The Rescue Bill: What It Costs to Fix a Vibe-Coded Product

The build took a weekend. The cleanup takes two quarters.

That second half never makes it into the pitch. AI coding tools genuinely collapsed the distance between an idea and working software, and I use them every day, so I'm not about to pretend otherwise. But plenty of founders heard the first half of that sentence and stopped listening. The invoices are landing now.

The numbers going around are rough. By the end of 2025, something like 10,000 startups had put AI-built apps into production. By mid-2026, more than 8,000 of them had needed a partial rebuild or rescue engineering to keep running. Typical cost: $50,000 to $500,000.

Treat the precision as marketing. The shape of it matches what I see when founders call me.

The debt nobody chose

Normal technical debt is a loan you took on purpose. You knew the shortcut. You wrote it down. You meant to pay it back.

The AI-generated kind arrives without anybody deciding anything. The code showed up, it passed the demo, it went in.

An analysis of 8.1 million pull requests found technical debt rising 30 to 41 percent after teams adopted AI coding tools. Code duplication up 48 percent. Refactoring activity down 60 percent.

Read those last two together. Your system is producing more near-identical copies of the same logic while the people maintaining it do less consolidating. Every copy is a place a future fix has to land, and nobody knows how many copies exist.

Then a change request arrives. "Add a discount code." Should be a day. It's 3 weeks, because the pricing logic lives in 6 places that mostly agree with each other.

What the rescue bill actually buys

Most of it is archaeology.

Before anyone can safely change the code, someone senior has to work out what it currently does. There's no spec to read. The person who prompted it can't tell you either, because they never read the output.

[ROGER: a specific example from a rescue you've done. What the "simple" change was, how long it should have taken, how long it actually took, and the thing you found in the code that explained why.]

Then there's the security tail. A scan of more than 1,400 vibe-coded production applications found 65 percent had security issues and 58 percent had at least one critical vulnerability.

That line item doesn't go in the rebuild estimate. It goes in a disclosure conversation, possibly with a customer.

The tells

You don't need an audit to spot this.

Your estimates stopped being useful. Wrong by 300 percent, in both directions.

There's a file nobody wants to touch. Your team routes around one part of the system, and the routing has a name.

Bugs come back. You fixed it, it returned, because you fixed 1 of the 6 copies.

Onboarding a developer takes months instead of weeks, and the honest reason is there's nothing to read.

The version that skips the rescue

Writing everything by hand isn't the answer. That ship sailed, and it was a slow ship anyway.

Generated code needs an owner. That's the point of the 70/30 split: agents do the mechanical 70 percent, the scaffolding and boilerplate and test fixtures, fast and cheap. A senior human owns the other 30 percent, deciding what shape the system should have, reading what came back, refusing the parts that don't fit.

And by reading, I mean reading. A rubber stamp gets you the same codebase with extra steps.

That 30 percent feels like friction when you're moving fast. It's also the whole difference between a codebase you own and one you're renting from a model you can't ask questions.

One more number, because it explains the dynamic better than anything else I've seen. Developer trust in AI coding tools fell from 43 percent to 29 percent over 18 months while usage climbed to 84 percent.

Your team is probably shipping code they don't trust, because the pressure to ship is louder than the doubt. You're the one who can turn that pressure down.

The weekend build is real. So is the bill. You just get to pick whether you pay it in review time now or rescue engineering later, and the exchange rate is terrible.