Industrial AI Is Making OEE Worse, Not Better

The plant manager at a Tier 1 automotive stamping facility outside Detroit had done everything right. He'd hired a systems integrator, spent $1.8 million on sensor infrastructure and an edge compute stack, trained a convolutional neural network on 18 months of production footage, and achieved what looked like a genuine technical win: 94.7% accuracy in detecting tool wear across 12 stamping presses. The model latency was 140 milliseconds per inference, well within the 500-millisecond window required for real-time intervention. By every measure that machine learning engineers measure things, the system worked.

His OEE collapsed by 3.2 percentage points in the first three months of production.