How to Scale a Computer Vision Pilot to Production in Manufacturing

SUMMARY
Most computer vision pilots in manufacturing stall in the integration, not the model. Scale by modeling vision data as reusable so it lands in your data lake, running detection at the edge with a compact model like RF-DETR, and handing the platform to the plants to run themselves.
Around 77% of vision AI implementations in manufacturing never make it past the pilot phase. The striking part is that this is rarely a technology failure. The models pick up the signals and generate the alerts.
What breaks is everything around the model: the disconnected systems, the teams that do not share data, and the assumption that a vision project is a software project instead of an operational change. Scaling a computer vision pilot to production is a people-and-process problem long before it is a modeling problem.
That is the throughline of a recent episode of the AI in Business Podcast, where Jeff Witt, Digital Transformation Leader at a Fortune 500 global leader in building materials and fiberglass composites, who has deployed computer vision in manufacturing across a large plant network. Jeff has watched pilots stall and pushed them through to production at scale, and his account of what actually changed is concrete enough to act on.
This article walks through what scaling a computer vision pilot to production really takes, why so many stall, and what the episode shows about getting past it.
What scaling a computer vision pilot to production means
A pilot proves a model can spot something on one line, in one plant, under conditions someone set up for it. Production means that same capability runs on its own, feeds its output into the systems people already use, and repeats across dozens or hundreds of sites without a specialist babysitting each one.
The gap between those two states is mostly integration. In Jeff's experience the camera systems sit on the manufacturing IT network, separated from the enterprise data pipelines and BI systems. A pilot can ignore th…