Does an AI agent connected to your internal documents respect your access rights?

In brief.
To answer quickly, an enterprise AI agent indexes internal documents in a search engine. If that index does not faithfully reproduce each user’s access rights, the agent becomes a way around them: it hands whoever queries it information that person could never have opened themselves. The classic example: a salesperson who asks the right question and obtains material from the M&A file. It is one of the most serious blind spots in internal projects.

An architectural detail with major consequences

Rights live in your source systems: document management, email, team spaces. The semantic index is a transformed copy of that content. If permissions are not propagated to this copy, then checked on every query for the user asking, the original partitioning disappears. And it has to be maintained continuously: rights change every day, with arrivals, departures and confidential projects.

The subtleties that trip up prototypes

Documents with mixed rights (a confidential appendix in a public folder); derived content, meaning AI-generated summaries and extracts, which rarely inherits the restrictions of its source; conversation logs, which may contain sensitive information wrongly disclosed and themselves become data to protect. Designing, auditing and certifying this partitioning end to end is a multi-year undertaking. A simple test to give your CISO: create a user with restricted rights and ask the awkward questions.

The Optivalue.ai approach

Optivalue.ai applies rights at the time of each query, logs who obtained what, and its deployment as a private AI per customer avoids the additional risk of pooling: your content never sits alongside another organisation’s.

Why is the search index a problem?
Because it is a copy of the content: if it does not inherit the permissions, it bypasses them.

‍Do generated answers need protecting too?
Yes: summaries, extracts and logs inherit the sensitivity of their sources.