The pattern is familiar. An executive sponsor funds a pilot; the demonstration impresses; the deck circulates. Six months later the pilot is still a pilot, and the team that built it has moved on. The organisation concludes — wrongly — that AI "isn't ready".
In our experience the technology is rarely what fails. What fails is everything around it: data that was hand-prepared for the demo, integrations that were never designed, governance questions deferred until a regulator or risk committee asks them.
Start where the pilot will have to live
A production AI system inherits the constraints of its environment: identity and access, audit, data retention, service levels. If the pilot ignores these, its success tells you little. The most useful early question is not "does the model work?" but "what would it take for this to run inside our systems, under our controls, for two years?"
The most expensive AI programmes are the ones that prove value in a demo environment nobody can ever leave.
Three disciplines that carry pilots into production
Data readiness before model selection. Assess quality, lineage and access control on the real data the system will use — not an extract. Most "AI failures" are data failures discovered late.
Integration as a first-class requirement. Decide early which systems of record the AI reads from and writes to, and design the human-in-the-loop checkpoints where judgement matters. Retro-fitting these doubles the cost.
Governance sized to the risk. Regulated organisations don't need to move slowly; they need controls proportionate to the decision the system influences. A document classifier and a credit decision do not deserve the same committee.
None of this is glamorous. All of it is what separates the organisations quietly compounding value from AI from those still commissioning pilots.