Every company brain demo ends with "and it knows everything about your business."
That's the pitch. To a CISO, it's also the entire objection.
Why the objection is correct
A system that ingests every document, ticket, contract and table has, by construction, assembled the most sensitive object in the company. Then you point an LLM at it and expose it via chat.
The security question isn't paranoia. It's the deployment blocker, and it's usually raised too late.
The four questions that decide whether it ships
| Question | Bad answer | Good answer |
|---|---|---|
| Does the index respect source permissions? | Flattened at ingest | Entitlement preserved per concept |
| Two users, same question — same answer? | Yes | No, and correctly different |
| When is entitlement evaluated? | Per corpus, at build | Per query, at compile time |
| Can you prove a specific person could see a specific answer? | Access logs | The SQL with predicates inline |
Row two is the sharpest test. If two people with different clearances get identical answers, entitlement was lost somewhere in the pipeline — and "we filter the response" means the retrieval already surfaced it internally.
Why retrieval architectures struggle here
Embeddings don't carry ACLs. When you flatten a corpus into vectors, source permissions are metadata at best, and reconstructing them at query time is approximate by nature.
Compile-time governance inverts the order: policy is evaluated as part of resolving the question, so an unauthorised answer is never assembled. Nothing to filter, because nothing was retrieved.
What to require
Policy attached to concepts, not documents. RBAC plus ABAC plus row and column predicates injected before execution. De-identified resolution paths for aggregate questions. And a point-in-time reproducible audit trail — which your annual assessment needs anyway.
The full breakdown — the security architecture, the privacy model, and how compile-time governance handles unstructured sources — is here:
👉 Company Brain Security: Deterministic Governance for Enterprise AI
Originally published at colrows.com/blogs/company-brain-security-privacy
Top comments (0)