The AI Receptionist That Should Not Have Been Built

The plan they walked in with

The brief was specific and confidently argued. An AI voice agent would answer every inbound call across four branches, qualify the caller, look up availability, and book the job. Leadership had a vendor quote, an internal champion, and a board slide showing recovered revenue from missed calls.

The number on that slide was the entire justification. It had been calculated by taking every unanswered call from the phone system report, multiplying by average ticket value, and assuming each one was a lost job. That arithmetic produces a very large number, and nobody had questioned it.

Our first request was not for a technical spec. It was for six months of call detail records and the corresponding job records, so we could match one against the other.

What the numbers actually said

Roughly two thirds of unanswered calls came from numbers that called again within twenty minutes and reached someone. Those were not lost customers. They were mildly annoyed customers who had already been counted as revenue on the board slide.

A further slice were existing customers calling about a job already scheduled, plus a meaningful volume of supplier and recruiter calls. Once we removed everything that was not a genuine new-work enquiry that never came back, unanswered calls accounted for about eleven percent of the revenue leak leadership believed they were solving.

The eighty-nine percent was somewhere else, and it was not hard to find once we stopped looking at the phone system. Estimates were being written and then never followed up. Nearly a third of quoted work over the prior year had no recorded second contact of any kind. Nobody owned the follow-up, so it happened when a service manager had a slow afternoon.

The part that would have cost them three times over

Cost was only half the problem with the original plan. The other half was that it could not have been built as scoped. The voice agent needed live technician availability to book a job, and their dispatch system exposed availability only through a nightly export. Booking against a schedule that was up to twenty-four hours stale would have produced double-booked trucks in week one.

Closing that gap meant either replacing the dispatch platform or paying for a custom real-time integration the vendor had not quoted. Both had been priced at zero because neither had been discovered. Adding realistic figures took the programme from the quoted number to roughly three and a half times our recommended alternative, on a timeline stretching past nine months.

We put that on one page: their plan with the missing line items filled in, against a staged alternative aimed at the eighty-nine percent. We did not tell them the voice agent was a bad idea forever. We told them it was the fourth-best use of the money and could not run correctly until the dispatch data problem was solved anyway.

What we recommended instead

Three pieces, sequenced by payback. Automated estimate follow-up first, because unsold quotes were the largest single pool and the work was mostly plumbing between systems they already owned. Then after-hours intake capturing enough detail overnight to schedule in the morning, which addressed the genuine share of the call problem at a fraction of the voice agent cost.

Third, a dispatch data fix, which was unglamorous and was the reason the original plan could not work. Getting availability out of the scheduling system in real time was a prerequisite for anything customer-facing, including the voice agent if they still wanted it later.

The roadmap was explicit that the voice agent stayed on the table. It moved from phase one to a decision point after phase three, at which point the integration would already exist and the honest cost would be a fraction of the original quote.

Why the assessment paid for itself before anything was built

The most valuable output was not the roadmap. It was the six weeks and the six figures that did not get committed to a project whose central assumption was wrong.

Leadership was not wrong to want the voice agent. They were working from a number that had never been tested, produced by a report that could not distinguish a lost customer from an impatient one. That is the ordinary case, not an unusual one.