Answers
What does it take to move an AI prototype into production?
Usually three things a prototype doesn't have yet: real monitoring so you actually know the moment it starts getting things wrong, a maintenance plan for the models and dependencies it quietly depends on, and a defined incident process for the day something inevitably breaks in production.
Monitoring means knowing the moment an AI feature starts getting things wrong, not finding out from a customer complaint. That means tracking the same accuracy and confidence metrics used during evaluation, in production, continuously — not just at launch — so drift shows up as a graph moving in the wrong direction rather than a support ticket.
A maintenance plan means someone owns keeping the models and their dependencies current: a model version deprecation notice from a provider, a library update that changes behavior, a data source that silently changes format upstream. A prototype nobody's watching for these things degrades quietly until it's visibly broken.
An incident process means having an actual answer, decided in advance, to "what happens when this breaks in front of a real user" — who gets paged, what the rollback looks like, whether the feature fails safely to a human fallback or just fails outright. A prototype that's never had to answer that question isn't production-ready yet, no matter how good its demo looked.