Summarize with
Share

Ask an owner what worries them about an AI answering their phone and almost nobody says the voice. They say this: it books something that is not real, and I find out at eight on a Saturday morning, parked outside a house where nobody ordered a plumber. Half a day gone, a truck out, fuel burned, and a tech who could have been on a paying job. That worry is correct, and it is the reason the safe default is not booking. A phone system should capture the call and text you. It should only write to your calendar when the caller has said yes out loud to a specific address and a specific time.

01The Saturday you cannot get back

A missed call and a phantom booking are not the same size of problem, and owners are right to weigh them differently.

When you miss a call, you lose the chance at one job. Bad, and I have written enough about what that actually costs using your own average ticket rather than somebody’s benchmark. But the loss is bounded. The customer went elsewhere and your Saturday stayed yours.

A booking that is not real costs you three things at once. The slot, which you turned down other work to protect. The drive, which nobody pays for. And the customer you moved to Tuesday to make room, who now thinks you are disorganized. One phantom job on a Saturday can cost more than four missed calls on a Wednesday, and it costs it in a way you feel personally, standing in a driveway at eight in the morning.

The failure modes are boring, which is why they happen. The system heard 1450 when the caller said 1415. The caller was ringing round for prices and never intended to commit. The person on the phone was a tenant who cannot authorize work on the building. Or they said “maybe Saturday” and something wrote that down as Saturday.

None of those need a dramatic AI failure. They are the ordinary noise of people talking on phones while doing something else, and any system that turns that noise straight into calendar entries will produce phantom jobs at a steady rate.

02What has to be true before a job is real

Five things, and a booking that is missing any of them is a guess wearing a calendar entry’s clothes.

A clipboard with a job sheet showing an address and a time slot beside a set of van keys

The caller has to be someone who can authorize the work. Tenants, adult kids ringing about a parent’s house, and office staff at a commercial site all call in good faith and cannot all say yes to a bill.

The address has to be right and reachable, including the unit number, the gate code, and which of the two houses with the same number on that street it actually is.

You need a rough idea of what is wrong, in the caller’s own words, because “water coming through the ceiling” and “tap drips a bit” need different vans and different amounts of time.

Somebody has to be able to let you in. A confirmed appointment at a house nobody will be at on the day is a phantom job with extra steps.

And the time has to be one they actually agreed to, not one they floated. “Probably Saturday” is not a booking. It is a message.

Any decent phone answering setup can collect all five. The question is what it does when it only gets three, and that is where the good ones and the reckless ones separate. I went through the shape of a call that gathers these properly in what a good call to a new caller sounds like.

03A message beats a calendar entry

The safe mode is simple. The AI picks up on the second ring, gets the caller talking, collects what it can, and sends you a text. You read it between jobs and you decide.

What lands on your phone should be short enough to read at a traffic light. Name and number. Address as the caller gave it. What is happening, in their words, not summarized into something tidy. How urgent they say it is. When they said they are free. And a link to the recording, so if anything reads oddly you can hear the thirty seconds that matter.

That is thirty seconds of your attention instead of half a Saturday. And the decision stays with the person who knows whether you have a slot, whether that street is worth the drive, and whether the last job on that block went sideways.

Speed still matters here, and I do not want to talk anybody out of answering fast. Harvard Business Review timed 2,241 US companies responding to inbound inquiries and found the average first response took 42 hours. Only 37% got back within an hour. 23% never responded at all. The MIT work on response time found a five minute reply was about 100 times more likely to connect with the person than a thirty minute one. Answer in seconds. Just do not confuse answering instantly with committing instantly. One is free, the other spends a truck.

04What the confirmation actually sounds like

There is a version of booking that is safe, and it hinges on one habit: read it back and make the caller say yes.

A phone screen showing a short text message with a caller name and address

It sounds like this. “So that’s 1415 Oak Street, apartment 3B, Saturday morning between eight and ten, and it’s the water heater making a banging noise. Is that right?” The caller either confirms it or corrects it. If they correct it, you read it back again. It costs eight seconds and it catches almost everything a phone line mangles.

Two more rules make the difference. If the caller hesitates, goes vague, or the line is bad enough that the address came through half heard, the system drops out of booking and into message mode. Nothing partial gets written anywhere. A message that says “caller wants Saturday, address unclear, hear the recording” is useful. A calendar entry with a wrong address is a trap you will fall into later.

And if the caller asks whether they are talking to a machine, the answer is yes. Every time, no hedging. An AI that dodges that question has already told you what else it will do when a conversation gets awkward.

