Article · 2 June 2026
AI on the shop floor without science fiction
Where to start an AI deployment on the floor: data, takt, a process owner. No digital-factory brochure, no POC that dies after the demo.
Most conversations about AI in manufacturing start with robots and “Industry 4.0”. Useful deployments start with duller questions: where takt is lost, where a claim costs more than a minute of downtime, and who on the shift will own the exception.
Do not start with the model
A model without a source of truth in MES or ERP is a slide. First: one process, one line, one KPI. Quality, scrap, changeovers, failure prediction — pick what you already measure, even if you measure it badly.
If you cannot say how many claims return from a given operation, you do not have an AI problem. You have a counter problem.
A POC you are allowed to kill
A good floor POC lasts 4–6 weeks and has a kill criterion. If false positives eat takt, you switch the station off. That is not failure. It is the only honest way not to deploy theatre.
Camera and PLC data are dirty. Put that in the plan, not the retro.
Human-in-the-loop is a feature
An operator who does not understand a red semaphore will route around it in a week. Show 20 frames, a defect code, a “not this” button. Those clicks become custodianship.
AI stays on the floor when it has a shift lead. Not when it has an innovation budget.
