The demo went brilliantly. The model hit 94% accuracy on the test set, the steering committee applauded, and the project got a green light. Eighteen months later, it has never made a single decision in production. If this story sounds familiar, it should — by most industry counts, roughly four in five enterprise AI initiatives never cross the gap between pilot and production.
Having taken dozens of systems across that gap, we can report that the causes are remarkably consistent — and almost none of them are about model quality.
The three walls
The first wall is data plumbing. Pilots run on curated extracts; production runs on live pipelines that break, drift, and arrive late. If the pipeline feeding your model isn't engineered to the same standard as the one feeding your financial reporting, the model's accuracy is a hypothetical.
The second wall is ownership. A pilot can be owned by a data science team. A production system that makes real decisions needs an owner accountable for those decisions — someone on the business side whose numbers move when the model is wrong. Projects without that owner stall in an endless loop of "one more validation round," because nobody is empowered to accept the residual risk that any live system carries.
The third wall is operations. Models degrade. Inputs drift. Regulations change. A production AI system needs monitoring, alerting, retraining triggers, rollback paths, and an audit trail — the MLOps muscle that turns a clever artifact into dependable infrastructure. This is unglamorous work, which is precisely why it gets skipped.
The playbook
Start pilots on production plumbing, even when it's slower — a model proven on the real pipeline has already crossed half the gap. Name the business owner before training begins, and define with them what accuracy is sufficient, because sufficient-and-live beats perfect-and-stalled every time. Budget the operations work as a first-class line item, roughly equal to the modelling itself. And instrument everything from day one, so the conversation about model performance is a dashboard, not a debate.
None of this is exotic. It is engineering discipline applied to a discipline that has been allowed to live outside engineering for too long. The 20% of projects that reach production aren't luckier — they're just built, from the first week, as if production was the point.