Why existing permissions matter
Copilot acting on behalf of a user does not create new access. It works with the access that user already has. That distinction matters. A file that has been sitting unnoticed on an overshared SharePoint site for years could potentially appear in a Copilot answer in seconds. AI agents that operate using their own identity or a maker's credentials introduce different considerations and require appropriate governance.
This is why secure AI adoption should start with readiness, not with agents. Before organizations expand their use of Copilot and AI agents, they need to understand the data, identities, permissions, and security controls those technologies depend on.
So, how can organizations identify where potential gaps may exist? A useful starting point is to ask five questions.
Five questions to ask before scaling AI
- Do we know which AI apps employees use today, and has each one been marked as sanctioned, tolerated, or unsanctioned?
- Do we know how many SharePoint sites are open to “Everyone except external users,” and has each one been reviewed?
- Have we run an oversharing assessment in Microsoft Purview and acted on the highest-risk findings?
- Are our most sensitive SharePoint sites labeled with a Microsoft Purview sensitivity label?
- Is prompt injection protection active for email, are supported local AI agents protected, and does the security team know how to respond?
These questions reflect areas addressed in the self-assessment included in our eBook, including shadow AI, data exposure and oversharing, sensitivity labels, SharePoint access and agent governance. From there, organizations can begin building a more structured approach to secure AI adoption.