How does an AI receptionist actually work when you're on a job?

An AI receptionist is a voice agent that answers your phone, listens, works out what the caller needs, and books the job into your calendar during the call. Your number forwards to it: every call, or only the ones you don't pick up. The caller talks and it talks back; when it can book the job, the appointment exists by the time they hang up. What it can't book reaches you as a message, with the details already taken.

What follows is the mechanism underneath that, step by step, and the parts nobody advertises get their turn too.

What happens when the agent picks up?

Once your number forwards the call - straight away, or after your own phone has rung out - the agent should pick up within a few rings and greet the caller in your business name, saying it is an AI assistant and that the call is recorded. From there the call runs as a loop that repeats every time the caller stops talking.

Step	What it does	Typical time
Speech to text	Turns the caller's words into text	200-400 ms
The model decides	Works out the next question or action	300-1,000 ms
Text to speech	Starts speaking the reply in the agent's voice	150-500 ms
Network	Carries audio back and forth between the services	100-300 ms
Processing	Queues and passes data between them	50-200 ms
Turn detection	Waits out a silence to decide the caller has finished	200-800 ms

Those are the typical ranges in Hamming's January 2026 latency breakdown. Add them up and a typical agent takes 1 to 3.2 seconds end to end. Hamming's optimised ranges come to 670 to 1,450 milliseconds.

Even the quick end of that is slow by human standards. In Stivers and colleagues' study of ten languages in PNAS, the average gap between one person finishing and the next starting was 208 milliseconds, and the most common gap was no gap at all. So even a well-tuned agent answers a beat later than a person usually would.

The question is when that beat starts to cost you, and there is a number for it. Roberts and Francis played 380 listeners the same "sure" in reply to a request after silences of 200 to 1,200 milliseconds and asked how willing the person answering sounded. Ratings dropped off at 600 milliseconds, and the fall became statistically significant between 700 and 800. Those were requests between friends, so for a business on the phone read it as a guide. Still, an optimised agent's quickest reply, 670 milliseconds, already sits past the point where those ratings start to drop, and Hamming puts the industry median at 1.4 to 1.7 seconds.

It is also the easiest thing to test before you buy. Ring a vendor's demo line if it has one and talk over it mid-sentence. Listen for two things: how long it pauses before answering and whether it stops when you interrupt.

How does it work out what the caller actually needs?

It works from what the caller means. "There's water coming through the kitchen ceiling" and "I think a pipe's gone" describe the same job: a leak, probably urgent. A phone menu needs the caller to know which of your options their problem is. An agent works it out from what they say.

Then it takes down what the job needs:

• Who they are. Their name, spelled back letter by letter before it goes on a new record, and the number they're ringing from, which comes from caller ID.
• Where the job is. The postcode and house number, with the street read back.
• What is wrong, in their own words.
• How urgent it is, noted on the booking.

What a well-designed agent will not do is promise you'll be there in the hour. It has no idea where your van is, so it only offers times your calendar shows free.

What happens if the caller says they can smell gas?

It stops trying to book anything. A gas smell, a carbon monoxide alarm, or a headache that lifts when someone steps outside all end the booking. SwiftGuard's agent says plainly that it cannot handle an emergency, tells the caller to leave the building and, once they're outside, to ring their national gas emergency line or the emergency services. Fire, or someone unconscious or in danger, means hanging up and calling 112.

Ask any vendor what their agent does with that call. If the answer involves offering a slot on Thursday, keep walking.

How does it get the job into your calendar?

The times it offers come from your real calendar, and the slot the caller picks becomes an event in it, with the job as the title and the address as the location, so it sits on the phone in your pocket like anything else you've booked.

Two parts of that are harder than they look.

The address. A street name is easy to mishear on a phone line, and a misheard one still points at a real house that belongs to someone else. So the safer design asks for the postcode and house number, reads the street and town back, and waits for a yes before it books. It also has to tell a match to the house from a match to the street: SwiftGuard's address lookup flags anything short of the exact house as unconfirmed instead of quietly treating it as certain.

The returning customer. Someone you've worked for before shouldn't have to give you their whole address again. SwiftGuard checks a returning caller before it opens their record - their postcode and house number, or a code texted to the phone they're ringing from - and then asks one question: is this for the address we already have?

A voicemail is a job you still have to chase. A booked slot is a job that is already yours.

What do you get after they hang up?

A record of the call, and the appointment if it booked one. The record holds the recording, the transcript and, unless the summary step fails, a few sentences on what the caller wanted, so you can read it between jobs instead of listening back to someone describing a smell.

The calls it can't book arrive as a message instead. A caller on a landline can't receive the code it texts to check a new customer, and a withheld number can't be texted at all; those, and anything else it can't sort out, reach you with the details taken, for you to ring back.

Voicemail was never much of a callback list. On Invoca's call data, fewer than 3% of callers pushed to voicemail leave a message (why callers skip the beep). How many of your own calls go unanswered in the first place is something you can count on your own phone this week. How this compares with a human answering service is a post of its own.

What do you have to set up before any of this works?

Three things, and only the first one is fiddly.

1. Forwarding. Two choices, and you pick one: every call goes to the agent, or only the calls your own phone doesn't answer - rung out, engaged, or switched off. You first confirm your business number with a code SwiftGuard texts to it, so for now it has to be a number that can receive a text, which in practice means a mobile: most landlines can't receive one, so a landline business number usually can't be confirmed yet. Then you dial two short codes from the business handset: one clears any old forwarding, the other switches yours on. SwiftGuard's forwarding setup is built for Dutch lines only for now, and until your SwiftGuard number is assigned, the code on your setup screen is a placeholder that forwards nowhere. The step-by-step is here.
2. Your calendar. It needs to see when you're free before it can offer a slot at all. Connecting Google Calendar takes about a minute, and Google Calendar is the only one SwiftGuard connects to today.
3. Your hours. The agent only offers times inside your business hours and far enough ahead for you to get there. New accounts start at Monday to Friday, 08:00 to 18:00, with a two-hour lead time, and two hours is the wrong setting for emergency work: turn it down if you take call-outs.

Where does an AI receptionist fall down?

On the calls that were never routine. Be honest with yourself about these before you switch anything over:

• Bad lines and heavy background noise. A caller on a motorway hard shoulder is hard for a person to hear too, but a person can improvise more.
• A caller who wants to argue, or who is upset and wants a person to hear it. The agent should offer to take a message for you to ring back rather than grind on.
• A price on the phone, sight unseen. SwiftGuard's agent won't quote or estimate one. It takes the problem down in the caller's words, and you confirm costs when you ring back.
• Anything genuinely unusual. The right failure is a proper message for you, with nothing invented to fill the gap.

In the EU there is also a rule to know about. Article 50 of the AI Act requires the providers of AI systems that interact directly with people to build them so those people are told they are dealing with an AI, unless that would be obvious anyway. It has applied since 2 August 2026. SwiftGuard's agent says so in its opening line.

Telling people is only part of it. In Gartner's survey of 3,566 B2B and B2C customers, run in February and March 2026, 87% said it is essential for a company using generative AI for customer service to offer a way to reach a human agent. SwiftGuard's way to a person is a message: the agent takes one, and you ring the caller back. It cannot put them through to you live. What customers actually do when AI answers has the rest of the numbers.

Everything above describes a machine doing the dull few minutes at the front of a job: who, where, what's broken, when can you come. That was never the skilled part. It was just the part that had to happen while you were under a sink with your phone in the van.

That's the job SwiftGuard does. If you want to see what it books and what it hands back, the plans start at €69 a month, ex VAT.