Supported Copilot experiences can route requests to different models or offer model selection. Copilot Studio can offer external providers when administrators allow them. Model-agnostic does not mean any model in every product.
Understand where model selection happens.
Some applications choose models through service-managed routing; others expose a choice to users or makers. A model offered in one product is not automatically available in another. Even where selection is exposed, the surrounding application's tools, supported context, and access policies continue to shape the result.
Reasoning capability, response speed, context capacity, and cost can vary. A larger context window is not a substitute for retrieving relevant information, and stronger reasoning does not guarantee a correct answer or a properly authorized action.
Compare choices against a task.
Use the same representative inputs and scoring criteria when comparing supported options. Keep retrieval settings, instructions, and tool permissions stable where possible. Otherwise, an apparent improvement may come from better context or a changed workflow rather than the model itself.
Separate task quality from operational requirements. A model that produces a better draft may still miss a latency target, increase review effort, or exceed the budget at expected usage. Evaluate these tradeoffs together instead of choosing solely on a benchmark or release announcement.
Example: extraction versus analysis.
Extracting a short set of fields from a standard document may favor predictable structured output and low latency. Reconciling conflicting evidence across several documents may need more reasoning and careful source inspection. The best supported choice can differ even when both tasks appear in the same application.
A switch also needs regression cases for empty sources, ambiguous requests, malformed tool responses, and denied access. Test whether downstream systems still receive the structure they expect and whether uncertainty remains visible to the reviewer.
Treat a model change as a release.
Confirm current availability, licensing, region, data handling, and administrator controls. Record the evaluated version or configuration where the service exposes it. Maintain a fallback or pause procedure appropriate to the application. These steps make change explainable without assuming that every service supports version pinning or unrestricted model interchangeability.
Further reading
See the model-selection guidance for supported options, and AI Governance for decisions and evidence needed when a configuration changes.