AI in manufacturing: which use cases actually pay

Machine learning methods have been available for years and are readily accessible in libraries. What manufacturing companies lack is labelled history: scrap reasons recorded consistently, downtime reasons that do not mostly land on “other”, and timestamps that line up between MES and machine control.

Every use case therefore starts with a data-readiness assessment. It is unspectacular and it decides the outcome: a model on poor history reliably produces plausible wrong predictions.

Optical inspection is the most accessible case. The data requirement is manageable — a few hundred to a few thousand labelled images per defect class — and the effect is directly measurable. The precondition is reproducible lighting, which fails more often on the rig than on the algorithm.

Scrap prediction from process parameters requires cleanly linked process and quality data over at least six to twelve months. Where that link exists, the leverage is considerable; where it first has to be created, that is where the real project scope lies.

Remaining-life estimation on plant needs failures in the history — several per failure mode. Well-maintained plant produces too few events, which makes the case difficult at exactly the companies most able to pay for it.

Demand forecasting in procurement is computationally undemanding and organisationally difficult, because it presupposes reliable sales figures from the commercial side. The resistance rarely lies in the technology.

A model nobody owns decays. Data distributions shift with every tool change and every supplier switch. Set no owner, no monitoring of model quality and no retraining interval, and you are running a prototype in earnest.

So the concept always includes the uncomfortable question: what happens when the model is wrong? For visual inspection that means a defined second check; for a prediction, a decision rule that involves a person. Only then does a pilot become a piece of operating equipment.