AI products pass through five stages: idea, prototype, pilot, production, and operation. The stage that kills most of them is the move from pilot to production — where a thing that demos well must become a thing people depend on, with ownership, observability, and a review path.
The five stages
- Idea. A thesis about where AI moves a number. Cheap. Plentiful. Worth almost nothing on its own.
- Prototype. A working sketch that proves the capability exists. Demos well. Fools everyone.
- Pilot. Real users, real data, narrow scope. The first honest signal.
- Production. Depended upon. Instrumented. Owned. The graveyard stage.
- Operation. Running, monitored, improving. The only stage that pays.
The stage that kills products
Pilot → production is where most AI products die. Not because the model fails — because nobody owns the move from "it works when I run it" to "other people rely on it without me." That transition needs observability, an evaluation harness, a defined review path, a rollback story, and a named owner.
A prototype answers one question impressively. A product answers a thousand questions reliably and tells you when it shouldn't answer at all. The distance between those two is the whole job.
How to cross it
Treat production as a stage with its own scope and budget, not a victory lap after the demo. Name the owner before you build. Instrument the baseline before you optimize. Design the review path before you ship. The teams that plan for this stage cross it; the teams that assume it cross nothing.