What it does

Sets out how Microsoft 365 Copilot handles your tenant data: what it can see, where prompts and responses go, and what Microsoft does and doesn't do with them.

Key facts

• Copilot only surfaces content the user already has permission to open. It respects existing SharePoint, OneDrive, Exchange and Teams access.
• Prompts, retrieved data and responses stay inside the Microsoft 365 service boundary and are not used to train the foundation models.
• Interactions are protected by the same commitments as your other Microsoft 365 data, including the Data Protection Addendum.
• Grounding data pulled via Microsoft Graph inherits the source item's sensitivity and permissions.

When to use / skip

This isn't something you switch on. It's the baseline you explain to security stakeholders before they'll sign off on a rollout. Read it, then move straight to oversharing and labelling work.

Configuration decisions

• Whether users can turn on optional connected experiences and web-grounded queries.
• Whether plugins and third-party connectors are allowed, since those extend the data path.

Gotchas

• "Copilot respects permissions" is true and misleading. It respects current permissions, which in most tenants are far too broad. The privacy model doesn't fix your sharing sprawl.
• Bing web queries are handled under separate terms from the tenant boundary. Don't conflate the two when briefing legal.

Consultant notes

• The privacy model is your easy win in the security review, but it lulls people. Follow it immediately with a permissions audit or you'll be firefighting oversharing later.
• Get legal to read the DPA and EU Data Boundary commitments early. They ask the same three questions every time and it's better answered on paper.
• Don't let anyone treat "not used for training" as the end of the data-governance conversation. It's the start.

---

Review when the Copilot privacy documentation is next refreshed.