March 17, 2026 · Quelle Works
Access Control Is the Hard Part of Enterprise AI Search
Placeholder description: connecting an AI tool to your documents is the easy part — preserving who's allowed to see what is where most implementations quietly fail.
The part that gets skipped in a demo
[Placeholder copy.] It’s straightforward to point a retrieval layer at a document store and get impressive-looking answers in a demo. It’s a different problem entirely to guarantee that a search result never surfaces content the person asking wasn’t already permitted to see.
Where permissions actually live
[Placeholder copy.] Source systems like SharePoint, Drive, and Notion already encode permissions — at the file, folder, or workspace level. A retrieval layer that doesn’t map those permissions through to query time isn’t a shortcut, it’s a new access control gap that didn’t exist before.
What proper implementation looks like
[Placeholder copy.] Permissions need to be checked at query time against the same source of truth the underlying system already uses, not cached into a separate, quietly-drifting copy. This is slower to build and less demo-friendly than ignoring the problem, which is exactly why it’s usually the differentiator between a real implementation and a prototype.
Related: M&A technical due diligence case study, where scoped, temporary access was the entire point.