An AI phone answering service answers your business line with a computer that talks. Not a menu, not a recording, not "press one for bookings" — a conversation, where the caller says what they want in their own words and something on the other end works out what to do about it.
That sentence has been true in marketing copy for about a decade and genuinely true for about two years. It is worth understanding which parts are now real, because the gap between the demo and the Tuesday afternoon is where the money is.
What actually happens on the call
Four things, in about a second and a half, repeatedly, for as long as the call lasts.
The caller's speech becomes text. A recogniser turns audio into words. This is the stage that fails most often and most invisibly, because phone audio is narrowband, compressed, and full of a hairdryer. "Two PM" comes back as "2 AM". A service name comes back as a person's name.
A language model decides what to do. It reads the text, the conversation so far, and what it knows about your business — your services, their durations, who can do them, your hours — and picks an action.
Tools do the work. Checking availability, holding a slot, writing the booking. This is the part that separates a system that books from a system that chats: it has to reach into the same calendar your front desk uses, not a copy of it.
Text becomes speech. The reply is spoken back, ideally fast enough that the caller does not think the line has dropped.
We go through this in more depth in how voice AI understands a phone call, including the part nobody talks about, which is knowing when the caller has finished speaking.
What it handles well
Bookings, reschedules and cancellations. The core of it. A caller asks for Thursday afternoon, it offers real times from a real calendar, they pick one, it is in the book before they hang up.
The same six questions, forever. Are you open Sunday. Do you take walk-ins. How much is a gel fill. Where do I park. Every business has a short list of questions that consume an enormous amount of front-desk attention, and a system that answers them accurately buys back more time than the bookings do.
Volume at the worst moment. Three calls at once during a rush is where human coverage breaks — not because anyone is bad at their job, but because there is one of them and three of you. This is the failure mode an AI receptionist genuinely solves, and it is worth more to a restaurant at seven on a Friday than the after-hours coverage is.
Recognising a returning caller. Knowing that this number belongs to someone who was in six weeks ago, for the thing they had done, with the person who did it.
Where it falls short, and you should hear this before you buy
It mishears. Every system does. What separates a usable one from a liability is whether a mishearing becomes a question or a confident mistake. We have written up five calls to make before you trust an AI receptionist — all five are failures we found in our own system before a customer did.
It should not handle everything. An allergy question, a complaint, an unusual medical request: these need a person, and a system that improvises an answer is worse than one that admits it does not know. The measure of a good one is how gracefully it hands over.
It speaks one language well. Ours speaks English today. If your callers need another, ask directly rather than accepting a roadmap answer.
What it costs, and how the pricing models differ
This is where the category is genuinely confusing, because three pricing models are all described with the same words.
Per minute. Inherited from human answering services. You are billed for talk time, so your busiest month is your most expensive month.
Per call or per booking. Cleaner, but it means a caller who rings back twice costs you three times.
Flat, by the size of your business. What we do: priced by team size, with the calls and texts uncapped, so the bill does not move when December arrives. Ours starts at ninety-nine dollars a month for up to five staff and the pricing page has the rest.
None of the three is inherently right. What matters is knowing which one you are signing, and modelling it against a busy month rather than an average one.
How to tell a real one from a phone tree with better marketing
Three questions, and you can answer all of them from a single call to the vendor's demo line.
Ask for a vague time. "Tomorrow afternoon." A phone tree cannot cope. A real system offers you two or three genuinely spread-out options.
Interrupt it mid-sentence. A real system stops within about half a second and deals with what you said. A scripted one finishes its sentence.
Ask it something it cannot know. A price for something not on the menu. The honest answer is that it does not know and will have someone call you. Anything else is a system that will eventually invent a price for a real customer.
Getting one onto your line
You do not need a new number and you should not accept a setup that requires one. Conditional call forwarding hands over only the calls you do not pick up, which means you can start with nothing but the overflow and the after-hours calls, and leave your published number untouched. The call forwarding guide has the codes by carrier.
Start there. Forward the calls that are currently reaching voicemail, and leave everything else alone. If it works, widen it. If it does not, you have lost nothing that was not already going unanswered.
Ours answers on (941) 430-4049 if you want to try the three questions above on it right now.