The call I would never hand entirely to AI

Imagine a customer calling about an invoice they believe is wrong. They have the invoice number in front of them and want someone to explain the charge. A receptionist can collect their name, callback number and the disputed item. The person responsible for billing should decide what happens next.

For an AI receptionist, that boundary belongs in the call plan before the first customer uses it. Write down the requests it may handle, then name the person responsible for requests outside that scope. A request for a human deserves a clear route too.

Start with one ordinary business situation. For the imagined invoice call, an intake note could say: “Customer disputes the second line item and requests a billing callback.” Avoid turning that into “Customer owes the full amount” or promising a refund. The caller has described a problem; the business has not resolved it.

Give the human enough context to respond. A useful handoff includes the caller’s own description, the relevant reference number and the next step the receptionist communicated. Collect only the details needed for that task. Decide in advance what to say if the responsible person cannot answer, and avoid promising a callback time the team has not agreed to meet.

Before using the workflow with customers, rehearse the handoff with a colleague. Try a straightforward request, a disputed charge and a caller who asks for a person. Check both sides: what the caller hears and what the staff member receives. Treat an unsuccessful transfer as an open problem, even if the intake conversation sounded polished.

Choose one call type for that rehearsal. Write down the transfer trigger and the receiving person’s role. Keep the test narrow enough that someone can inspect the full conversation and decide whether the handoff works.