When Microsoft 365 Copilot comes up, admins and data protection officers almost always ask the same thing first. How does Copilot decide which files to show, and how do we keep sensitive data private? I hear this in nearly every Copilot pilot I accompany. The answer is the Copilot semantic index. There's no mysterious AI magic involved. It's a structured process that respects user permissions and your organization's data governance.

The Copilot semantic index is a permission-aware catalog and not a black box

Every tenant gets a semantic index that Microsoft builds and maintains automatically in the background. You can picture it as an advanced catalog of all indexed content from SharePoint, OneDrive, Exchange and other sources. Classic full-text search matches words. The semantic index also understands topics, relationships and context. When someone asks Copilot a question, Copilot goes through the Microsoft Graph API, which checks permissions and pulls the relevant content from this index (see Microsoft's documentation).

The key point is that Copilot only ever surfaces files and messages the user could already open directly. This is built into the system, not just stated in a policy. Every Copilot prompt triggers a permission check through Graph. In my experience, privacy officers focus on exactly this part, and they're usually skeptical at first. Copilot cannot get around access controls, which is reassuring. The downside is that any oversharing in SharePoint or OneDrive shows up in Copilot right away. If files or folders are shared too broadly, Copilot finds them as easily as a user clicking through their apps would. I've written up practical ways to tighten permissions and cut down oversharing in my article on the Copilot oversharing deployment blueprint.

Freshly created files need some time before they reach the index

People regularly wonder why brand-new files or recently edited documents don't appear in Copilot straight away. That's down to how indexing works. SharePoint and OneDrive content is crawled and indexed continuously, but never instantly. Copilot (and Microsoft Search) only see a file once it has been fully indexed and added to the semantic index. Depending on the size of the tenant and how much is going on, that can take anything from a few minutes to several hours. This isn't a bug. It's a technical limit you should simply be aware of.

There is a small workaround. Opening and editing a file directly in SharePoint Online sometimes speeds up indexing, though nothing guarantees it.

Practical tip: For quick Copilot testing, upload test files to a well-known SharePoint site, wait five to ten minutes, then try again in Copilot.

Microsoft Search settings and restricted content discovery limit what Copilot sees

Many people forget about the Microsoft Search settings, yet the semantic index is closely tied to whatever Microsoft Search includes. A site, library or file that's excluded from search (through compliance or governance controls) is invisible to Copilot as well. The strongest lever you have here is restricted content discovery for SharePoint. It lets you deliberately exclude whole sites or libraries from indexing, and therefore from Copilot search. For confidential projects or sensitive team spaces, I consider this essential. Details, pitfalls and real-world experience are in my article on restricted SharePoint search and Copilot oversharing.

This simple diagram shows how a Copilot query flows through the semantic index, Microsoft Graph and the Search settings:

flowchart TD
  User["User asks Copilot"]
  Copilot["Copilot prompt"]
  Graph["Microsoft Graph checks permissions"]
  Index["Semantic index searches content"]
  Search["Microsoft Search settings"]
  Result["Copilot result"]
  User --> Copilot --> Graph --> Index --> Result
  Index --> Search

Figure: A Copilot query passes through the semantic index and the Search controls.

Copilot makes existing oversharing obvious

What surprises many organizations most is how clearly Copilot exposes old, overly generous sharing that nobody had noticed. Copilot only answers with what each user can access. The catch is that this set of content is often far larger than anyone assumed, particularly where "Anyone with the link" sharing or excessive group permissions exist. All of a sudden confidential files pop up as Copilot suggestions, simply because nobody secured them properly.

If you want to fix this, concentrate on these essentials:

  • Regular SharePoint and OneDrive permission reviews
  • Sensitivity labels and DLP policies
  • Use of restricted content discovery
  • Governance rules for new teams and groups
  • Clear communication to content owners

More on sensitivity labels and protection mechanisms is in my article on Copilot, sensitivity labels and Purview protection.

Web grounding follows completely different rules

Keep all of this separate from web grounding. When it's turned on, Copilot can add Bing search results from the public web to its answers. The rules here are not the same at all, because web grounding isn't filtered by your tenant's permissions. Most organizations have to control this exposure, and you do that through the Copilot and Bing settings. I've covered web grounding from an admin's point of view, including what data actually leaves Microsoft 365 and which controls work in practice, in this article on controlling Copilot web grounding and Bing.

Better governance leads to fewer Copilot surprises

Once you see how the Copilot semantic index, Graph and the Search settings fit together, you can decide up front what Copilot can and can't find for each user group. In practice, Copilot tends to reveal governance gaps that have existed for years, because users suddenly realize how much "sensitive" content sits in overshared libraries. My advice is to run a thorough review of permissions and content before you roll Copilot out widely. The technical foundation is solid. What really decides the outcome is the human side, meaning governance and clean permissions. If you want to read more about how Copilot, the semantic index and privacy interact, have a look at the official Copilot overview and the Copilot privacy controls.