Skip to main content

Model Governance

Do not select the most capable model by default. Maintain an approved portfolio that matches models and deployments to use-case sensitivity, quality, cost, latency, throughput, availability, and processing-location requirements.

Model Register

For each approved model connection, record:

FieldWhy it matters
Provider and model familyContract, capability, support, and risk ownership
Deployment name and typeRuntime configuration and processing boundary
Region or data zoneResidency and operational review
Approved use casesPrevents accidental use in unsuitable scenarios
Preview or production statusPreview terms and behavior may differ
Input/output modalitiesDefines data types the model can receive and produce
Evaluation resultShows quality and safety evidence for the intended use
Quota, limits, and cost ownerSupports capacity and budget control
Retirement or review datePrevents unmanaged model drift and end-of-life risk

Portfolio Pattern

Use high-capability reasoning models for complex planning and analysis, smaller efficient models for high-volume extraction or classification, embedding models for retrieval, and specialized audio or image models only where required. Separate highly regulated or deterministic workloads from general assistants.

If a model router is used, approve the routing mode and allowed model subset. The router does not remove the need to review every possible underlying model and the deployment's processing location.

Change Control

Re-evaluate a production agent when the provider changes a model version, a preview becomes generally available, a model is retired, routing membership changes, or deployment type changes. Preserve the previous evaluation result and rollback decision.

See Azure AI Foundry and Model Deployment and Data Residency.