A recognisable pattern has established itself inside large organisations: an impressive proof of concept, enthusiastic sponsorship, and then a project that never quite becomes a system anyone depends on. The model is almost never the reason.
The demo was built on the easy tenth of the problem
Proofs of concept are built on clean, hand-picked data. Production meets the other ninety percent: the scanned document at an angle, the record with the customer's name in the wrong field, the format that changed in 2019 and was never migrated. Accuracy that looked excellent on the sample degrades sharply, and the gap between the two numbers is where confidence in the project dies.
Nobody specified what happens when it is wrong
A system that is right most of the time needs a defined path for the rest. Who reviews the uncertain cases, how quickly, and what does the customer see meanwhile? Pilots rarely answer this because in a pilot a person is watching everything. In production that person is a cost line, and if the workflow was never designed the errors reach customers unmediated.
Integration is the real project
The model is a small component. The work is authentication, permissions, audit logging, monitoring, a rollback plan, and making it talk to a system that was written twenty years ago and has no API. That is ordinary engineering, it takes ordinary engineering time, and it is consistently left out of the estimate because the demo did not need any of it.
The measure was never agreed
"Improve efficiency" cannot be evaluated. "Reduce average handling time for this claim type by a fifth without increasing appeals" can. Projects without a number to hit run until the sponsor's attention moves.
None of these are AI problems. They are the same problems that have always separated a prototype from a system, and they are worth naming early because they are all solvable if someone owns them.