IT teams: AI prototypes to production, no rebuild

The demo took a weekend and production took the rest of the year

The problem

AI coding tools have moved the hard part. Getting something working is now genuinely fast, and a small team can put a functioning prototype in front of stakeholders inside a week. The prototype does what it promised, the demo goes well, and the project gets funded on the strength of it.

The usual answer

Then the gap opens. The code was written to demonstrate behaviour, not to hold traffic, and the things it skipped are the things that take the time. There is no test suite, so nobody can change anything with confidence. Authentication was hardcoded for the demo. Nothing is instrumented, so when it misbehaves in production there is no way to see why. The data model works for ten records and falls over at ten thousand. Secrets sit in the repository. None of this failed review because there was no review. What usually happens next is that a second team quietly rebuilds it properly, the timeline doubles, and the original demo becomes a spec document that everyone resents.

How we approach it

We work from what has already been built rather than starting again. That means putting a real backend architecture under it, writing the test coverage that lets anyone touch the code, closing the security gaps a review will find before the review finds them, and standing up infrastructure with deployment, monitoring and rollback that an on-call rota can actually live with. Where the prototype made a choice that will not scale, we replace that piece and keep the rest.

What changes

The version that ships is recognisably the version people saw and approved. The speed the prototype bought does not get handed back, and the team on call afterwards inherits something they can operate.