Three quiet revolutions AI is running on the shop floor

The stock question we ask on factory visits is always the same: “where are you using AI?” The answer usually begins by pointing at a screen — a large dashboard, colourful charts, current KPIs. Then we narrow the question and the picture changes: how many times a day does anyone look at that screen, and what happens after they do?

The real AI is working elsewhere in the plant. Usually quietly, and usually in places nobody called an “AI project.”

First observation: line balancing stopped being a project

Line balancing is classically an engineering project: collect data, build a model, run scenarios, deploy a new layout six to twelve months later. The result can be good — until the product mix changes and the same project is needed again.

What we see in the field is different: a recommendation engine that learns from the hourly flow of work orders keeps operator loading current during the day. Nobody calls it an optimisation project. The shift supervisor looks at the list in the morning, changes two rows, and the day starts.

Second observation: scrap is discussed before it happens, not after

The scrap report is traditionally a backward-looking document. You look at it at month end, name a percentage, hold a meeting.

Models that read operator, machine and product together reverse that order. When scrap risk rises for a particular combination of the three, the warning arrives mid-batch — not at month end.

The subtlety matters: the warning is a call to adjust, not an accusation. The difference between a system saying “this operator is bad” and one saying “scrap has risen on this combination before, check the parameter” decides whether the recommendation is accepted at all.

Third observation: the anomaly comes from purchasing

The third observation is not on the line but on the side feeding it. Breaks in supplier price curves — a line departing from its own history, a quote spread narrowing suddenly, a lead time stretching quietly — are signals the human eye misses.

None of them is a decision on its own. But the sentence “this supplier’s last three quotes do not match its own pattern” is what gets a buyer to pick up the phone. The call is still theirs to make.

What they have in common

None of the three is a separate screen. The recommendation sits where the decision is made: in the work order list, on the batch card, on the quote comparison screen. Every recommendation moved to a dashboard becomes a recommendation nobody reads.

Production AI is not a report; it is a referee embedded in the hourly decision loop. The production module takes that as the default, and our whole approach to AI rests on the same principle.