Choose the task people actually need
An assistant is easier to evaluate when its job is clear. Answer common service questions. Gather the context for a requested callback. Help a visitor find the right page. Pick a useful task before adding more abilities.
Look at the questions your team already receives. A repeated question with a clear, approved answer is a better starting point than a vague instruction to handle everything.
Separate approved knowledge from assumptions
Write down the information the assistant may use: current services, public contact details and approved guidance. Also write down what it must not invent, such as prices, availability, guarantees or details from another business.
Information changes. Decide who keeps the source current and how outdated answers will be reviewed. A polished conversation is not proof that the underlying facts are correct.
Make the handoff explicit
When a visitor wants a person, collect only the information needed for the agreed next step. Make clear whether the assistant has saved an enquiry, requested a call or actually confirmed a booking. Those are different outcomes.
Your team should receive enough context to continue the conversation without asking the visitor to start again. Keep ownership and the next action visible.
Test the boundaries as carefully as the greeting
Try incomplete details, unfamiliar questions, interruptions and requests outside the assistant’s scope. Check that failures explain what happened and offer a useful way forward.
Start small, review real behaviour with appropriate access and improve the workflow. Adding more tools is useful only when it improves the job the assistant is there to do.