative models change quickly, assign an owner for the inventory, set review dates, and record every addition, version change and retirement so that both layers update together.
Trussed provides the request layer. It enforces policy in the execution path, allowing, redacting, blocking, rate limiting or escalating individual model calls for human approval, and it generates audit-ready evidence from that enforcement. It runs self-managed in the customer cloud, with configurable retention and data residency. Trussed has run 100+ models under one policy plane and has 18+ months of continuous production in a regulated enterprise. Its customers are in insurance, pharma, healthcare and education. Native Azure controls remain the right tool for the provisioning layer. Whichever tooling you choose, the test is the same: can you show which policy version permitted or stopped a specific call?
Layered enforcement of an approved model list
Inventory
One approved record of models, versions and use cases
Provisioning layer
Controls on which deployments exist and who can create them
Request layer
Per-call validation of the model actually reached
Evidence
Allow and block decisions tied to a policy version
Show examiners what your AI was allowed to do
See how policy enforced in the execution path produces audit-ready evidence of allowed and blocked model calls.
Request a Demo