Most AI pilots fail before a model is chosen.

The failure is almost never the model

The AI projects that quietly die share a shape. Someone saw a demo, a tool was bought under time pressure, it was pointed at a workflow nobody had mapped, and no one recorded what performance looked like beforehand. Twelve months later there is a subscription nobody can defend and nobody can cancel, because cancelling would mean admitting it never worked.

Not one of those failures is a modelling problem. They are all decision problems, made months before any model was involved.

How use cases get scored

Every candidate use case gets scored on the same five axes, and the scoring is done against your workflows rather than a vendor's case study.

• Hours reclaimed per month, measured from the actual process rather than estimated
• Tolerance for error, meaning what a wrong answer costs and who catches it
• Run cost at your real volume, including the month volume triples
• Data sensitivity, and whether the workflow can legally leave your systems at all
• Time to a measurable result, because a nine-month pilot is not a pilot

Build, call an API, or buy

Decided per use case, with the numbers written down. Most teams reach for the two extremes and both are usually wrong: building a model when a well-prompted API call and a fortnight of integration would have done it, or buying a platform whose one useful feature you could have shipped yourself.

Buying is right more often than engineers like. Building is right more often than vendors admit. The only way to know is to price all three paths over five years and read the result honestly.

Data governance, before rather than after

Which systems a model may read, which it may write to, what leaves your infrastructure, what is retained by whom, and who is accountable when an output is wrong. Agreed and documented before the first integration ships.

This is the section teams want to skip and the one that ends AI programmes when it is skipped. In healthcare, insurance and financial services it is also the section a regulator will ask about.

A pilot you can actually kill

Every pilot ships with a baseline recorded before launch, a success threshold agreed in advance, and a date on which it is judged. That makes "it is working" a number rather than a feeling, and it makes cancelling a normal outcome rather than an admission of failure.

Roughly a third of the use cases that reach scoring do not survive it. That is the service working.