Note
Access to this page requires authorization. You can try signing in or changing directories.
Access to this page requires authorization. You can try changing directories.
Copilot Studio offers different types of models. These model types are based on their intended use and availability. AI model selection isn't a one-time design decision. Models are introduced, updated, made generally available, selected as defaults, and eventually retired. An agent that works well with one model can behave completely differently with another model, even a model in the same model family.
Treat model lifecycle management as an ongoing operating practice for every production Copilot Studio agent. Establish a repeatable process to discover model changes, evaluate candidate models, prepare for retirements, migrate affected agents, and monitor quality after deployment.
Changing the model an agent uses is rarely just a model selection change. A newer model can interpret instructions more literally, choose tools differently, produce different response length and formatting, and change latency. Plan each model change as a migration that includes evaluation, instruction and tool refinement, approval, and post-deployment monitoring.
The governing principle is to architect for flexibility, operate conservatively, and gate every upgrade with evaluation. Upgrading to every new model risks silent regressions. Avoiding all model changes guarantees an emergency when a retirement arrives.
Apply the following lifecycle to production agents:
- Discover new, updated, default, and retiring models.
- Inventory the agents, environments, owners, and business processes that depend on each model.
- Evaluate candidate replacement models against an established baseline.
- Approve the migration by using documented quality and operational criteria.
- Deploy through the organization's Application Lifecycle Management (ALM) process.
- Monitor production results and add newly discovered scenarios to the regression suite.
- Repeat the process as models and agent requirements evolve.
This article covers discovery and inventory. The series continues with:
- Decide when to upgrade an AI model: Upgrade decision, prerequisites, and responding to retirements.
- Run and validate a model migration: Evaluation, approval, deployment, and monitoring.
Model lifecycle management requires coordination between agent owners, makers, platform administrators, testers, security and compliance teams, and release approvers. Assign ownership before a model change creates an urgent migration.
Understand the model landscape
Before you can plan a model change, you need to know how Copilot Studio classifies models, which models your organization can actually use, and which agents depend on each one.
Understand model release types
Copilot Studio identifies models by release and availability classifications. These classifications help determine how to govern a model and where to use it. Model names, release stages, regional availability, and retirement status change over time. Always check model availability by region for current information instead of relying on a static model list.
An agent that uses the default model moves to a new model whenever the default is upgraded, whether or not you planned for it. For high-risk and high-volume agents, select a specific model rather than tracking the default, so that every model change goes through your migration process.
Warning
Experimental and preview models can have limited availability, variable response quality, different latency or message consumption, timeouts, and regional data-processing considerations. Copilot Studio doesn't recommend them for production agents. If you publish an agent that uses a preview or experimental model and users interact with it, that use is still billed at the established rates.
Match the model use category to the agent's purpose
Copilot Studio tags each model with a use category that describes what the model is optimized for. Choosing the right category for the agent's workload affects quality, latency, and credit consumption.
- Deep: Optimized for deliberate, multistep reasoning and tool-supported workflows. Best for complex analytics, policy analysis, and document synthesis. Carries the highest latency and credit consumption.
- Auto: Covers mixed workloads by routing queries dynamically. Best for helpdesk and employee agents with unpredictable or varied query complexity. Latency and cost vary per turn.
- General: Optimized for speed and cost on everyday chat and light grounding. Best for drafting, summarizing, FAQ-style answers, and simple action automation. Lowest latency and credit consumption.
Learn more in Model use categories.
Important
The most common upgrade mistake is a use category mismatch, such as moving a high-volume FAQ agent from a general model to a deep model because the deep model scores better. Answer quality can improve marginally while latency and credit consumption rise sharply. This change is a net regression in user experience and cost.
Understand external models and administrator controls
You can use models from external providers such as Anthropic, xAI, and Mistral as an agent's primary model. Learn more in Choose an external model as the primary AI model.
Administrator settings control which models makers can select in an environment. A model that's documented as available might still be unavailable to the agent you're migrating if the required setting isn't turned on.
| Administrator setting | Effect on model availability |
|---|---|
| Preview and experimental AI models | Turn on before makers can select preview or experimental models in an environment. |
| Move data across regions | Required for cross-geo models. The tenant administrator manages this environment-level setting in the Power Platform admin center. |
| External models | Turns on external providers for an environment or environment group. You must also allow access to each provider separately in the Microsoft 365 admin center. This requirement makes external models the one class that needs two independent administrator actions. |
Note
Preview and experimental models and external models are governed by separate settings. Enabling one type doesn't enable the other. An administrator can allow preview and experimental models while blocking external models, or the reverse.
Before planning a migration, confirm the candidate model is available to the maker in the target environment. The model list in Copilot Studio reflects your administrator settings and is the ground truth for what a specific agent can use. Learn more in Admin controls for AI model selection.
Review model availability regularly
Review primary AI model for your agent regularly. It's the authoritative source for the current model list. New models appear there when they're introduced, and existing models update as they become generally available, become the default, or are retired.
