AI can generate a working-looking app in an afternoon. We love that. What we don't love is what happens three months later, when it breaks in production and nobody in the building can explain why. Lately, that's who calls us.
First, let's be clear about where we stand: we use AI every day. It's a genuinely powerful tool, and it's made parts of our work faster and better. This isn't an anti-AI piece. It's a warning about a specific, increasingly common failure mode, the app that was vibe coded into existence and is now quietly falling apart.
"Vibe coding" is the new habit of building software by describing what you want to an AI, accepting whatever it produces, and shipping it, without anyone on the team truly understanding the code underneath. It feels like magic. You type a sentence, an app appears. For a demo or a weekend prototype, that's fantastic.
The problem starts the moment that prototype becomes something a real business depends on. Because the AI optimized for "looks like it works," not for "will still work under load, with real data, a year from now, when someone needs to change it."
AI writes code fast. Engineering is what makes that code survive contact with reality.
We've now been hired to rescue several of these projects. The symptoms are remarkably consistent:
Individually, each of these is fixable. Together, in a system nobody understands, they turn every small request into a frightening one.
Here's the uncomfortable truth: the AI did exactly what it was asked. It was asked to make something that looks like it works, and it did. It was never asked, and couldn't be held accountable, to make something maintainable, secure and correct. That part is engineering, and engineering is a discipline, not a prompt.
An experienced engineer brings the things a model on its own doesn't: a mental model of the whole system, judgement about trade-offs, an instinct for where things will break, and accountability when they do. Skip that, and you don't have a shortcut, you have a debt that comes due at the worst possible moment.
When one of these lands on our desk, we don't just paper over the cracks. We do what should have happened the first time:
Ironically, we still use AI heavily to do this, just with an engineer holding the wheel, reviewing every line and answering for the result. That's the difference. AI as a tool in expert hands is an accelerator. AI as the whole plan is a liability.
If you built something with AI and it's working, wonderful, keep going. But if you're starting to feel that quiet dread, changes that break things, a system nobody can explain, bugs you can't reproduce, don't wait for it to fail in front of a customer. That feeling is the debt coming due.
We can tell you honestly what's salvageable and what isn't, and rebuild it on a foundation that will still be standing in five years. No hype, no fear-mongering, just engineering.