The Burst Pipe Call That Went to Voicemail at 11 p.m.

23:04, a burst pipe in Graz

Helga is 74 and has lived in the same flat in Graz for thirty years. At 23:04 on a Thursday in January, a pipe in her utility room bursts, and water starts running across the tiles towards the hallway. She calls the number on the sticker on her boiler, the firm that has serviced it every autumn for twelve years.

It rings five times. Then a recorded voice tells her the office is closed, and asks her to leave a message. She does, and then she looks for another number.

Lukas, the technician on call that night, is kneeling in front of a boiler in a care home across town, where the heating failed at nine. His phone is in the van. He sees the missed call at 23:40 and rings back. By then another company's van is outside Helga's building.

The firm is invented, a heating and plumbing company with nine technicians, and so are Helga and Lukas. The voicemail is not. Most trade businesses that run an on-call rota from one mobile will know this night, and they usually hear about it from the customer, days later.

Nine technicians and one number

During the day, calls go to Andrea, the office manager. When she is on another line, they ring through to whichever technician is next on the list, usually on a ladder or under a boiler. After five, the company number forwards to the on-call technician's mobile, every night, all night.

The calls are a mix nobody can plan. Most are routine: a boiler service, a dripping tap, a quote for a new bathroom. A few are real emergencies, a burst pipe or a smell of gas. And at two in the morning, some are people who want an appointment in March and assume the number reaches an office. The on-call technician answers all of them, because he cannot tell which is which until he picks up. When he is already on a job, like that Thursday, nobody answers at all.

Thomas, the owner, hears about Helga on Monday, when her letter cancelling the service contract arrives. What he wants is not a generic AI receptionist. It is a rule: at night, only real emergencies reach a person, everything else is booked properly, and anyone who wants a human can reach one. Many of his customers are Helga's age and will not talk to a machine for long, so the way to a person has to be short and obvious.

A protocol before a voice

Thomas and I talk on a video call from Pristina that week. I ask for the on-call rota and a week of call logs from the phone system, by hour. The logs are ordinary and telling. Most evening calls are routine. Two in the week were real emergencies. Five went unanswered, all while the on-call technician was already on a job.

The voice is not the first thing I build. The first thing is a table that Thomas, Lukas and the other technicians write with me, one row per situation, saying exactly what happens. A caller who asks for a person is transferred at once, with no further questions. A smell of gas gets a fixed instruction: leave the building, do not use switches or open flames, and call the gas emergency number, 128 in Austria, from outside, while the on-call technician is alerted by call and SMS. A leak gets: close the main water valve and keep away from electrics near the water, then a transfer to the technician. No heating means checking the power, the water pressure and any error code on the boiler display. If an infant or an elderly or sick person is in the home in winter, the technician is called. Otherwise it is the first slot next morning.

The model's job is to recognize which row a call belongs to. What happens next is fixed in code. The agent reads safety instructions word for word, as the protocol has them. It never improvises advice, never diagnoses a fault, and never estimates what a repair will cost.

One sentence is fixed before any other: the opening. It says this is the company's automated assistant, that the call is recorded, and that the caller can ask for a person at any time. Article 50 of the EU AI Act covers voice agents like this, and the Commission's guidelines say the disclosure has to come at the first interaction. The Article 50 guide has the rest.

One Thursday night, twice

To test the protocol, we replay the kind of night the call logs describe, call by call: once as it runs today, once with the agent answering.

Today, Lukas's phone rings for all six. With the agent, it rings for three, and each of them is a call he would want to take. The other three are handled properly without him, and the one that went to voicemail reaches a person.

The man who asked for a person

A public phone number is not a place to experiment, so the agent starts small. For the first two weeks it only takes the daytime calls Andrea cannot pick up. It books nothing and takes messages with a summary. Every afternoon, Andrea listens to the recordings and marks each mistake.

On the fourth day she marks the one that matters most. An older caller, halfway through giving his address, says he would rather speak to a person. The first version answers politely: of course, but may it first take the house number, so the colleague has it? He hangs up. The model had understood him perfectly. It had simply decided that one more question was helpful.

So that decision leaves the model. Recognizing the request is still its job, but what follows is code: a transfer, at once, with no further question. After hours, a non-emergency caller who wants a person is offered a callback at a set time instead of waking Lukas, and the promise lands on Andrea's list for eight o'clock. Every request is counted, because it tells Thomas how his customers feel about the assistant.

The recordings teach one more rule. Styrian German, a caller standing in a wet cellar, a signal that fades: the agent reads back every critical detail, and after two failed attempts to understand, it transfers or offers a callback rather than guessing. Only then do the evenings follow, with bookings as requests Andrea confirms in the morning. Direct booking comes last, once her confirmations stop changing anything. Emergency handling never gets more freedom than the protocol gives it.

What Lukas hears before the caller does

Underneath, the pieces are plain. After five, the company number forwards to a Twilio number. A real-time voice model holds the conversation, full duplex, so callers can interrupt it the way they would interrupt a person, in German or in the other languages the firm's customers speak.

When a call is an emergency, the transfer is warm. Lukas's phone rings, and before the caller is connected he hears two sentences: what is happening, what the caller has already done, and where. If he does not answer within 30 seconds, the backup on the rota is called, and the caller is told what is happening instead of listening to ringing.

The address gets more care than anything else in the call. A Hauptstraße exists in many of the villages around Graz, so every address is read back with its town, geocoded, and checked against the service area. Staircase and door numbers are kept as access details, and Lukas sees the caller's own words next to the geocoded result. An address that does not match is never booked blind.

The firm keeps its schedule in a shared Google Calendar, so routine jobs are written there as holds. A slot is held for ten minutes while the caller confirms, so two callers cannot take the same one, and if the calendar cannot be reached mid-call, the caller gets a booking request and a callback time rather than a promise the system cannot keep. Every caller gets an SMS summary, and every call becomes a job entry with its recording and transcript. The whole design is in the blueprint for an after-hours call agent.

08:00, Andrea's list

On another cold Thursday, a pipe splits in a terraced house outside Graz at 22:50. In the first sentence, the caller hears that she is talking to the company's automated assistant. Within a minute she has closed the main valve. Lukas's phone rings, he hears two sentences, and he is on the line before she has finished mopping.

At eight, Andrea opens her list: the night's calls in order, each with a summary, a recording and what happened. Two booking holds to confirm, one of which she moves because it needs a part that is not in stock. A callback promised for half past eight to a man who wanted a person. A request for a bathroom quote. One emergency call-out, created by Lukas, with the transcript attached.

Lukas was woken once that night, for the pipe. Nobody quoted a price on the phone, and nobody who asked for a person had to ask twice.

If your number forwards to a mobile tonight

For simple booking, buy. Smith.ai, Rosie and Goodcall answer calls, take messages and book appointments for small businesses at a monthly price. In the US, ServiceTitan sells an AI voice agent to its customers, and Avoca is built for home services on the US field service platforms. If all you need is fewer missed calls, one of them is enough.

A build wins when the rules are the product: a protocol your technicians wrote, rotas with backups, transfers that brief the technician first, bookings in the European trades software or calendar you actually run, and callers in more than one language. It is usually a single-purpose AI agent build, priced mostly by your scheduling system and how detailed the protocol is. The full blueprint covers the failure modes and the rollout, my voice agent build guide covers the technical side, and build or buy for an AI receptionist covers when a product is the better call.

Frequently asked questions