Why most AI pilots never reach production — and the five things that do
We looked back at every engagement we've run. The projects that shipped had less to do with the model than with evaluation, ownership and scope.
By Mohan Pasappulatti
Every company we talk to has run a pilot. Far fewer have run one that made it into the work. The gap is not where people expect it to be: it is almost never the model.
What actually stops a pilot
Looking back across our engagements, the projects that stalled shared a short list of causes — and none of them were about capability.
- No one agreed what good looked like before the build started.
- The pilot ran on a copy of the data, and the real data was worse.
- It had a sponsor but no owner, so nobody carried it past the demo.
- The scope was a department, and the decision rights were a company.
- There was no way to tell whether it was still working a month later.
The question is not whether the model can do it. It is whether anyone will notice when it stops.
What the ones that shipped did differently
The projects that reached production started from the outcome and worked back to the technology — and they wrote down how they would measure it before anyone opened an editor.
This piece is placeholder writing; the full version is still being drafted.