01 / Organize knowledge around real questions
Collect recurring questions and separate general information, pricing, reservations and exceptions. Each topic needs an owner-approved answer and a recent review date. Resolve conflicting copies before the assistant uses them. A smaller, consistent knowledge base is a more useful starting point than an unreviewed collection of old messages.
02 / Separate static knowledge from live state
Opening hours and pickup directions may be knowledge-base content. Tomorrow’s available vehicles require the correct operational source. Without that integration, the assistant should explain the limit and provide the appropriate check or handoff. It must not turn an old description into a current availability confirmation.
03 / Put exceptions behind human approval
Refunds, individual discounts, emergencies and policy changes need explicit handling. The assistant can collect necessary context and explain that a person must review it. It should not say an action has completed without a confirmed system result. Identify the queue owner and what happens outside staffed hours.
04 / Test more than the ideal conversation
Try incomplete details, language changes, requests outside policy and several questions in one message. Check whether clarification and handoff happen appropriately. Assess answers against the approved business source and observe real usage costs after a limited activation. A scripted preview shows an intended process, not evidence that a live model has passed these tests.
Your working checklist
Tick these off as you prepare. This checklist stays on this page and does not submit information.
This is a project-planning guide. Your business rules and integrations are confirmed in the individual delivery scope.
Prepare your project brief ↗