Do sensitivity labels stop Copilot reading files?

That’s a genuine concern, and you’re wise not to assume “we labelled it” equals “Copilot can’t use it.” Labels classify; Copilot still grounds on what the signed-in user can already open — unless you add a policy that says otherwise.

Here’s what’s happening. Microsoft Purview can target the Microsoft 365 Copilot and Copilot Chat DLP location. With Content contains → Sensitivity labels and the action Prevent Copilot from processing content, labelled files and (supported) emails are excluded from the response summary: Copilot doesn’t use the item’s content, but the item can still appear in citations. Separately, sensitive information types (SITs) can block prompt processing or block web search grounding when the prompt itself contains sensitive data. You can’t mix SITs and sensitivity labels in the same rule — use two rules in one policy. Copilot still runs in the user’s security context: no DLP magic invents access the person never had. Learn’s Contoso-style guidance is exactly this pattern — keep Highly Confidential / Personal out of summaries while Copilot stays useful on Internal and Public.

This is Third Law in organisational clothing: the easier it is for people (and therefore Copilot) to reach sensitive work, the easier it is for that work to show up in a chat summary. 12 Panes puts Security → GRC in Steady before broad seat spray.

Practical options:

1. Labels first, then Copilot DLP. Publish a clear taxonomy; auto-label where you can; add a Copilot-location DLP rule for Highly Confidential (and anything regulated). Run simulation, review Activity Explorer, then enforce. Best default for most organisations.
2. Prompt and web-search SIT rules. Stop credit cards / national IDs in prompts, or allow work grounding but block external web search when SITs appear. Complements labels; doesn’t replace permission hygiene.
3. Rely on labels alone (no Copilot DLP). Classification without the Copilot location still helps humans and other controls; it does not reliably keep labelled content out of Copilot summaries. Incomplete for your stated goal.
4. Block Content Analysis Services on a label (PowerShell). Narrow Office-app behaviour; Learn notes content can still be available to Copilot in other scenarios (for example Chat / Teams). Treat as specialist, not your only fence.

The caution: Restricting Copilot from processing labelled files/emails via this DLP location is typically an E5 / Purview Suite class capability — check Purview service description for your SKU. Prompt safeguards are broader. Calendar invites aren’t covered the same way as files/emails; open-file skills in Word/Excel/PowerPoint evaluate at open. Product names move — re-read Learn about DLP for Microsoft 365 Copilot before you promise the board “invisible.”

I’d suggest one sentence for Security and Adoption together: labels tell us what the file is; Copilot DLP decides whether that content may enter a summary — and citations may still name the document, so oversharing remediation still matters.

Steady before scale: fence the Highly Confidential set, then widen Copilot. You’re asking the right questions.

DIGITAL SKILL WORKS LTD / 12 PANES · You ask, we answer