Which part first: the order that decides whether anyone keeps using it

Start where the documentation is best, not where the pain is worst

This is counter-intuitive and it is the whole trick. The area with the most pain usually has the worst documentation - that is often why it hurts - and an assistant launched there will be wrong in front of the people who complained loudest. Start instead where the answers genuinely exist and are current, even if that area was not the complaint. Being right about expenses policy buys the credibility to later be useful about something harder.

Narrow means one subject and one audience, not one department

"Finance" is not a scope; "how to submit an expense claim, for staff" is. A scope that small can be made genuinely correct in a day, tested exhaustively by three people, and judged honestly. Departmental scopes contain a dozen subjects of wildly different documentation quality, which reintroduces the original failure at a smaller size.

The signal to widen is the unanswered log going quiet

Widen when the questions arriving outside the current scope outnumber the ones inside it, and when the inside ones have stopped producing corrections. That is evidence the scope is finished. Widening on a calendar date instead - "phase two starts in March" - is how an unfinished scope gets buried under a second one, and the first corrections never get made.

Tell people the scope, in the assistant's own words

An assistant that says what it covers and what it does not is trusted within that boundary. One that attempts everything is judged on its worst answer. State the scope in the opening message and have the assistant decline outside it by name - "I only cover expenses and travel booking; for anything on contracts, ask Dana" - which is more useful than a guess and protects the rollout.