The First Test Call to a Roofer's New AI Phone Line Hung Up on the Owner
Quick answer: A Maryland roofing contractor called his own new AI phone line about storm damage, and the call went wrong six ways, starting with the line hanging up on him when he asked about hours. Almost none of it was the AI being unintelligent. It was settings, wiring and instructions nobody had written down, and the worst problem was that nobody had called it before he did. Over the next four weeks we logged 52 problems on that line, and zero of them needed a smarter AI.
Six problems in one phone call
This summer, a roofer we work with called his own new AI answering line. He played a homeowner with storm damage, the call that matters most to his business, and then he told us what happened. One call turned up six problems:
1. He asked what the hours were, and it hung up on him. 2. He asked for a person. It didn't transfer him, and it didn't say it couldn't. 3. It never asked for his last name. 4. It never asked for his email. 5. He said storm damage, and it tried to steer him into the regular list of services instead of treating it as a storm call. 6. It talked too fast.
The seventh problem wasn't in the call. It was that the owner was the first person to test it. Nobody had picked up a phone and called that line before he did, and that one is on us.
October is when this matters. Wind season is starting, the calls pick up, and a lot of owners are deciding whether to trust an answering system before winter. So here is what that call actually taught us.
Most of it had nothing to do with the words
When an AI phone line misbehaves, the instinct is to rewrite the script, or to decide the AI just isn't smart enough. We had the same instinct. Our first guess on the hang-up was that the line didn't know the hours.
It did. The hours had been in its instructions the whole time. The hang-up came from a setting, not from a gap in what it knew. If we had only added the hours to the script, the same thing would have happened on the next ordinary question it didn't expect.
The rest broke down the same way. The transfer failed because nothing had been connected to pass a call to a person at all. The fast talking was a setting too. Three of the six problems had nothing to do with what the line says.
The other three were instructions nobody had written down. It skipped the last name and email because it had only ever been told to ask for a name. It mishandled the storm call because storm damage wasn't one of the kinds of calls it had been told about. The owner knew all of that. It just hadn't made it from his head into the system.
52 problems, and none needed a smarter AI
Over the next four weeks we wrote down every problem on that line, no matter how small. There were 52.
Most were wiring: something not connected, a step that didn't fire, a message that went nowhere. A good share were missing instructions, like the storm call. A few were limits of the platform itself. And a handful, which I'll come back to, were fixes that didn't hold.
The number that matters is zero. Not one of the 52 needed the AI to think harder or understand more. Every one of them was fixable by someone checking the connections and writing down what the owner already knew.
That changes what you should worry about. If you're comparing AI answering products and asking which one is smartest, you're asking the wrong question. The one that works is the one somebody has actually set up, connected and called.
A fix that passed once is not a fix yet
Here is the part that stung. About two weeks after that first call, a set of fixes had been written down as done. Then we went and looked at the actual records, and no real caller had booked an appointment through the line by any path. Not one.
Two things had gone wrong. One test call had passed, and it had been written down as proof of a fix two separate times. The fix hadn't held. Separately, a corrected version of the script was sitting in a file on our side while the live line kept answering customers with the old one.
Neither of those is an AI problem. They're the kind of mistake any shop makes when it marks a job finished from the truck instead of walking the roof. The version that counts is the one your customers hear, and the only way to know what they hear is to call.
What this does not tell you
This is one roofer's line, built by us, over a few weeks. It does not tell you that AI answering works for every roofing company, and it does not tell you it doesn't. It tells you what broke on this one and why.
Testing doesn't make a phone line perfect. It catches what you think to check. Real callers will still find things you never imagined, and someone has to keep reading what the line writes down.
It also isn't an argument that a human answering service is safer. A person on the other end can rush, skip the email, or fail to reach you too. The self-check below works on any of them.
And it says nothing about the calls that need you personally. An answering line can take a message and get the details right. It doesn't replace the owner on the call that needs a judgment.
A self-check you can run on any answering system
Whatever answers your phone when you can't, whether that's an AI line, an answering service or a teenager on the office cell, run this before a customer does it for you.
Call it yourself. From a number it doesn't recognize, playing a real homeowner.
Ask something ordinary it might not expect, like your hours. It should stay on the line and answer, or say it'll get you an answer.
Ask for the owner. It should either reach you or say plainly that it can't right now. Silence is the failure.
Call about storm damage. Does it treat that as the call it is, or push you into a menu?
Listen to the pace. Could your oldest customer follow it on a landline?
Read what it wrote down. Full name, a way to reach them besides the number they called from, and what they called about.
Do it again after anyone changes anything. Then check again a few days later. One good call proves very little.
If it fails one of these, good. You found it instead of a homeowner with water on the ceiling.
FAQ
Do AI phone agents work for roofing companies?
They can take calls, capture details and book appointments, but only after someone has set them up, connected them to the rest of your system and called them repeatedly. On the line in this post, every problem we found came from setup, not from the AI lacking intelligence.
What usually goes wrong with a new AI answering line?
On this one: hanging up on ordinary questions, failing to transfer without saying so, skipping contact details, misreading storm calls and talking too fast. Most traced back to settings, connections and missing instructions.
How often should I test my answering line?
Before it goes live, after every change, and again a few days after each change to make sure the fix held. In storm season, call it yourself at least once a week.
Brian runs Omnibus Victis AI, working with home services businesses across Maryland.
