3 min read

The AI assistant that answers everyone as the admin

An AI assistant can only respect who may read a document if it knows who is asking. Keeping permissions in sync is becoming a setting; matching identities across every tool is still a decision, and AI prototypes, built to work, skip it.
Hundreds of keys on numbered hooks with coloured room tags, every room's key on one board, the way one AI assistant account can hold access to every tool.
Photo by Faris Mohammed / Unsplash

Take an internal AI assistant built to answer questions about company documents. It reads the shared drive through one service account, a technical login that can open every folder, so in the demo it answers everything, for everyone. The HR folder is restricted in the drive, but the assistant reads it with its own account and answers anyone who asks. Nothing in the logs looks wrong.

The exposure is a scenario; the setup behind it is not. I have seen it in large enterprises and in scaleups that run their data solutions across many tools: AI prototypes built to work, with little time for what production needs, and permissions that turn out to be the hardest thing to add afterwards.

Whether someone may read a document depends on who is asking, not on the document. An AI assistant that does not know who is asking answers everyone with the same knowledge. The vector store holding the assistant's copy of the documents is where this shows, but the problem is identity.

Syncing is becoming solved. Identity is not.

Managed retrieval services now copy permissions along with the content and keep them in sync with each data source. That is worth using. But a permission says who may read a document, and the service can only apply it if it knows who is asking, in the form every source system recognises. The shared drive knows the user by email, the wiki by an internal ID, the CRM by a username, and each keeps its own groups. When identities do not match, the service fails silently: usually as missing answers, sometimes as the wrong ones.

The prototype runs as everyone

A prototype is built to answer, so it reads through one account that can open everything, and the answers look right to everyone who tries it. Adding the user's identity "later" then becomes a redesign across the application, the index and every connected source. I have seen how hard that retrofit is, and it gets harder with every tool the data solution spans.

Enterprises spent years building identity management to tie their tools together. A scaleup meets the same problem the day one AI assistant sits across all of them, usually without anyone whose job it is.

Nobody owns the end-to-end question

Single sign-on belongs to IT, each source system to its own team, and the AI application to product. Nobody owns the question that matters: may this user see this answer? Filtering results by permission does not settle it on its own, because the retrieval service trusts whatever identity the application hands it. If the application does not authenticate the user properly, the filter enforces nothing.

The leadership mandate

The CTO's job is to treat identity as data governance for the whole organisation, not as a feature of one AI application, and to settle it before the first prototype, while it is still cheap. A managed service handles the syncing for common tools. What it cannot decide for you is which identity is the user's, and whether every tool agrees on it.

The path forward

Before an AI assistant reads a new source, check four things:

  1. One identity provider decides who each user is, for every tool.
  2. Every connected tool knows the user by the same identifier.
  3. The application authenticates the user and passes that identity with every question.
  4. When the identity cannot be matched, the assistant returns nothing rather than everything.

Then test every demo as a restricted user. If the prototype cannot be shown to someone with limited access, it is not ready for production.

Keeping permissions in sync is becoming a setting. Carrying the right identity from every tool to the AI assistant is still a decision someone has to make, and the cheapest moment to make it is before the first demo.