Why AI pilots stall

Four things to check before you fund one. All four are visible before anyone writes code.

We see a lot of stalled pilots. The causes repeat. We now check for four things before funding one of our own, and we'd suggest you do the same.

1. Nobody agreed what "working" means

The demo looked good. Then someone asked how accurate it was and there was no number, because there was no evaluation set. Without one, every stakeholder carries a different bar in their head and the project can never be called done. Write down the metric, the threshold, and the data it's measured on before you fund anything. If that's hard, that's your answer.

2. The demo data wasn't the real data

Demos run on clean samples. Production has missing fields, three date formats, scanned PDFs at an angle, and a long tail nobody mentioned. The quality drop between the two is where most pilots die. Pull a random sample of real inputs and run the demo on that before anyone sees a slide.

3. Nobody did the math

A pilot that costs more per case than the person it was meant to help isn't a pilot. It's an expense. Model cost, review cost, and the engineering time to keep it alive should be estimated against the value per case before the build starts. The numbers will be rough. Rough is enough to kill the bad ideas.

4. Nobody owned it after the demo

The team that built it moved on. The team that was meant to use it never asked for it. A system without an owner doesn't reach production no matter how well it works. Name the owner on day one, and make sure they were in the room when the problem was chosen.

What to do with this

None of these checks require building anything. They take a few days and a willingness to hear "not yet." That's most of what a feasibility study is.