The challenge
A typical organization has a huge amount of content: reports, contracts, memos and attachments collected over the years from different systems and formats. Some of it is sensitive and some public, but no one has ever marked which is which. When Copilot and agents start rolling out, this suddenly becomes a security question: in principle the agent might retrieve and show anything, including content the user should not have access to.
How the solution is built
The solution does not start with restricting the agent but with classifying the material. First, what content is sensitive, who owns it and who should really have access to it is worked through. With Microsoft Purview, a classification is built that marks the content’s sensitivity level and access rights machine-readably, not as a note written into the document’s name.
Only after this is the agent technically limited to seeing only the material the asker has a right to. This means the same agent gives different users different answers to the same question, depending on what they can access. Not because the agent is inconsistent, but because the access rights behind it genuinely differ.
What the solution aims for
The use of AI expands without the security risk expanding with it. The organization can show what material the agent has access to and what it does not, instead of trusting that no one notices a wrong hit.