The compliance side is calmer than most owners expect, as long as the call is one the customer started. In February 2024 the FCC clarified that AI generated and artificial voices fall under the TCPA, and that outbound calls using them generally need the called party’s prior express consent. Inbound calls do not have that problem, because the person dialled you. That is one of the reasons a phone answering system should stay inbound and leave your customer list alone. Nothing about the number itself changes either, which I covered in what happens to your phone number.

05The jobs an AI should never touch

Some calls should not reach a booking question at all, and the boundary needs to be absolute rather than clever.

A gas smell, chest pain, or a sparking panel goes straight to a human number. Immediately, no qualifying questions, no message, no “someone will get back to you.” If a system tries to be helpful on that call, it is not saving you a dispatch, it is building you a lawsuit.

It should never diagnose. A caller describing a noise wants to know what is wrong, and the correct answer is that a person will look at it. A guess made on the phone at eleven in the morning becomes an argument in a driveway on Thursday, and you are the one standing in the driveway.

It should never quote a price. Not a range, not a “usually around.” Every trade has jobs that look identical on the phone and differ by a thousand dollars once the panel comes off.

Past those three, there is a softer list where a message is nearly always the better output. Insurance and restoration work, where scope is unknown until somebody stands in the room. Commercial jobs that need a purchase order. Anything in the week after a storm, when the job is bigger than the caller thinks it is. And any caller who cannot tell you what is actually happening, which is not their fault and is exactly why you send a person. If you want the wider version of what this technology does well and where it stops, I wrote it up in can an AI actually call and book your leads.

06How to test a vendor on this in ten minutes

You do not have to take anybody’s word for it, including mine. Phone the vendor’s demo line yourself and try to break it.

A contractor sitting in a truck cab listening to a call recording on a phone

Mumble the address. Say a street number quickly and slightly unclearly and see whether it reads it back or just accepts it. Then give it a soft date. “Maybe Saturday, or Sunday, I’m not sure yet.” A careful system turns that into a message. A careless one turns it into an appointment.

Then ask the vendor three questions. What share of your calls end as a booking versus a message? If the answer is that almost everything gets booked, that is not a strength, that is the phantom job rate they have not measured yet. Can I hear a recording of a call where the caller was unclear? A vendor with nothing to hide will play you a messy one. And who eats the cost when it books something that is not real? Most contracts say nobody does. That may be fine, but you should know it before you sign rather than at eight on a Saturday.

Last thing, and it is the one owners forget to check. Make sure you can hear every recording and read every transcript yourself, without asking. A system you can audit is one you will trust in six weeks. A dashboard that only shows you totals is one you will stop believing the first time a job goes wrong.

Frequently asked questions

So does it ever book anything, or does it just take messages?

It books when the picture is complete and the caller confirms it out loud. Repeat customer, address you already have on file, a job type you do every week, a slot they said yes to after hearing it read back. That covers a decent share of calls. The rest come to you as messages, which is the correct outcome rather than a failure.

My customers want to book right there. Won’t a message annoy them?

Some do want a firm time, and for straightforward jobs they get one. For the rest, what people actually want is to know they have been heard and that somebody is coming back to them. A call that ends with “I’ve got all that, the owner will text you within the hour with a time” lands far better than voicemail, and better than a booked slot that gets moved the next morning because it never fit your schedule.

Doesn’t this just move the work to me, reading texts all day?

You are reading a short message and replying yes or no, mostly while you would otherwise be checking your phone anyway. Compare that to the current version, which is listening to voicemails at seven in the evening, calling people back who have already hired somebody else, and losing the ones who never left a message. The reading takes seconds. The calling back took your evening.

What stops a real booking from being forgotten by the customer?

Confirm it in writing while it is fresh. The caller gets a text with the address and the window straight after the call, so a wrong address surfaces the same day instead of on the morning. Keep the recording attached to the job so anybody dispatching can hear what the customer actually said rather than reading somebody’s summary of it.

Are the recordings a problem legally?

Recording rules vary by state, and some require every party on the call to consent rather than just one. The workable answer is to disclose it at the start of the call, keep the recordings where you can produce them, and check your own state’s rule rather than assuming a vendor has handled it. It is a short conversation with whoever does your paperwork, and worth having before you turn recording on.

The bottom line

The thing that makes an answering system safe is not how good the voice sounds. It is what it does when the information is incomplete, which on real phone calls is most of the time. A system that always books will eventually put you in a driveway on a Saturday outside a house that never called. A system that captures the call, reads back anything it writes down, and hands you the doubtful ones as a message keeps the decision where it belongs, with the person who owns the truck.

We book appointments for roofing companies: we find homeowners looking for roof repair or replacement, answer their replies and book them into your system. If that would help your team, see how it works at rushdial.com/roofing.