Roughly 95% of enterprise AI pilots never reach production. The number gets quoted as though it describes a technology problem. It does not. The technology usually worked — that is why it became a pilot.
The pilot was never the hard part
A pilot is optimised for one thing: showing that the idea has merit. Clean data, a friendly scope, a demo audience that wants it to succeed, and no obligation to handle the awkward cases.
Production reverses every one of those. Messy data, the full scope, users who do unexpected things, and an absolute obligation to handle the awkward cases — because in production the awkward case is someone's money.
The four places pilots die
The data was borrowed. The pilot used an export someone pulled by hand. Nobody costed the pipeline that would keep it fresh, and the pipeline is often larger than the model work.
Nobody owns it. A pilot has a sponsor. Production needs an owner — someone accountable when it is wrong at 2am. That name is frequently missing, and the project quietly stalls at the point where it would be assigned.
The integration was assumed. "It plugs into the CRM" turns out to mean a decade-old system with no usable API and a team with a full roadmap.
Risk arrived late. Legal, compliance or security see it for the first time near the end. The questions are reasonable, the answers were never designed in, and retrofitting them costs more than the build.
What the 5% did differently
The pattern is unglamorous. They picked a narrower problem than they wanted to. They costed the data pipeline before the model. They named an owner at the start. And they brought the awkward questions forward — what happens when it is wrong, who reviews it, what evidence exists — while those were still cheap design decisions instead of expensive retrofits.
None of that is technically difficult. It is just the opposite of how a pilot is normally run, which is why so few pilots convert.
The practical implication
If you are about to start a pilot, decide up front what would have to be true for it to reach production, and check those things first. If the answer is a data pipeline nobody has scoped or an integration nobody has confirmed, you have learned the most valuable thing available for the price of a conversation.
That is roughly what the Reality Check does: one week to find out whether the thing is buildable, what it costs, and what stops it — before the pilot budget is committed.
This is the part we do — the crossing from a demo to a system that survives production.