Why Industrial AI Pilots Die on the Plant Floor
A vendor demos a beautiful AI dashboard in a conference room. The charts animate, the executives nod, the pilot gets approved. Six months later the tool has four logins — all from the same office, none from a control room. Somebody quietly writes “results inconclusive” and moves on. I have watched a version of this more times than I can count, and the autopsy almost never finds a technical cause. The model worked fine. The tool just never made it onto the floor.
It was built for the org chart, not the shift
The person who is supposed to use it is mid-task — gloves on, an alarm going, a window of maybe ninety seconds. If the tool asks for a login they do not have, a form with six fields, or a tab they have to go find, it loses to the radio and the operator who has done this for fifteen years. Industrial AI does not get rejected in a meeting. It gets rejected in the moment it adds a step.
It never earned trust
An operator who has been right for twenty years will not hand judgment to a black box that is confidently wrong even once. One bad answer during a real upset and the tool is finished — not because it is usually wrong, but because in that environment “usually” is not good enough. Trust is earned by being narrow, being right inside that narrow lane, and saying “I'm not sure — check with the reliability engineer” instead of bluffing. Most tools are built to sound impressive. The ones that survive are built to be trusted.
It was designed for the day shift
Tools get designed around the people in the room for the demo — daytime, office, salaried. The night shift, the weekend crew, the contractor who showed up for the shutdown are the ones who actually need a fast answer, and they are the ones nobody designed for. If a tool only works for the people who were in the pilot meeting, it is not a deployment. It is a screenshot.
It was a science project, not a workflow
A lot of industrial AI is a capability looking for a home. It can do something clever, but it does not live anywhere in an actual task. The tools that stick do one specific job at one specific moment of decision — and then get out of the way. Cool is not the bar. Used is the bar.
What actually works
None of this is about better models. It is about building for the moment of decision instead of the org chart, fitting the workflow that already exists instead of demanding a new one, earning trust by being narrow and honest, and measuring adoption — who is actually using it, on which shift — instead of counting features. The unglamorous truth is that the hardest part of industrial AI was never the AI. It is the floor — which is, not coincidentally, the part most AI people have never stood on.
That is the whole reason I build the way I build: I start with the workflow and the person, not the model. If that is the problem you are wrestling with, it is most of what I do.