Mining Your Own Calls: Using AI to Find Out Why You Lose Jobs

Every service business has a theory about why it loses jobs. Usually it's price. Ask the owner and you'll hear that the market has gone cheap, that a competitor is buying work, that customers don't understand what good work costs. Sometimes that's true. But the theory almost never comes from evidence — it comes from the two or three calls that stung the most. Meanwhile, the actual answer is sitting in a folder of call recordings nobody has opened since the day they were made.

You already have the data

If you run call tracking, a CRM with a phone integration, or a hosted phone system, you probably have recordings or transcripts of most of your inbound calls. Ninety days of them is a lot of material: every objection a customer raised, every question your front desk couldn't answer, every quote that went out and never came back. Until recently that pile was useless, because reviewing it meant paying someone to listen to hours of audio and take notes. Reading a hundred transcripts is now a task you can hand to a model and get back in an afternoon.

Before you do anything with recordings, check how you're allowed to use them. Consent rules for recording calls vary by state and by province — some places require only one party to consent, others require everyone on the line. Whatever your disclosure says at the start of the call is the boundary you're working inside. If your recordings exist for quality and training, analyzing them for lost-job patterns generally fits; sending them to a third-party tool is a different question, and worth ten minutes with whoever handles your contracts.

Tag the calls, don't summarize them

The instinct is to ask the model to summarize each call. Don't. Summaries read well and tell you nothing you can count. What you want is tagging: give the model a fixed list of categories and make it assign each call to one, with a short quote as evidence.

Build the list yourself, from what you already suspect. For a roofing company it might be: price objection, timeline objection, insurance or financing question, wanted a service we don't offer, out of service area, never reached a human, got a quote and went quiet, booked. Five to ten categories, mutually exclusive, written in your own language. Then the instruction is simple: read the transcript, pick one category, quote the line that made you pick it, and flag anything that doesn't fit.

That last part matters more than it sounds. The calls the model can't categorize are where the real discovery is. If eleven calls come back as "doesn't fit" and eight of them are people asking about a service you quietly stopped offering last year, you've just found something no summary would have surfaced.

Run it on a real batch — a quarter's worth, not a week's — and ask for the output as a table you can sort. One row per call: date, who took it, category, the quote, whether it booked.

Read the results without fooling yourself

Now the part where most of this goes wrong. A model will produce a confident, tidy breakdown of two hundred calls whether or not its tags mean anything. Before you act on the table, pull fifteen rows at random and listen to those calls yourself. If the tags match what you hear, trust the rest. If they don't, your categories are ambiguous — usually two of them overlap — and you rewrite the list and run it again.

Then watch for the traps. Small categories aren't findings; if a reason shows up four times in a quarter, that's noise, not a pattern. Recorded calls are also a biased sample: they miss the customers who never called because your listing was wrong or your ad ran at midnight. And a stated objection isn't always the real one. "Too expensive" is what people say when they mean they weren't sure you'd show up, or they didn't understand what was included, or they'd already decided on someone else and wanted to end the call politely. What the tagging gives you is a map of where to look, not a verdict.

The useful comparison is between people, not just categories. If price objections are spread evenly across everyone who answers the phone, you have a pricing or positioning issue. If they cluster with one person, you have a training issue, and it's a much cheaper thing to fix.

What to automate, and what stays your call

Automate the mechanical parts and nothing else. Transcription, tagging against your fixed list, the monthly table, an alert when a category jumps sharply — all of that can run on a schedule without you touching it. Automate the assembly of the evidence.

Keep the interpretation. Deciding that "too expensive" means you need to raise your close rate rather than cut your price is a judgement about your market, your capacity, and what kind of business you want. So is deciding which objection to address first, and whether the answer is a script change, a website change, or dropping a service line. A model can tell you that thirty percent of your lost calls mention timeline. It cannot tell you whether the right response is hiring a crew, quoting honestly and losing those jobs on purpose, or building a standby list.

And keep the listening. Once a month, sit with five lost calls end to end — not the transcript, the audio. The tags tell you what to look for. The tone tells you what actually happened.