What an AI receptionist actually does on a business call (and what it should not)

The first thirty seconds: answer, greet, identify

Every call an AI receptionist handles starts the same way. The line is answered on the first or second ring, the caller hears a greeting that names the business, and the receptionist asks an open question: how can I help today? The caller answers in their own words, and the receptionist works out which of its defined jobs the request belongs to.

That last step is the one people underestimate. A caller rarely says “I would like to book an appointment.” They say “my furnace is making a noise again” or “is Dr. Tremblay in on Thursdays?” The receptionist has to map that sentence onto a job — a service booking, a schedule question — and then confirm it back before doing anything: “It sounds like you need a technician to come out. Is that right?”

That confirmation loop is what separates a receptionist from a voice menu. A menu forces the caller to guess which digit matches their problem. A receptionist listens, proposes an interpretation, and lets the caller correct it in plain language.

The jobs an AI receptionist does well

The useful list is shorter than the marketing suggests. An AI receptionist is good at bounded, repeatable tasks where the information it needs is either in the caller’s head or in a system it has been connected to.

• Booking, moving or cancelling an appointment against a calendar it can read and write, including asking for the details the booking needs (name, callback number, reason, preferred time).
• Taking a message with a written transcript and delivering it to the right person or shared inbox, rather than leaving a voicemail nobody plays back.
• Answering routine questions from a script you control: hours, address, parking, what to bring, whether a service is offered, how to reach a department.
• Transferring to a named person, an extension or a queue when the caller asks for one, and announcing who is calling before the transfer completes.
• Collecting structured information before a human picks up, so the person who answers already knows the caller’s name, account or job address.
• Handling the same call the same way at 7 a.m., during lunch and at 9 p.m., without a change in tone.

The jobs it should hand off, every time

The second list matters more than the first, because getting it wrong costs trust. A well-configured AI receptionist is deliberately limited. It has a rule for each of the following situations, and the rule is always some form of “get a person.”

Anything that sounds like an emergency. A caller who says someone is hurt, a pipe has burst, or they smell gas should not be walked through a booking script. The receptionist should say clearly that it is transferring them, and transfer immediately to whoever is on call — or, where the business has decided this, tell the caller to hang up and dial 9-1-1.

Anything requiring judgement the business has not written down. A receptionist can say a clinic offers physiotherapy. It cannot decide whether a caller’s symptoms mean they need it today. It can quote a price you have given it for a standard service. It cannot estimate a job it has not been given a price for.

Anything where the caller is upset, confused or repeatedly not understood. Two failed attempts to identify the request is a reasonable limit. After that, the receptionist should stop trying and offer a person or a callback rather than a third rephrasing of the same question.

What the receptionist should never say

A short list of forbidden behaviours is worth writing into the configuration and reviewing in transcripts. The receptionist should not claim to be a person if asked directly. It should not promise a specific staff member will call back at a specific time unless it has checked a calendar that confirms it. It should not invent an answer to a question outside its script — “I don’t have that information, but I can take a message for someone who does” is the right response.

It should also not collect information it does not need. If a message only needs a name and callback number, asking for an email address, a date of birth and an account number adds friction and risk for no benefit. Configure the receptionist to ask for the minimum each job requires.

Finally, it should not hold a caller who has asked for a human. Some callers will say “I want to talk to someone” in the first five seconds. That request is honoured, not negotiated.

Drawing the line for your own business

The line between “handle” and “hand off” is not the same for a dental clinic, a plumbing company and an accounting firm. A useful way to draw it is to take a week of real calls — the ones your staff actually answered — and sort each one into three piles: could have been fully handled by a script, needed a person but could have been prepared by a script, and needed a person from the first second.

The first pile becomes the receptionist’s jobs. The second becomes its intake work: collect the details, then transfer. The third becomes its escalation rules. Most businesses find the first two piles are larger than they expected and the third is small but non-negotiable.

Review the line after the first month. Transcripts will show calls that were handed off when they could have been handled, and — more importantly — calls that were handled when they should have been handed off. Adjust the script, not the caller.

Common questions

Two questions come up in nearly every planning conversation about an AI receptionist.