AI-built products don't break at scale. They break at user one.

There's a myth that shipping fast with AI is fine until you scale, and then the cracks show. In my experience it's the opposite. The cracks show the moment someone who isn't you touches it.

The reason is simple: AI optimizes for "works in the demo." And the demo only ever has one user, you, doing the exact happy path. So the stuff that never comes up in a demo just never gets built:

Auth edge cases. Access control. Exposed API keys. Missing security headers. Dependencies frozen at whatever version they were generated with, each carrying known vulnerabilities. Certs that quietly expire. Input nobody validated because in the demo the input was always clean.

None of that is an architecture problem. On day one it all works. It keeps working right up until a real user does something you didn't script, or six months pass and the thing quietly rots while everyone assumes it's fine.

That's the part "just use AI to build it" skips. Generation got faster. Decay didn't get slower. Someone still has to care about the unglamorous stuff before and after launch, and that someone has to actually understand what was shipped.

Curious where others land on this: what's the most boring, unglamorous thing you've seen take down a shipped product? The stuff nobody thinks to check until it's already broken.