Every phone system vendor now sells something called an AI receptionist. The demos all sound the same; the products are not. The gap between a good one and a bad one is the gap between a teammate and a robot that embarrasses your business on live calls. So here is a concrete bar - what the technology should actually do for a real business, and the lines it should never cross.
What it should do
Answer instantly, every time. The entire economic case is that intent decays in seconds. That means every call - after hours, holidays, and the overflow ring your office cannot pick up at 2 PM on a Tuesday - answered on the first ring, not a fallback that engages after the damage is done.
Know who is calling. A receptionist that greets a ten-year client like a stranger is worse than voicemail. A real one matches the caller against your client records and starts with context: the open appointment, the upcoming visit, the balance due.
Answer from your data, not the internet. Hours, service area, what you do and do not offer, the prices you chose to publish - all of it from the business profile you configured. And when it does not know, the only correct answer is 'let me take a message,' not a confident guess.
Book real appointments. Reading your actual calendar availability, offering real slots, holding the slot, and sending a confirmation text. 'Someone will call you back to schedule' is a message-taking service with better marketing.
Take structured messages and leave a paper trail. Name, callback number, reason, urgency - captured as fields, not a blob of transcription. And every conversation should land on the client's thread as a transcript with a recording, so a human can audit exactly what was said.
What it should not do
It should not pretend to be human. Callers forgive an assistant that is honest and effective. They do not forgive discovering they were deceived. Disclosure costs nothing and protects trust - and in a growing number of places, it is the law.
It should not improvise on money, medicine, or law. Quoting prices you never configured, estimating a repair sight unseen, or offering anything resembling medical or legal advice is how an AI turns into a liability. A hallucinated quote is a policy your business now has to honor or walk back.
It should not trap callers. 'Representative' must always work. The moment a caller wants a human, the AI's only job is a graceful handoff - transfer if someone is available, a structured message if not. An AI that argues for one more turn is optimizing for the vendor's metrics, not your customer.
And it should not handle emergencies alone. Gas leak, chest pain, water pouring through a ceiling - these need immediate escalation rules to a human or emergency guidance, configured before launch, not discovered after.
The handoff is the product
The difference between a gimmick and a teammate is what happens at the edges of its competence. A great AI receptionist hands a live transfer to a human with the context attached, turns an unanswered transfer into a structured message, and turns a promised callback into an actual task on an actual person's list. When you evaluate one, spend most of your time on the edges. The happy path is easy; every vendor has that demo.
How to evaluate one before you trust it
Call it yourself, ten times, and be difficult. Mumble. Interrupt it mid-sentence. Ask for a price you never configured. Ask for a human immediately. Describe an emergency. A production-ready system passes the hard calls, not the scripted ones.
Then read every transcript for the first two weeks and verify that booked appointments actually landed on the calendar. The transcripts will teach you what your callers actually ask - which is useful far beyond the AI.
The bar
An AI receptionist should behave like your best front-desk person on their calmest day: perfect memory, zero hold music, no sick days - and the judgment to get a human the moment it matters. Anything less answering your phone is not automation. It is a liability wearing a headset.
Frequently asked questions
A good one answers every call instantly - after hours, holidays, and midday overflow - recognizes returning callers from your client records, and answers questions from the business profile you configured. It books real appointments against your real calendar and captures structured messages with a transcript and recording a human can audit. When it does not know something, it takes a message instead of guessing.
Only if it reads your actual calendar availability, offers real slots, holds the slot, and sends a confirmation text. A system that promises 'someone will call you back to schedule' is a message-taking service, not a booking one. Verify this during a trial by checking that test bookings actually land on the calendar.
They should always disclose, and in a growing number of places it is legally required. Callers forgive an assistant that is honest and effective, but they do not forgive discovering they were deceived. Disclosure costs nothing and protects trust.
Call it yourself at least ten times and be difficult: mumble, interrupt it mid-sentence, ask for a price you never configured, ask for a human immediately, and describe an emergency. A production-ready system passes the hard calls, not the scripted demo. Then read every transcript for the first two weeks to confirm it behaves the way it did in testing.