How to Stop Copilot From Surfacing Sensitive Files in Microsoft 365

Here is the sentence that makes every IT leader pause during a Copilot rollout: Copilot will only show a user files that user already has permission to open. That is reassuring for exactly one second, until you realize how much sensitive material people can already open in your SharePoint.

Copilot does not create new access. It inherits the access each person already has, and it reads far faster than any human ever will. A finance spreadsheet sitting in a broadly shared folder, an HR document with an open sharing link, a contract library everyone in the company can browse. None of that is a new permission. It is all already visible. Copilot simply makes it findable, in seconds, through a chat box.

The fix most organizations skip is sensitivity labels. A sensitivity label is a tag you attach to a file or a site that says what kind of information it holds, like Confidential or Internal Only. Once a label exists, you can attach rules to it: who can open labeled files, whether they can forward them, and critically, whether Copilot is allowed to use them when it builds an answer.

Here is a practical setup you can run in about half an hour. It will not solve every data governance problem your company has, but it will put a real guardrail in front of Copilot this week.

First, create two or three labels in the Microsoft Purview portal, which is where Microsoft 365 data protection settings live. Keep the set small. Labels like Confidential, Internal, and General are enough to start. A ten-label taxonomy looks impressive in a slide deck and dies quietly in real life because nobody remembers which label applies.

Second, publish a label policy that turns those labels on for the right people. A label policy is simply the rule that decides who sees which labels in their Office apps. Start with your pilot group rather than the whole company, so you can see how the labels behave before everyone is labeling files.

Third, set the Confidential label to block Copilot. In the label's settings, you can exclude labeled content from being used by Copilot and other AI tools. This is the single highest-value step in the whole setup, and it takes about two minutes.

Fourth, apply the labels where they matter most. You have two options. You can let people label their own files, which works if your team will actually do it, or you can use auto-labeling, where Microsoft scans files for patterns like Social Security numbers or credit card numbers and applies the label automatically. Auto-labeling costs extra licensing, so start manual in the pilot and add automation when the pilot proves its value.

Finally, test it the way your users will experience it. Ask Copilot a question that would normally surface one of the labeled documents, and confirm it does not. Then ask about an unlabeled public document, and confirm it does. Testing both directions is what separates a guardrail that exists from a guardrail that works.

There are two pitfalls worth naming. The first is label sprawl, which is what happens when every department invents its own labels and nobody can keep them straight. Three labels that everyone understands beat fifteen that nobody uses. The second is forgetting the guest and external sharing settings on your SharePoint sites, because a label on a file does not help much if the whole library is shared with anyone who has the link.

If you take one thing from this post, take this: Copilot makes your existing permissions searchable, so your existing permissions need to be worth searching. Labels are the fastest way to draw the lines you should have drawn years ago.