Full Article
A striking share of enterprise AI pilots never reach production. In our reading of why, the model is almost never the binding constraint.
1. Nobody owns the definitions
The data exists across six systems and three of them disagree about what a "customer" is. This is not an AI problem, it is a data governance problem, and it usually consumes more of the timeline than the modelling.
2. The pilot proved the wrong thing
A demo on curated sample data proves the model works on curated sample data. It says nothing about the messy scanned document, the edge case, or the 3am failure. Scope pilots around the hard cases, not the clean ones.
3. Compliance was consulted last
The system works, then legal asks where the data goes, the answer involves a third-party endpoint in another jurisdiction, and the project pauses indefinitely. Involve the blocker at design time; it is far cheaper than rebuilding.
4. Nobody costed production
Pilot volume is a rounding error. Production volume is a budget line. Projects get killed in month four by a bill nobody modelled.
5. There was no kill criterion
Without an agreed definition of failure, a pilot cannot conclude - it just quietly loses its sponsor. Agree in writing, before work starts, what result would mean stop.
The pattern behind the patterns
All five are decided before any code is written. That is why we front-load scoping, data readiness and written success criteria, and why we would rather lose a deal at the scoping stage than deliver a pilot that was never going to ship.
Tags
