Users forgive an AI feature that says it does not know. They do not forgive one that was confidently wrong about something consequential. That asymmetry should drive most of the product design.
Show the work
Citation is the cheapest trust mechanism available. If an answer comes from a document, link the document. Users who can check are far more willing to rely on answers they never actually check — the availability of verification is doing the work.
Refusal is a feature
Define the categories where the system must decline, enforce them in code rather than in a prompt, and make the decline useful: say what it cannot help with and route to something that can. A clean handover reads as competence. A vague apology reads as a broken product.
Keep the human in the loop where it matters
- Automate the retrieval and drafting; keep the approval where a mistake would be expensive.
- Make the reviewing step fast enough that people do not route around it.
- Log what was approved and what was corrected — that record is your next evaluation set.
Set expectations in the interface
The words around an AI feature do more for trust than the model behind it. Calling something an assistant sets a different expectation from calling it an answer, and the expectation you set is the standard you will be judged against.