The gap is rarely the model. It is usually the eight steps around it that nobody costed.
A pilot proves a model can do a task on a clean sample. Production asks a different question: can the whole path — intake, routing, review, exception handling, audit — absorb the change without anyone’s Tuesday getting worse. We have watched three organisations run a successful pilot and then quietly shelve it, every time because the surrounding process was never in scope.
The fix is unglamorous. Measure the current path by hand before you build anything: volume, cost per item, error rate, where the time actually goes. Then scope the workflow, not the model.