Pilot purgatory is a funding artefact
Pilots do not stall because the technology underperformed. They stall because pilot money and production money come from different pockets, and the second pocket belongs to someone who never asked for the pilot.
- The block is the budget, not the technology. Pilot money and production money come out of different pockets.
- The owner of the production budget never asked for the pilot, and faces a permanent cost increase against a target they already carry.
- Ask before you start: which budget line does production come from, and who owns it by name. If nobody can answer, the pilot will stall however well it performs.
- If the accuracy held and the decision still never came, stop looking at the model.
A pilot that works and then goes nowhere is the most common failure state in enterprise AI, and it is almost never a technical result. The pilot was funded from a discretionary pot: an innovation budget, a transformation office, a CIO reserve, sometimes a vendor credit. That pot exists to buy learning, and its approval threshold is low because the amounts are small and the owner is a sponsor who wants things tried.
Production is funded from a different place. It comes out of a line owner's operating budget, and that person did not ask for the pilot, was not in the steering meeting, and now faces a permanent cost increase against a target they are already carrying. The pilot result is not evidence in the conversation that decides their budget. It is a story about someone else's experiment.
The diagnostic
Ask one question before the pilot starts: which budget line does the production version come out of, and who owns that line by name. If nobody can answer, the pilot will stall regardless of how well it performs. You are not running a test of the technology. You are running a test of an approval path that does not exist yet.
The answer has to be a person, not a function. “Operations” is not an answer. “It would sit in the shared services envelope” is not an answer. A name is an answer, and the useful version of the question is whether that named person has seen the number and said the word yes.
What this changes
- Recruit the production budget owner as a co-sponsor before the pilot, not as an audience after it. Their objection at month six is information you could have had at month zero.
- Design the pilot to produce the evidence that owner needs, which is usually unit cost and failure rate under their volumes, not the accuracy figure a steering committee enjoys.
- If no owner will take the line, stop. A pilot with no landing budget is a training exercise, and it should be funded, scoped and described as one. There is nothing wrong with buying learning on purpose.
The counter-case
Occasionally the block really is technical, and the tell is specific: the pilot succeeded on a curated slice and degrades on the real distribution. That shows up as accuracy falling when you widen the input, not as a decision that never gets made. If the number held and the decision still did not happen, the problem is the pocket.