3.1.6. Third-Party AI Providers and Copilot Connectors
💡 First Principle: Both topics are about Copilot reaching outside Microsoft's own processing. Third-party AI providers change who processes your prompts; connectors change what data Copilot can ground on. Each needs an admin to decide whether the benefit is worth the different terms, data paths and licensing.
AI providers fall into three legal classes:
| Class | Terms | Default |
|---|---|---|
| Microsoft-hosted models (for example Azure OpenAI) | Data stays with Microsoft | — |
| AI subprocessor — processes data on Microsoft's behalf | Under Microsoft's DPA, Product Terms and enterprise data protection | Varies by provider |
| AI independent processor — processes data under its own terms | Outside the DPA and Product Terms | Off until an admin enables it |
Anthropic as a Microsoft subprocessor is the case most likely to be tested:
- It's on by default for most commercial-cloud tenants, but off by default for tenants in the EU, EFTA and UK.
- Anthropic models are excluded from the EU Data Boundary and in-country processing commitments, and aren't available in GCC High, DoD or sovereign clouds.
- Manage it at Copilot > Settings > View all > AI providers operating as Microsoft subprocessors, as an AI Administrator or Global Administrator. Disabling it can stop some features working.
- Anthropic models with data retention are a separate, independent-processor offering: off by default everywhere, governed by Anthropic's own terms, with inputs and outputs stored for up to 30 days.
Independent-processor providers are managed under Copilot > Settings > Data access > AI providers operating as independent processors. Assignments are made per provider, not per model, to users or groups (up to 999 assignments per provider), and changes take hours to propagate. Power Platform has its own additional control for external models in Copilot Studio.
Copilot connectors (formerly Microsoft Graph connectors) bring external content into Copilot and Microsoft Search. Configure them at Copilot > Connectors > Gallery, as an AI Administrator (prebuilt synced connectors also need admin access to the external system). Microsoft offers 100+ prebuilt connectors, such as ServiceNow, Salesforce, Confluence, Jira and Google Drive.
| Connector type | How it works | Who sets it up |
|---|---|---|
| Synced — tenant configuration | Indexes content, metadata and access control lists into Microsoft Graph | Admin |
| Synced — self-serve | Indexes limited recent content per user after each user consents | Admin enables, user connects |
| Federated | Fetches live at query time through MCP, with no index — but data still flows at runtime | Admin |
Licensing: content from synced connectors appears in Microsoft Search for any Microsoft 365 license, but grounding Copilot on it requires a Copilot license. Every user who queries a federated connector needs a Copilot add-on license (or Microsoft 365 E7) — Copilot Studio and pay-as-you-go licenses aren't supported for federated connectors.
Operational points: incremental crawls don't reprocess permission changes, so run a full crawl after any ACL change; turn off Copilot visibility to hide a connector's results quickly (exclusion takes up to 30 minutes); and the parsed-text item size limit is 30 MB.
⚠️ Exam Trap: "Federated connectors don't move data, so any license works" is wrong on both counts — data flows at query time, and every querying user needs a Copilot license. And Anthropic is not "off unless you opt in" for most commercial tenants — it's on by default outside the EU, EFTA and UK.
Reflection Question: A German company asks whether Copilot will send its users' prompts to Anthropic, and whether that stays within the EU Data Boundary. What would you tell them about the default and about the boundary, and where would you change the setting?