My First AI Client Doesn't Use Computers.

For the last 6 months I'd been trying to land my first freelance client as an AI consultant. A Koerner Office podcast had planted the idea. So I pitched people. Posted about it. Told anyone who'd listen I could help their business with this stuff. Nothing landed.

Then a client walked up to me while I was paying to get my car fixed.

I'd taken it to a small auto shop. The owner isn't a tech guy. He's the kind of operator I respect: honest about what the car needed, honest about what it didn't, fast, running most of the place himself. He did good work. So when he asked what I do, I told him the truth. I work with AI all day. We stood at the counter and talked about how it helps a business like his, not the version on the news.

That conversation was the whole sale. I just didn't know it yet.

The client I wasn't looking for

Somewhere in the back and forth, he brought up his reviews. He wanted to know if there was a way to fix a couple of bad ones. I told him the truth: I'm not sure, but I can find out and get back to you. A few days later I sent him a short PDF: how to handle bad reviews, the exact steps, no charge. I wanted to hand him something useful before I ever asked him for a dollar.

Then I made the offer. I'd bought Vapi and Twilio months earlier for my washer and dryer rental business and never used them. Sitting there paid for, doing nothing. He's the only person answering his phone, and when he's under a car it rings out. So I offered to point that unused setup at his shop and build him an agent that picks up when he can't.

He asked if I could really build that. He gave me his personal number. I said I'd reach back out when I had something working. Then, because I got excited and jumped ahead of myself, I nearly skipped the part where this is a business. He was the one who said it: could he try it for seven days first? I would have offered that anyway. He beat me to it. Start date, the first of July.

He told me what he wanted, and he doesn't use computers

Here's the part I keep thinking about.

The mechanic told me exactly how the agent should work, and he'd never touched a tool like this in his life.

He didn't want to be replaced. He was clear about it. He answers his own phone every time he can, because that's how he runs his shop. The agent only exists for the moments he misses: he's driving, he's on the other line, his hands are inside an engine. In those moments he wants it to pick up, capture everything he'd need, and hand it back so he can call the person right away and know at a glance whether it's urgent. A net under him. Not a substitute for him.

That's the thing the AI news keeps getting wrong. The guy who supposedly "doesn't get technology" understood the correct relationship with it on the first try.

The AI vendors have a whole genre of blog post about this, and every number in it comes from a company selling the fix. Strip all of it away and one number is left: the average repair ticket is $428. Every call he can't get to is $428 walking out the door, and until now, no way to even know he lost it. He never read one of those posts. He knew it already, in his hands, because he's the one who can't answer while he's under a car. The market the vendors sell as a statistic, he lives as a fact.

And this is San Antonio. Plenty of his customers call in Spanish. So I built it to answer in both.

What shipping one real thing costs

None of this looked like the demo.

A Vapi agent is three models stacked together: one that hears, one that thinks, one that talks. I spent a week going back and forth with Claude and ChatGPT on which combination stays fast without getting dumb, because the cheap models sound warped and the smart ones lag. I tested it against my own phone and email over and over. Early versions repeated my name like a malfunction. I rewrote the script until it stopped sounding like a machine reading a form, made the lead land in an inbox, made it handle both languages.

Then came the part nobody demos: getting his actual phone to ring my agent.

The day before it went live, I sent him the forwarding code. It didn't work. "Oh crap," I thought. I went to Claude to find what I'd missed. The owner was gracious about it: no problem, let me know when you figure it out. I went back in, found a different code, and that one forwarded every call straight to the agent, which is the exact opposite of what he asked for. He wanted to answer first. I sent him the code to cancel it and told him I'd drive over.

I went straight from my other job. Tried it myself in the shop. Same failure. Back to Claude one more time, and there it was: it's a business landline, and business accounts use a different code. I sat in that shop for about 25 minutes typing carrier codes into a phone that wasn't mine, until his landline finally rang and my agent answered it.

That's what "AI is transforming small business" looks like on the ground. A guy on a stool in an auto shop, over on his lunch break, making one phone ring.

The rest of the alphabet

I haven't been paid yet. It's a trial. He still isn't sure it'll be useful to him, and I should ask him more about what useful means, because that answer is the actual product.

But something already changed, and it wasn't the money.

I'd been stuck in a frame and hadn't even noticed the walls. I thought the work had to be X, or Y, or Z. Then I helped one real customer with one real problem, and A through W showed up, a whole world that had been sitting there the entire time. Turns out the sales pitch is simple: get curious about someone's problem, help them with it, then actually charge for the help.

You can't brainstorm your way to the next thing. The opportunity was never in your head. It's standing at the counter while you pay for something else, and you only see it after you've shipped one real thing for one real person.

The client I spent all that time hunting for showed up the day I stopped pitching and started helping.

Sources:

Chris Koerner's The Koerner Office — the podcast