Behind the Scenes
·
7 min read
·
Our AI Quotes Every Lead in Under Five Seconds. It Closes About 1 in 50.
By
Noah Satho
·
Milpitas, California
Why One Quote Took Five Minutes
The inquiry lands in one system, our pricing and our packages live in another, and the two of them don't talk to each other. So somebody has to read what car the person drives out of a form, work out what size class that car falls into, look up what that size costs, assemble the proposal in our shop software and send it, one at a time, all of them by hand. None of that is difficult, and that's exactly what's wrong with it, it's the kind of work that takes five minutes of somebody's full attention and leaves nothing behind except a quote a machine could have written. We measured what that was costing us before we changed anything, and it was worse than we assumed. Across 158 leads between April 16 and May 8, 23.4% of them ever received a proposal at all. Not 23.4% became sales, 23.4% got a price. Everybody else asked us what something cost and never heard an answer from us, and the people who did hear back usually waited a day or two for it, from a guy who by that point barely had time to call them. Since the script went live on May 12 it has sent 625 proposals. They come off the same price list he was reading from by hand, so they aren't a cheaper version of what he was sending, they just go out in seconds instead of waiting for a gap in somebody's day.
The Hard Part Wasn't the AI, It Was a List We'd Been Keeping for Years
We do paint protection, and what a job costs depends on the size of the car, so the whole system rests on getting one question right, which is what size is this vehicle. That turns out to be something you can't calculate. A Lamborghini Urus is an SUV by any measurement you could take of it, and it prices as an exotic, because for a handful of marques the badge settles the class before size ever comes up. A Porsche Carrera T prices as an exotic too and the number 911 never appears anywhere in its name. And any car thirty or more model years old prices as a classic whatever shape it is, which puts a 1966 Mustang and a 1972 Suburban in exactly the same place, two cars that have nothing else in common at all. There's no formula that gets you to any of that. It's knowledge that accumulated in our heads and in a document we'd been adding to for years, over 1,300 vehicles deep by the time any of it got automated, and it's also the part a general purpose AI gets wrong while sounding completely certain about it. That's the reason the answer has to come out of our own list instead of out of a model's general knowledge. So most of the work went into the matching, reading whatever a stranger typed into a form at midnight, in whatever way they decided to type it, and landing it on the right row of a list that already existed. The list was the asset, and the script was the cheap part, which is not how any of this usually gets talked about. And the most valuable thing we put in it isn't the match, it's the refusal. When the script isn't confident about a car it doesn't guess, it stops and hands that lead to a person. We learned why that matters the hard way. A Porsche came through, the matcher fell back to guessing off the make, and the guess went out under what that car should have been quoted. Once somebody is holding a number, that number is real to them, and you don't get to quietly correct it afterward. So now it holds anything it isn't sure about and a person picks those up in the morning, which is slower, and is the right answer.
It Didn't Replace Our Sales Rep, and It Was Never Going To
Out of those 625 proposals, the ones that became a sale with no human involved at any point are about 1 in 50. That sounds like a failure rate and it's actually the whole point. Nearly everything written about AI right now implies the machine replaces the salesperson, and it doesn't, not here anyway, not on four figures of work on a car somebody actually loves. People want to talk to a human being before they hand that over, and honestly they should, we'd think less of somebody who didn't. What the script replaced was the typing. Same rep, same closing ability, he just isn't the bottleneck anymore, and he spends the day on the phone instead, answering the questions people actually have and helping somebody decide about a car they care about. That's the part I'd want anybody automating something in their own business to take from this. I don't want AI to replace human connection, I want it to free us up from mind numbing tedium and allow people more time to connect with one another. Five people work here and we answer like a shop three or four times our size, which is worth something, but it isn't why I built any of it. I can afford to speak with my clients at pickup for 30 minutes, because I can get so much more done with agentic AI and vibe coded automations handling admin work for me, instead of trying to cut our conversation short because I have too much work piling up on my plate. Automate the queue, not the relationship.
Does an AI decide what to charge me?
No. It reads from the same price list our team reads from, so it isn't inventing a number or negotiating with you, it's doing a lookup and filling out a proposal. What it decides is which size class your vehicle falls into, and if it isn't confident about that it refuses to guess and hands your inquiry to a person instead.
Am I going to be talking to a bot?
Only the first quote is automatic. Everything after it is a person, and if you reply to that email or call us you get one of the five of us. The reason we automated the quote is so there's more of a person available for the rest of it, not less.
What if it gets my car wrong?
It's built to stop rather than guess, so the cars it can't confidently identify get held and picked up by a person in the morning. If a quote you already received looks wrong for the car you actually drive, tell us and we'll fix it, we would rather correct a number than defend one.
How fast do I actually hear back?
The proposal itself is built and emailed in under five seconds. A human reply takes as long as it takes somebody to read your inquiry and think about it, which is the part we're deliberately not automating.
What did you build it with?
Claude Code wrote it, in Python, and it runs on its own without anybody starting it. It reads the inquiry out of our CRM, matches the vehicle against our own size list, and builds and sends the proposal in our shop management software. There's no AI writing the price and no AI talking to you.
Did this cost anybody their job?
No. We're five people and we still are, nobody was replaced and nobody was planned to be. What changed is that our sales rep got about four hours of his day back, and he spends them talking to customers instead of typing their details into a second system.