AI + legal docs without leaks: how we built it in DocuChat
In DocuChat, we looked at two approaches for AI in legal documents: an on-premise LLM and contractual guarantees from the provider. We also added ground truth verification to reduce the risk of hallucinations and fake citations.

In DocuChat, we ran into the classic 2025 problem: everyone wants AI, but dumping contracts into the cloud is a bit much, especially when they contain trade secrets. We looked at 2 approaches: on-premise LLM and contractual guarantees from the provider. The first gives you maximum control; the second is faster to launch, but then it becomes a question of trust and legal boundaries.
Another challenge: when checking wording, AI sometimes hallucinates and confidently cites a law that does not exist. Yes, like a colleague who “definitely remembers,” but no 😅 So we added ground truth verification: answers are compared against confirmed sources to reduce the risk of fake citations.
And yes, without authentication and access control, you should not ship systems like this to production at all. In the era of the AI Act and constant news about data leaks, that is no longer a feature — it is the baseline.