A pilot is a demonstration that a capability is possible. Production is a commitment that it is reliable, owned and affordable at volume. These are different problems that reward different behaviour, and a pilot built to demo well is engineered to avoid the very questions production demands.
A pilot optimises for the wrong thing: a curated dataset and a forgiving demo path. Production demands reliability on the long tail, a named owner, a real evaluation harness, an aligned workforce and a known unit cost. The causes arrive in a fixed order: ownership, evaluation, change management, run-cost.
Sunk pilot cost with zero return, an AI roadmap that loses board confidence, and talented teams demoralised by work that never ships.
A second failed push to production costs more than the first, and the organisation accumulates a portfolio of successful demos instead of running systems, while competitors who shipped compound their advantage.
Tracking pilots in slides and a spreadsheet status tracker tells you a pilot exists, not whether it can ship. It hides the four real blockers: no owner, no held-out evaluation set, no change-management plan, and an unmodelled run-cost. A green status cell is not production readiness.
Refuse to start a pilot until you have named the production owner, defined the evaluation metric, identified whose work changes, and produced a cost-per-task estimate at full volume. For a stalled pilot: a production-readiness diagnostic that names the specific blockers, then a recovery plan that fixes them.
Our pilot-to-production recovery is a fixed programme: a two-week diagnostic against a production-readiness scorecard, then we stand up the missing evaluation harness, assign one accountable owner, model true run-cost, and ship. If the honest answer is to kill it, we say so before you spend on recovery.
Organisational, not technical: ownership, evaluation, change management and run-cost, in that order. The model is rarely the blocker.
We will tell you in the diagnostic, with reasons, before any recovery spend. Roughly one pilot in five is better retired or rebuilt than rescued.
Typically eight weeks from the diagnostic to a controlled production go-live, depending on the blockers we find.