Most AI pilots die of vagueness

The pilot went fine. The demo impressed the steering committee, the vendor wrote a case study, and six months later not one production process runs differently. If that story sounds familiar, the problem was probably not the model. It was the mandate.

"Explore AI" is a mandate that cannot fail and cannot finish. It names no process, no owner, no baseline, and no number that should move. A pilot built on it produces exactly what it asked for: an exploration.

Name the process or don't start

Our rule for applied AI is blunt: scope to a named process before any model work begins. Invoice matching. Outage-report triage. Contract clause review. A named process comes with an owner, a data trail, an integration point, and a working definition of done. Everything a pilot needs to become an operation.

Naming the process also surfaces the honest question early: is this worth automating at all? Sometimes the answer is no, and finding that out in week two is a bargain.

Production is mostly plumbing

The model is the visible fraction of the work. The rest is data engineering: pipelines that refresh, definitions that agree with finance, access that survives an audit, and monitoring that notices drift before your users do. Skip the plumbing and the smartest model in the building is a screenshot in a slide deck.

A pilot that cannot name its process is a demo with a budget.

In regulated settings the bar is higher again. A utility that wants AI in its reporting chain needs every output to be explainable to a regulator, which means model choices, data lineage, and human checkpoints are design requirements from day one. Vague mandates do not survive that conversation. Named processes do.