Why AI pilots stall before production — and what changes when they do not

A capable engineer can build a persuasive AI prototype in a few days. It answers questions about a sample of company documents, it summarises well, and it demonstrates obvious potential. Leadership sees it and approves further investment. Then, months later, nothing has reached production.

This pattern is common enough to be predictable. It is worth being precise about the cause, because the usual explanation — that the model was not good enough — is almost never the real one.

A prototype typically runs against a curated sample of data, with permissions ignored, in a single language, with one user, no audit requirement, no cost ceiling and no definition of acceptable accuracy. Production has none of those conveniences.

The specific things that stop a prototype becoming a system are consistent: retrieval that has to respect existing entitlements so a user never sees through AI what they could not open directly; integration with systems that have real authentication, rate limits and failure modes; a way to tell whether a change improved or degraded quality; a defined owner for accuracy, cost and incidents; and a support model for when it misbehaves at 2am.

None of these are model problems. They are engineering and operating-model problems, and they are why AI delivery looks more like platform engineering than data science.

The practical fix is to invert the sequence. Instead of proving the model works and then asking how to operationalise it, start by identifying the single hardest production constraint — usually permissions, integration or evaluation — and prove that first.

If retrieval cannot respect your access model, nothing else matters. If you have no way to score output quality, you cannot safely change anything after launch. If the process has no human owner at the point of consequence, security and legal will not approve it regardless of accuracy.

A prototype that resolves the hardest constraint and demonstrates modest capability is far more valuable than an impressive demo that avoided every real question.

A production-ready AI capability has a defined scope with measurable success criteria, grounded output with citations to approved sources, permission-aware retrieval, an evaluation set that runs on every change, cost and latency visibility per application, an escalation path to a named human owner, and a support model with the same seriousness as any other production system.

That list is unglamorous. It is also the difference between a pilot and a capability.