Nobody wants to talk to a robot. They want somebody to pick up.

The first time we tested a voice bot on our own phone line, it told a caller we did flat roofs. We do not do roofs. We do not do anything with roofs. It said it with total conviction, booked a site visit, and put it in the calendar for the following Tuesday at eight.

I remember listening back to the recording and thinking, well, that is that. Not because the voice was bad. The voice was fine. It was the confidence that killed it. A human receptionist who does not know something says “let me find out and call you back,” and that sentence is doing an enormous amount of work. It protects the caller, it protects you, and it buys time. Our robot had no idea that sentence existed.

So we shelved it. For about two years we told people who asked that AI phone answering was not ready, which was true, and that they should get a human answering service instead, which was mostly true and also cost them three hundred a month for someone in another state reading their business name off a card.

What actually changed

Two things, and neither of them is the thing the demos focus on.

The first is latency. In 2022 there was a gap of about a second and a half between you finishing a sentence and the thing responding. That gap is where the illusion dies. Not because it feels robotic, but because a second and a half of silence on a phone call means something specific to a human being: it means the other person is either confused or annoyed. So callers would repeat themselves, the bot would start talking over them, and the whole call would collapse into two people apologising. Now the gap is short enough that the call has a rhythm to it, and that single change did more for these systems than any improvement in the language model.

The second is grounding, which is a boring word for the only thing that matters. The system does not answer from general knowledge about plumbing. It answers from your price list, your service area, your call out fee, your hours, the four questions you always ask before you agree to look at a job, and the two kinds of work you turn down. When it does not have something, it is built to say so rather than to fill the silence. Turns out a model that says “I do not have that in front of me, let me have somebody call you back” is not a limitation. It is the entire product.

What a call sounds like now

Here is a shape we hear a lot. A late afternoon call, someone with water where water should not be, who has already tried two other numbers.

Notice what is not happening there. It is not pretending to be a person called Sarah. It says what it is in the first breath, every time, on every call, and we will not build it any other way. Partly because in a lot of places that disclosure is the law now, and partly because the one thing worse than a caller realising they are talking to a machine is a caller realising they were lied to for ninety seconds.

In practice almost nobody hangs up over it. What people hate is being told they are important and then held for nine minutes. Being told plainly that you are talking to an assistant, and then having the assistant sort your problem in under two minutes, turns out to be fine.

The question everybody asks second

The first question is always what it costs. The second is always, every single time: what happens when it gets something wrong.

It is a fair question and the answer is architectural, not aspirational. The system has a narrow band of things it is allowed to commit you to. Booking a slot, quoting a fixed call out, confirming whether an address is in your area, taking a message. Outside that band it does not improvise, it escalates. A question about a price it does not hold becomes a message with a flag on it. An angry caller becomes a message with a flag on it, marked urgent. A conversation that goes somewhere strange becomes a message with a flag on it and, if you want, a live transfer to your mobile.

What it is genuinely bad at

This is the part that gets left out of the sales page, so here it is.

• Emotionally loaded calls. A bereavement, a flooded home someone is standing in, a customer who is frightened about money. It is polite and it is useless. Those should ring straight through to a person, and we set them up to do exactly that where we can spot them.
• Anything involving judgement about your money. Discounts, payment plans, whether to waive a fee for a good customer. If you have not defined the rule, it should not invent one on your behalf.
• Multi property, multi trade quoting. If the answer to how much depends on six things somebody has to go and look at, the honest reply is that somebody will go and look at it.
• Disputes. An invoice argument does not want efficiency, it wants a person who can say sorry and mean it.
• Heavy accents on a bad line, still. Better than it was, not solved. It fails by asking you to repeat yourself, which is annoying but is at least the failure you want.

If most of your calls fall into those categories, an AI receptionist is not your problem to solve and we will tell you so on the first call. What we usually find instead is that seventy or eighty per cent of inbound calls to a service business are four or five repeated shapes. Are you in my area, what does it cost, can you come out, where is my invoice, can I move Thursday. That is the part worth handing over.

The honest arithmetic

This is not a sales machine that works while you sleep. Most businesses we look at get very few genuine three in the morning enquiries, and the ones they do get are usually emergencies that want a human.

The value sits in far more boring places. It is the call at ten to five while you are packing up. It is the two hours on Saturday morning when the people who cannot phone during the week finally get a chance to. It is the twenty minutes you are up a ladder with both hands busy. It is the fourth call in a row on a Tuesday, when the first three have already eaten the time you had set aside for paperwork. Every one of those is a real enquiry from a real person who is, at that moment, deciding whether to try the next name on the list.

We put a rough figure on that with you before we build anything, using your own numbers rather than ours. There is a calculator on the home page that does the same sum in about fifteen seconds, and it is deliberately conservative about it. Every missed call is not a lost job. Some of them call back. Some were tyre kickers. Argue with the numbers, that is what they are there for.

If you only do one thing after reading this

Call your own business line. Not from your mobile, from a number nobody there recognises, at a genuinely inconvenient hour. Half seven on a Thursday evening, or Sunday lunchtime.

Listen to exactly what a stranger hears. How many rings before something happens. Whether the voicemail greeting has your current prices in it, or a name of somebody who left. Whether it even records. Then ask yourself honestly what you would do, in their position, with a wet ceiling and a phone in your hand.

Most people who do that exercise stop asking us whether the AI sounds natural enough, because they have just heard what the alternative sounds like. If you want, we will do it with you and read the transcripts out loud, which is a slightly brutal fifteen minutes but a useful one. Details are on the contact section, and there is no presentation.