Blog · The basics
How an AI phone call actually works, from dial to summary
One call, followed from the moment you ask to the moment the summary turns up.
By the CosVoice team · · 5 minute read
Thursday, ten to three
Graham wants a table for four tomorrow night. The restaurant only takes bookings by phone, and he has a site visit in ten minutes. So he tells his assistant, Lloyd, and goes back to work.
What happens next is simpler than people expect. The assistant takes the instruction and dials the restaurant from its own phone number. It says who it is and why it's calling, has an ordinary conversation, and agrees a time inside the limits Graham gave. It reads the details back, hangs up, and writes down what happened. That's how an AI phone call works. The rest of this page is the same call, slowed down.
The instruction
Graham's message was one sentence: table for four, Friday, 7:30, and 8:15 is fine if they're full. He could have said it out loud by ringing his line, texted it, typed it into his dashboard, or asked the AI assistant he already uses.
That sentence is the brief, and everything in the call hangs off it. The window matters most. Without "8:15 is fine", any counter-offer has to come back to Graham as a question, and the table may be gone by the time he answers. With it, the assistant can say yes on the spot. (There's a whole article on writing a call brief if you want to get good at this.)
The dial
The call goes out from the line's own number, a normal number with an area code. Nothing special is needed at the restaurant's end. No app, no account. Their phone rings, same as ever.
Sometimes nobody picks up, or a machine does. Then the call is logged that way, as voicemail or not reached, and Graham hears about it instead of assuming he has a table. If there's a phone menu, the assistant works through it. If there's hold music, it waits.
The first ten seconds
A person answers. The assistant opens with one sentence that says who's calling and on whose behalf. Something like: "Hi, this is Lloyd, I'm Graham's assistant. I'm calling to book a table for tomorrow evening." It's short on purpose. The host is holding a phone in a noisy room and wants to know what this is about.
If the host asks whether they're talking to an AI, it says yes. It never claims to be a person. Then the call carries on, usually with "okay, what time were you after?"
The back and forth
This is the part people picture as clunky, a robot waiting for the beep before it speaks. It isn't like that. The voice is natural, and with the newer voices the host can cut in mid-sentence and the assistant stops and listens, the way a person would. (The classic voices are quicker to answer and crisper on a phone line, but they finish the sentence first.)
The host asks the usual things. What name? Graham. A number in case anything changes? The assistant gives the line's own number, not Graham's mobile. An email for the confirmation? The line has its own address, so it gives that. If the host mentions a policy, say the table is only held for fifteen minutes, the assistant acknowledges it and notes it down.
If you'd rather hear this than read about it, there's a recording of a table being booked on the demo page.
The counter-offer
"We're full at 7:30. I can do eight." This is where a call succeeds or fails, and where the brief earns its keep. Eight is inside Graham's window, so the assistant takes it.
Had the host offered 9:30, it would have asked once whether anything closer was possible. If not, it would ask about a waitlist or a hold and leave the decision to Graham.
And if the restaurant wants a card to secure the table? It doesn't have one to give. It never reads out a card number and never agrees to a charge. The call gets logged as needing the owner, and Graham sorts out the deposit himself.
The hang-up
Before it goes, the assistant reads the booking back. Friday, eight o'clock, four people, under Graham. This is dull and it matters. A misheard day, or a four written down as a two, gets caught while the host still has the book open.
Then thanks, goodbye, and the line drops. The whole call took about as long as it would have taken Graham. The difference is where Graham was while it happened.
What lands in the inbox
A written summary. It has the outcome in plain words (an alternative time confirmed, 8pm in place of 7:30) and the details: date, time, party size, the name, the fifteen-minute rule. The transcript is there too if he wants to check exactly what was said.
What we keep is the written call, not the audio. The content is sealed to the line, and calls never train models. The trust page has the longer version.
If Graham's own AI assistant is connected and can see his calendar, it reads the same summary and puts dinner in the calendar. If not, he does that himself in ten seconds. Either way nobody is relying on memory, which was the point of writing it down.
Common questions
How long does an AI phone call take?
About as long as the same call made by a person. A simple booking is a minute or two. A call with hold music takes as long as the hold does.
Does the business need special software to receive an AI call?
No. It's a normal phone call to their normal number. If they can answer the phone, they can take the call.
Is the call recorded?
You get a written summary and a transcript. We keep the written call, not the audio, and call content is never used to train models.
What if the person on the other end hangs up or says no?
Then that's what the summary says. Every call is logged with an honest outcome, including declined, not reached and voicemail, so you're never left guessing whether the thing got done.
Read next: What an AI caller can and can't do · Restaurant reservations by phone · How CosVoice works
Hear it for yourself.
Call our own AI agent and ask her anything, or get a number in your area code in about a minute.