A pilot that clears its accuracy target and still never reaches production is the most common outcome in enterprise AI. The postmortem usually blames the model. It is almost never the model.
What is missing is the operating model: who receives the output, what they are expected to do with it, what happens when it is wrong, and who is accountable when it is wrong twice. None of that is a modelling problem, and none of it can be retrofitted after a pilot has already shaped everyone's expectations.
The teams that get to production write the exception path first. They decide what a low-confidence result looks like on screen, who it routes to, how long that person has, and what evidence is kept. The model is then built to fit a workflow that already works on paper.
Audit trails matter for the same reason. If a decision cannot be reconstructed six months later, it cannot be defended to a regulator, an auditor or a customer — and a system that cannot be defended will not be allowed to make decisions that matter.