How to Select a Preview Response Model in Copilot Studio

Microsoft Copilot Studio can let eligible makers choose a preview model for an agent's generative responses. Open the agent's Settings, find the Responses section, and select a model under Primary response model. The preview feature is currently limited by region and environment policy, and Microsoft advises against using preview or experimental models for production agents.

Before you change the model

• Confirm availability: The generative-response preview described here is available only to eligible makers in supported regions. The models shown can differ by environment.
• Check administrator policy: An administrator must allow preview and experimental AI models for the environment.
• Review data location: Microsoft states that data processed by a preview model may be processed or stored outside the organization's geographic boundary.
• Use a test agent: Preview models can have limited capacity, higher latency, errors, or future availability changes.

These controls are separate from model choices in Prompt Builder and from newer agent experiences. If the interface does not contain a Responses section, confirm which Copilot Studio authoring experience and agent type you are using.

Select a primary response model

1. Open Microsoft Copilot Studio and select the agent you want to test.
2. Open the agent's Settings page.
3. Find the Responses section.
4. Open Primary response model and choose an available default or preview model.
5. Save the setting if Copilot Studio asks you to do so.
6. Open the test chat and run the same set of questions you used with the previous model.

You can return to the default model from the same control. Record the original selection before testing so that rollback is straightforward.

What the primary response model affects

The selected model is used when the agent generates answers instead of following a manually authored dialog path. A different model may change response quality, summarization, latency, and how consistently the agent follows instructions. It does not automatically improve weak knowledge sources, permissions, or agent instructions.

If the agent should answer questions beyond its configured knowledge sources, review the setting that permits use of the model's general knowledge. Enabling that option increases the range of questions the agent can answer, but it can also produce ungrounded responses. Leave it disabled when answers must come only from approved organizational sources.

How to compare a preview model with the default

Use a fixed evaluation set instead of judging models from one impressive response. Include:

• common user questions with known correct answers;
• questions that require citations or information from connected knowledge;
• ambiguous requests that should trigger clarification;
• out-of-scope questions the agent should refuse or redirect;
• long conversations that test consistency and context retention;
• prompts containing sensitive or unauthorized requests.

For each response, score factual accuracy, source grounding, instruction following, completeness, tone, latency, and any unexpected disclosure. Run important prompts more than once because model output can vary.

Why the preview model may be unavailable

• The maker or environment is outside the supported region.
• The environment administrator has disabled preview and experimental models.
• A cross-region data setting required by the selected model is not enabled.
• The model has limited capacity, is temporarily unavailable, or has been removed.
• The agent uses a different authoring experience with model selection in another location.

Do not bypass an organizational restriction by moving production data into an unapproved test environment. Ask the Power Platform administrator to confirm the applicable policy and data-residency requirements.

Production guidance

Preview and experimental models have not completed the same release process as generally available models. Microsoft warns that they may slow down, time out, behave inconsistently, or stop being supported. Use them for controlled evaluation, remove sensitive test data, and keep a documented rollback path.

For current region tables, model tags, and interface instructions, consult the Microsoft Copilot Studio model-selection documentation. Availability changes faster than a static list of model names, so the options displayed in your tenant should be treated as authoritative.