Writing · Field notes
Why enterprise AI pilots stall, and what to do about each one.
Pilots rarely die because the model was wrong. They die at one of five predictable points. Here is each one, and the fix.
· 6 min read
Enterprise AI pilots rarely die of a technical death. The model usually works. The pilot dies somewhere between working and shipping, at a point that has nothing to do with machine learning and everything to do with how enterprises actually make decisions. After twenty years delivering for Shell, Freshfields, JP Morgan and T-Mobile, I can name the five places it happens. Each has a fix, and the fix is almost never a better model.
Stall one, no owner.
The pilot was run by a team that does not own the budget line it would need to scale. It succeeds, everyone claps, and then there is no one whose job it is to take it forward. The fix: name the production owner before the pilot starts, not after it succeeds. If no one will own it in production, the pilot is a demo with a longer runway.
Stall two, the review no one prepared for.
The agent works, then legal, security and compliance ask their questions, and the answers do not exist. The pilot does not fail the review. It waits for it, indefinitely. The fix: answer the review's questions during the build, not after. Treat data ownership, audit trails and the classification under the EU AI Act as design inputs, so review is a checkpoint rather than a wall.
Stall three, the wedge was too wide.
The pilot tried to prove the whole vision at once, so it was too big to evaluate, too big to govern and too big to trust. The fix: pick a wedge narrow enough that one person can judge whether it worked, prove it, then widen. One brick, laid and inspected, beats a wall no one signed off.
Stall four, no number the board recognises.
The pilot reported model metrics, and the board thinks in cost, risk and revenue. The two never met. The fix: agree the business number before you build, the one that will appear on a slide the board actually reads, and report it monthly from day one.
Stall five, success had nowhere to go.
The pilot worked and then stopped, because no one had decided what the next step was if it succeeded. Momentum died in the gap. The fix: name the next wedge before the first one finishes, so a win has somewhere to travel.
None of these is a model problem. All of them are decided before the first line of code, in who owns it, how narrow it is, and which questions you chose to answer early. That is the actual work of getting AI into production, and it is why the hard part of this job was never the AI.
The letters
Occasional letters on getting AI into production.
Work email only.