Use the following sources together:
| Source | Description |
|---|---|
| Select a primary AI model for your agent | The primary source for model availability and new model introductions: model names, use category tags, release tags, per-region availability, cross-geo flags, retired status, US Government cloud availability, and administrator controls. |
| The model list in Copilot Studio, on the agent's Overview page under Model | What's actually available to a specific agent in your environment, given your administrator settings. |
| Continue using a retired AI model | How the retired-model compatibility window works and how to turn it on. |
| Microsoft 365 Message center and Power Platform admin notifications | Tenant-targeted change and retirement announcements. |
| Copilot Studio release plans and What's new in Copilot Studio | The forward-looking model and capability roadmap. |
| Early release cycle environments | Advance validation of platform and model changes before they reach business-critical environments. |
| Manage Copilot Studio credits and capacity | What your tenant uses, and at what consumption, by model. |
| Model provider upgrade guidance | Behavior changes between model generations, and the prompt changes that address them. |
Also trigger a review when:
- A relevant model becomes available in preview or generally available.
- The default model changes.
- A model retirement or automatic upgrade is announced.
- A model becomes available in the organization's region.
- The organization enables cross-geo processing, external models, or preview and experimental models.
- Production monitoring identifies a quality, latency, reliability, or consumption concern that another model might address.
Maintain a model and agent inventory
Use the agent inventories provided in the Power Platform admin center, the Power Platform CLI, or Power Platform APIs to identify agents that use a specific model. Use this information to begin model lifecycle communications with the affected business and technical owners.
Use either of the following views in the Power Platform admin center:
| Power Platform admin center view | How to use it |
|---|---|
| Manage > Copilot Studio > Model column | Review agents across the tenant and identify the model configured for each agent. Filter or export the results to find agents that use the model scheduled for retirement. |
| Licensing > Copilot Studio > Environment > Message consumption details > LLM Model column | Select an environment and review message consumption by LLM model. Use this view to identify environments, agents, and recent consumption associated with the retiring model. |
Query the inventory API to find agents by model
The Power Platform admin center views are effective for reviewing and exporting results manually. Query the inventory API instead when you want to collect the same information programmatically, so that agent enumeration can be scripted, scheduled, and repeated across the whole tenant instead of downloading reports by hand. Organizations with a large agent estate can use this approach to refresh the affected-agent list on demand during a retirement migration and keep it current between lifecycle events.
The inventory API returns the agent name, display name, environment, and configured model in a single tenant-wide query, so no correlation with another data source is required.
Before you run the query:
- Sign in with an account that has the Power Platform administrator or Dynamics 365 administrator role.
- Acquire an access token for the
https://api.powerplatform.com/resource. - Confirm that agent inventory is enabled for the tenant.
Send a POST request to the resource query endpoint, filter on the microsoft.copilotstudio/agents resource type, and project the fields you need, including properties.model:
POST https://api.powerplatform.com/resourcequery/resources/query?api-version=2024-10-01
Authorization: Bearer <access-token>
Content-Type: application/json
{
"TableName": "PowerPlatformResources",
"Clauses": [
{
"$type": "where",
"FieldName": "type",
"Operator": "in~",
"Values": ["'microsoft.copilotstudio/agents'"]
},
{
"$type": "project",
"FieldList": [
"name",
"properties.displayName",
"properties.model",
"environmentId = tostring(properties.environmentId)"
]
}
],
"Options": { "Top": 200 }
}
The response returns one record per agent. Field names in the response replace the dot with an underscore, so properties.model is returned as properties_model:
{
"totalRecords": 158,
"count": 200,
"data": [
{
"name": "00000000-0000-0000-0000-000000000000",
"properties_displayName": "Sample Agent",
"properties_model": "GPT-5 Auto",
"environmentId": "00000000-0000-0000-0000-000000000000"
}
]
}
The response includes totalRecords and, when the results are truncated, a skipToken value. Pass that value back in Options.SkipToken and repeat the request until all records are retrieved.
Group the collected records by properties_model to see where each model is used across the tenant. The following example shows the agent count by model for a tenant, excluding agents that use the Copilot Studio default model or run in the Microsoft 365 Copilot experience:
Model Count
----- -----
Claude Sonnet 4.6 24
GPT-5 Chat 22
GPT-5.5 Chat 5
GPT-5 Auto 4
Claude Sonnet 4.5 3
Claude Opus 4.6 2
Claude Opus 4.7 1
Claude Opus 5 1
Claude Sonnet 5 1
GPT-4o 1
GPT-5.6 Reasoning 1
When a retirement is announced, filter the same result set on the retiring model to produce the list of affected agents, their environments, and their agent IDs. Use the environmentId value to map each agent to a named environment, and route the results to the owners of those environments. Learn more in Respond to a model retirement for the full retirement response.
Learn more in:
Use Power Platform CLI for environment-scoped detail
Use pac copilot list when you need the agents and solution context for a single environment, such as when you're preparing a migration in one environment:
pac copilot list --environment <environment-id-or-url>
The command returns the agent name, Copilot ID, component state, managed status, solution ID, status code, and state code. This output doesn't include the model, so use the inventory API to identify agents by model. Use pac admin list to retrieve environment names and IDs.
What to record for each agent
The model alone isn't enough to plan a migration. Record the following for each agent, so that when a retirement is announced you already know who to contact and what work each agent needs:
- Agent name, agent ID, environment, and environment type, such as development, test, or production.
- Business criticality.
- Business owner, technical owner, tester, and release approver.
- The configured model and its release tag.
- Whether the agent uses the default model or a specific selected model.
- Cross-geo processing requirement and regional constraints.
- Whether the agent is powered by the standard harness or the GitHub Copilot harness, which determines the evaluation test methods available to it. Learn more in Choose a harness.
- The regression test set location and the date of the last baseline run.
- Whether the retired-model window is in use, when it expires, and who approved it.
Next step
With the model landscape understood and your agent inventory in place, use the decision criteria to determine whether an upgrade is needed.