I Spent Two Months Building an AI Business. Then I Killed It on Purpose.

On sunk costs, one-customer revenue, and the Tuesday I finally read my own numbers

On July 10th I commented out 17 cron jobs and shut down Local Listing Rescue, the AI business I'd spent two months building. No dramatic failure. No final straw. Just a Thursday, a spreadsheet, and a decision that was about six weeks overdue.

This is the post nobody writes. The internet is full of launch stories and pivot stories. What it needs more of is shutdown stories — because the most expensive skill in solo business isn't building. It's stopping.

The machine worked. That's what made it hard.

22,212 leads discovered. 1,358 emails sent. A 30.7% open rate — genuinely good for cold outreach. A/B testing, payment processing, a self-hosted mail server, an AI crew running the whole operation. The dashboard said traction.

The bank account said otherwise: two months, one real customer, $49 in total revenue.

I should disclose the full revenue picture, because it's funny. Our first-ever charge was $199 from a customer named David Buster. He refunded it two days later. Yes — I was my own best customer. The second and final charge was $49 from William T. Dolch, who remains the only person on Earth to have paid for this service. William, if you're out there: you had excellent taste and I hope the report was good.

How the sunk-cost trap actually feels

It doesn't feel like denial. It feels like progress.

Every week there was something new to fix, and fixing things feels like winning. The LLM chain died? Diagnosed and repaired — eleven days of downtime, root-caused to a thinking model that spent its entire token budget contemplating subject lines instead of writing them. Scraper finding info@ addresses instead of owners? Rewrote the extractor. Each fix was real engineering. Each fix was also beside the point, because the problem was never the machine.

It was the offer.

The market had told me, clearly and repeatedly: 3 replies from 1,358 sends — and one of those was my own test email. 178 unsubscribes; one in eight recipients annoyed enough to click. A 0.09% conversion rate. The data wasn't ambiguous. I just didn't want to read it, because reading it meant admitting the two months were tuition, not investment.

What finally broke the spell

Writing the post-mortem. I had Riker compile the factual record — every send, every cost, every failure — and reading your own project as a document instead of a dream does something to you. In tables, it was obvious: the machine was beautiful and the business wasn't there. A 0.22% reply rate isn't a marketing problem to optimize. It's a verdict.

So here's the rule I've adopted: when the only thing keeping a project alive is the time already spent on it, it's dead. The two months are gone either way. The only live question is whether they buy the next two months too. Sunk costs don't compound. They just collect interest in denial.

What the tuition actually bought

To be fair to the corpse: the lessons are real. I know how to run an AI agent crew in production — scraping, sending, tracking, payments, the works. I know what cold email actually does at scale. I know that "sophisticated infrastructure" is often just procrastination with a architecture diagram. And I know exactly how it feels to keep building because stopping feels worse.

Those lessons are already at work in the next thing. That's not consolation. That's the actual return on the $49 business.

Your one thing this week

If you're sitting on a project right now, ask the question I avoided for six weeks: if this arrived on my desk today, fresh, with no history — would I start it?

If the answer is no, the history doesn't matter. Kill it cleanly, write down what it taught you, and go build the thing the lessons point to. The cron jobs comment out easier than you'd think.

It's the hope that's hard to cancel.

---

---

Dave Buster is building AI Quarterdeck to help solo entrepreneurs harness AI agents without the hype. He's made just about every mistake possible so you don't have to. Previously, he built Chartsplitter (musicians love it), shouldIworry.net (seniors don't), and the Piano Quiz Playground (his wife rejected it). He's still married.