Beyond the announcements, a handful of applications genuinely survive the shop floor. They share one trait: a narrow scope.
We see a great many industrial artificial intelligence projects. Very few survive the move into production. Those that hold up share one characteristic: they solve a narrow problem whose expected answer can be checked immediately.
1. Visual inspection of surface defects
This is the most mature use. A camera, controlled lighting and a few thousand annotated images are enough to detect scratches, missing features or contamination with a consistency no operator sustains over an eight-hour shift. The hard part is not the model but the lighting, and the retraining procedure when the product changes.
2. Predicting setting drift
On a stable, well-instrumented process, a model learns the relationship between machine parameters and product characteristics. It warns the setter that a dimension is about to leave tolerance before the first reject appears. The value lies in the warning time, not in absolute accuracy.
3. Searching technical documentation
Finding the right instruction in three hundred pages of documentation, in plain language, saves maintenance teams real time. This use does not touch process control, which considerably reduces both risk and validation effort.
What does not work yet
Closed-loop process control by machine learning remains premature in most contexts: liability in the event of a defect is unresolved, and processes drift faster than models are retrained. We advise keeping a human in the decision loop until that drift is under control.
Back to the blog