The first thirty calls on the AI Receptionist.

When the AI Receptionist goes live, the first thirty calls are a learning artifact, not a finished system. Callers hesitate where you do not expect, they volunteer information you did not ask for, and at least one of them will try to test it ("are you a robot?"). The agent should pass that test comfortably in tradesperson language — and if it does not, that is the prompt file you edit, not a failing you hide.

The intake discipline behind the agent is intentionally narrow: name, callback number, service, location, and a preferred date and time. Five fields, asked in roughly that order, with a callback window so the office is never guessing when the customer wants the truck to arrive. Skipping any of them is fine for a single call and a disaster at thirty calls a day.

The bookings log that fills in while the calls are live is the second-order payoff. After thirty real calls, the pattern of which intake field most often stalls the conversation is visible — and that pattern is the one you build your follow-up flow around, not the field the vendor spec sheet emphasises.

If you go live with the agent, do three things in week one: read every transcript, edit the prompt that misfired, and write down the one call type you wish the agent had handled better. That last note is your next template, not your next complaint.