A confident wrong answer is the expensive kind.
Language models are built to produce an answer. In most work, that is useful. In compliance, an answer with nothing behind it can end up in a risk assessment, a contract or a reply to a supervisor. It reads just as well as a true one.
So the most important thing an AI can do here is stop.
What refusing looks like.
Take a simple question: does this party use subcontractors? Say no document in front of the AI answers it. Ours doesn't improvise. It says: "Needs evidence: no fact grounds this."
It shows what it read and what it could not check. Then the question goes to a person, with the gap in plain view. That person can ask for the document, or decide knowing it is missing.
Why this builds trust.
An expert does not trust an AI because it is usually right. An expert trusts it when they can see where it stops. Every claim has a source. Every gap is named. The person checking the work spends their time on the open questions, not on hunting for hidden ones.
Where the facts exist, it does the work.
Refusing is not the same as doing little. The AI reads contracts, certificates and registers, then drafts, chases and flags. It never approves, seals or signs. A person does.