What an AI receptionist should refuse to do

Most worry about AI answering the phone is about how it sounds. That is the wrong thing to worry about. A slightly robotic agent that says “I'll check and have someone call you back” costs you nothing. A warm, natural, articulate agent that invents a price costs you a customer and an argument.

Fluency is not accuracy, and it reads like it is

Language models are trained to produce plausible text. Plausible and true overlap most of the time, which is exactly what makes the gap dangerous: the wrong answer arrives in the same confident voice as the right one. Your caller has no way to tell them apart, and neither do you until somebody turns up.

You cannot prompt this away with “be accurate”. What works is narrowing what the agent is allowed to say at all.

The four refusals every agent should have

• Never state a price that is not on the list. “That depends on the work — they'll confirm when they see it” is a real answer, and it is always available.
• Never name a time without checking. Anything else sells a slot twice.
• Never say somebody is booked when only a request was taken. This is the one that produces a customer standing in your doorway.
• Never say whether a symptom is normal. A reaction to a treatment, a wound that looks wrong, a smell of gas — “that needs someone to look at it” is the correct answer, and it is not a failure of the agent.

The refusals that are specific to your trade

A general list is not enough, because the dangerous questions are different in every trade. A PMU agent must never say a pigment reaction sounds normal. A tattoo agent must never say parental consent makes a minor's tattoo fine, because that varies by jurisdiction and is not the receptionist's call. A spa's agent must never say a treatment is safe in pregnancy. A plumber's agent must never tell a caller who smells gas to wait for the morning.

None of those are tone. They are hard limits, and they belong in the brief the agent runs on rather than in the hope that a well-behaved model will work it out.

The one that gets missed: knowing when to stop

The hardest thing to build is an agent that abandons its own flow. Somebody says the word “bleeding”, or “I can smell gas”, or something that means they are in danger — and the correct behaviour is to stop taking the booking entirely, say the one useful thing, and get a person involved. An agent optimised to complete the booking will keep trying to complete the booking. That is the failure that actually matters.