A roofing receptionist that answers at 2am and will never quote a price

A homeowner with a leak calls three contractors. The one who picks up gets the inspection. This is not a subtle market dynamic, it is the whole thing. Every call that goes to voicemail after five is a lead being handed to whoever answers next.

But the fix is not just answering. The first question a homeowner asks is what it costs, and the honest answer is that nobody can price a roof from a phone call. Any ballpark given becomes the number you are held to, and it gets given by the person with the least information about the property. So a phone system that helpfully offers a range has not solved the problem, it has created a new one.

There was a third thing we found once we started building, and it turned out to be the most important. The FAQ document the business already had was actively sabotaging the phones. It pushed toward collecting an email address and hard-booking the appointment on every call, which is telemarketer behavior, and homeowners hang up on it. The receptionist's real job is warm capture toward a booked inspection. We had to gut that document to get there.

We built Sarah, a voice receptionist wired into JobNimbus. She answers at any hour, understands what the homeowner is describing, checks the service area, asks the qualification questions the estimator would ask in the order he would ask them, and books the inspection. The call record and the appointment land in JobNimbus, so nobody rekeys anything and there is no second system to check.

The trade knowledge is real, and that matters more than it sounds. She knows the difference between a manufacturer certification and a workmanship warranty. She knows what a square is, which is 100 square feet of roof. She knows that the enhanced GAF warranties require a qualifying full GAF system, which is why brands do not get mixed on those jobs. A system that gets any of that wrong says something embarrassing to a homeowner who knows better.

The hard rules are built in as constraints rather than instructions. She does not quote a price. Not a range, not a per-square number, not a most-jobs-run-about. She does not ask for an email. She does not invent a fact or a commitment. When something is outside her boundary, an angry customer, an insurance adjuster, a commercial job, she stops and hands the call to a person with the context attached.

Then there is the part that actually decides whether any of this ships: the grading. Every call is scored against a written rubric in two layers. A deterministic layer checks the hard rules first, and the caps mirror the trade's real risks: quoting any price caps the score at 50, asking for an email caps it at 50, fabricating a fact or a commitment caps it at 45. Then a model judges the softer parts against the same rubric a human would use, and it has to band-match a human grader on 9 of 10 labeled golden calls before we trust its judgment at all.

The system is deployed and it is dark, which is a deliberate state. It is switched off until the owners are ready, and that is the correct order of operations: build it, prove it, then let the business decide when it starts answering their phone. We are not going to publish a call volume number, because the honest one right now is zero, and a number we made up would be worth less than nothing.

What we can point at is what it took to get here: 132 passing tests, a 52 out of 52 migration slice when we moved the voice layer, and 21 tests that exist purely to prove the safety guardrails hold. Ten labeled golden calls with a written rubric, and a model judge that has to agree with a human grader before its scores mean anything.

The reason that matters is simple. This system talks to homeowners, in your name, when nobody is watching. The question is not whether the demo sounded good. The question is what happens on call four hundred, at 2am, when somebody asks a question nobody anticipated. The tests are the only honest answer to that, and they are why the owners get to decide when it goes live instead of finding out the hard way.