Ask a leadership team how many AI pilots it is running and the number is usually impressive. Ask how many are in production, with a named P&L owner and a measured baseline, and the number usually collapses to one or two.
We audited 180 stalled use cases across banking, retail, manufacturing, and the public sector. Three failure modes account for four out of five stalls, and none of them is about model quality.
“Name the baseline, the owner, and the destination system. Missing any one means it is still a pilot.”
The first is an unmeasured baseline. If nobody wrote down the cost per claim, the cycle time per order, or the conversion rate before the pilot started, the pilot cannot prove anything, and the sponsor who must fund scale-up has no number to defend. Fixing this costs a week and is the single highest-return intervention we have found.
The second is orphan ownership. Pilots are typically owned by a centre of excellence that has enthusiasm but no operational authority. Scale-up requires a business owner who is measured on the outcome after go-live. Handing over means transferring the model, the monitoring, and the failure playbook — and accepting that the business owner may change the use case to fit reality.
The third is integration debt. A model that works on a laptop often depends on data that arrives nightly, in a format nobody maintains, through a pipeline nobody owns. Productionising means building the boring middle: identity, access, logging, drift detection, and a rollback path. Teams that budget this as a first-class workstream finish. Teams that treat it as overhead do not.
The practical test we give clients is blunt. For any use case, name the baseline, name the P&L owner, and name the system the output lands in. If any of the three is missing, the use case is a pilot, not a product — and it should be resourced accordingly.
