For years, document classification was something regulated industries did and everyone else ignored. AI assistants change that. When a tool can read every file a person has access to and answer questions about it, the label on the document is what tells the tool — and the person — how to treat it.
The three pieces
Sensitivity labels
A tag applied to documents and emails — Public, Internal, Confidential, Highly Confidential — that can encrypt the content, add a watermark and restrict who can open it. Labels travel with the file. Copilot honours them: content with a label that excludes the user is not used in answers, and output derived from labelled content inherits the label.
Data-loss prevention (DLP)
Rules that watch for sensitive information — health numbers, credit-card numbers, client identifiers — and block or warn when it is about to leave the organisation, be shared broadly, or be pasted somewhere it should not go. DLP policies can now cover Copilot interactions too.
Copilot's own controls
Admins can exclude specific SharePoint sites from Copilot, restrict which apps it is available in, log prompts and responses for audit, and set retention on those logs. For a healthcare practice or a firm handling client files, these are the settings an auditor will ask about.
What to set up first
- A simple label taxonomy — four labels, clear names, one page of guidance.
- Auto-labelling for the obvious cases (anything containing a health number or a credit-card number).
- A DLP policy for the data type your business most needs to protect, in "warn" mode first, then "block".
- Copilot audit logging on, with retention agreed.
The realistic caveat
None of this replaces tidy permissions; it sits on top of them. And labels only help if people use them, which means a short, honest training session rather than a policy PDF. We implement this stack for customers in that order — permissions, labels, DLP, Copilot controls — and support it afterwards, so that "can AI see our client files?" has a documented answer.
