You do not let one client see another client's files. You do not let one business unit read another's. Your applications enforce it, your databases enforce it, your file permissions enforce it. Keeping those worlds apart is not a nice-to-have. It is the job.
Then you build one AI across all of them. One deployment, one index, one model, because standing up a separate AI for every client would be absurd. That is efficient, and it is the right call.
But a shared AI over separated data is only as separated as the AI makes it. By default, it makes it not separated at all. It indexes everything into one pile and retrieves by relevance. Ask it a question, and it can pull a passage from any client's data into the answer, and hand it to whoever asked. The walls you enforce everywhere else stop at the AI.
One question from client B, and client A's data is in the answer. That is not a flaw in the model. It is the absence of a gate.
