A demo works because everything in it was chosen to make it work. Clean data. A single, well-scoped task. No competing priorities in the room. No one asking what happens when the input is malformed, or late, or contradicts three other systems that were never mentioned in the pitch.
Real operational environments don't offer any of those courtesies. Data arrives incomplete. Priorities compete for the same person's attention. The workflow the AI is supposed to slot into already has its own informal exceptions, workarounds, and tribal knowledge that nobody wrote down because nobody thought to.
The gap between sandbox performance and operational deployment isn't a technology gap most of the time. It's a context gap — the demo was never exposed to the actual mess the real workflow lives in, so it never had to prove it could survive that mess.
The demos that do make it into real use tend to share one trait: someone deliberately fed them the ugly, incomplete, contradictory version of the data before anyone else saw the polished version. That's an unglamorous step, and it's the one most pitches skip, because it makes the tool look worse before it makes it look better.
I've started treating a clean demo as a starting hypothesis, not a finished result. The real test is what happens the first time the input doesn't cooperate.
Current Hypothesis
A demo fails in production for the same reason a bridge fails in an earthquake it wasn't designed for — not because the design was wrong, but because the design never accounted for the actual conditions it would face.
Not a conclusion — a working idea, revisited as evidence comes in.