Manufacturing
The most valuable systems on the floor are not on the network.
Some of the most valuable systems in a manufacturing operation exist in the heads of experienced operators and engineers. They are undocumented, unversioned, and scheduled to retire — and no MES upgrade addresses that.
A veteran supervisor knows which combination of readings warrants stopping a line, and which looks alarming but resolves itself. That judgment is real, it is worth money, and it is not written anywhere. When it leaves, the organization discovers how much of its operating capability was resident in individuals.
Where the cost shows up
How can manufacturers use AI in operational workflows?
By encoding the judgment experienced operators already apply and running it inside real plant constraints — safety protocols, union agreements, equipment limitations — with a supervisor approving before action and traceability running from sensor reading to that approval. The constraint set is part of the decision, not context around it.
Credibility on a plant floor is spent once. A system that recommends something unsafe or contractually impossible is discounted permanently, whatever its accuracy on the metrics it was trained against. This is the main reason plant AI fails, and it is a design problem rather than a modelling one.
Evidence
A West Coast manufacturer built a human-in-the-loop production supervisor AI capturing more than twenty years of operator judgment, with recommendations operating inside plant-floor constraints and full traceability from sensor data through supervisor approval. Separately, a manufacturer deployed explainable AI pricing under controlled governance, aligning sales, finance and operations on one methodology with human approval for strategic accounts. Both are in the case library.
What makes manufacturing different to work in
Physical consequence. A wrong call damages equipment or hurts someone, so autonomous action is rarely the right design.
Union agreements and safety protocols are hard constraints, not preferences to optimize against.
Equipment heterogeneity means a model validated at one plant may not transfer to another — treating it as portable is a common and expensive error.
The people whose judgment is being captured are reasonably wary of why. An engagement that has not answered that question honestly will not get the knowledge it came for.
Common questions
Will this replace our experienced operators?
No, and designing it that way tends to fail. The deployed model keeps a human approving. The objective is that an operator's judgment remains available after they retire, and that less experienced staff decide closer to the standard your best operator sets.
How do you capture knowledge from someone about to retire?
At the point of decision rather than in an exit interview. Tacit judgment resists documentation — the holder often cannot articulate it on request — but it is visible in the decisions they make, and the pattern emerges across many of them. See knowledge intelligence.
Does this work across multiple plants?
The captured judgment and the governance model transfer. The specific thresholds frequently do not, because equipment and process differ, and validating that is part of the work rather than an afterthought.