- Essays··11 min read
When the Model Works and the Line Does Not
Seventy-seven per cent of AI vision pilots in manufacturing never reach full deployment — not because the detection accuracy is wrong, but because the factory floor is. This essay traces six failure modes: environments the training set never saw, integrations that ran eight times over budget, thresholds nobody will own in writing, and models that drift when procurement switches supplier. The automation thesis survives the pilot. The line does not.
ai-in-manufacturingcomputer-visionindustrial-aimodel-deploymentRead - Essays··8 min read
Sixty Picks an Hour, and the 88% That Never Graduate
Siemens' Erlangen humanoid moved sixty totes an hour at better than 90% accuracy. Across enterprise, 88% of AI pilots never reach production at all. The marquee deployments are three dozen Tier-1 buyers reading from one address book; everyone else is watching pilots die in the gap between a vendor demo and a system that still runs on a Tuesday in August.
ai-in-manufacturinghumanoid-robotspilot-to-productionindustrial-aiRead - Essays··8 min read
Erlangen, Spartanburg, and the Quiet Half of Industrial AI
The cognitive layer of industrial AI (copilots, agentic optimisation, the operating-system metaphor) is real and mostly in pilot. The physical layer of vision, manipulation and motion is the part actually booking cycle time on the factory floor. That asymmetry is the story.
industrial-aipilot-to-productionembedded-visionhumanoid-roboticsRead