Ask a business owner how their AI work is going and you often get a version of the same answer: "We're piloting a few things." Ask when the pilot started and it turns out to be last winter. Ask what happens next and the room gets quiet.

That is the most common failure mode in small-business AI right now, and it does not look like failure. Nothing crashed. Nobody wasted a huge budget. A couple of people are getting real value from a tool. It simply never became part of how the company works, and it never will, because nobody defined what would have to be true for it to graduate.

A pilot exists to answer a question, not to run indefinitely

The point of piloting is to spend a little to de-risk a bigger decision. That only works if you wrote down the question first. "Does AI help us?" is not a question — it has no answer that would change anything. "Can we cut proposal turnaround from four days to one without the quality dropping?" is a question. It has a number, a comparison, and a deadline.

If your pilot has no such sentence attached to it, write one today, retroactively. It will take fifteen minutes and it will immediately tell you whether the last six months produced evidence or just activity.

Three things every pilot needs before it starts

An owner — one named person, not a committee and not "the team." Production criteria — the specific result that means yes, defined before you see any results, so you cannot move the goalposts to protect a tool you've grown fond of. And a decision date — a calendar day when someone says ship it, kill it, or run one more round with a stated reason.

Notice that none of these are technical. The reason pilots stall is almost never the model.

Killing one is a good outcome

Businesses treat a shut-down pilot as a loss, so they keep marginal ones alive to avoid admitting anything. That is backwards. A pilot that ends in a clear no cost you a few weeks and bought you certainty. A pilot that never ends costs you a subscription, a slice of attention, and the illusion that the question is being handled.

The honest caveat

Some work genuinely needs longer than a quarter — anything touching regulated records, client money, or safety should move slowly. The fix there isn't a faster timeline. It's a checkpoint date where you say out loud what you have learned so far and what is still unresolved.

This week

List every AI tool in use at your company. For each one, write the owner, the criteria, and the decision date. Anything you can't fill in is not a pilot. Decide whether to make it one or to turn it off.