The inspection line that couldn't scale
A use case in manufacturing: a computer-vision pilot stalled nine months by one unanswered question — who is accountable when it misses a defect?
A manufacturer's quality team manually inspects thousands of parts per shift. A computer-vision pilot has shown it can catch surface defects faster and more consistently than tired eyes — but nine months later, it still isn't in production. The blocker isn't technical. Leadership keeps asking who is accountable if the system misses a defect, and nobody has an answer.
Meanwhile, AI is arriving through the back door. Two maintenance vendors have shipped equipment with embedded predictive AI features. Engineers are using generative AI to draft documentation. None of it is registered, assessed, or owned.
A discovery engagement produces the first plant-to-enterprise AI inventory: the stalled vision pilot, the vendor-embedded maintenance AI, the generative AI use in engineering, and a demand-forecasting model feeding procurement — each documented with its data, its decisions, and its risk profile.
Every system is tiered by safety, quality, and operational criticality. The vision system — classified as quality-critical — receives validation thresholds, ongoing performance monitoring, a defined human override, and a named owner in the quality organization. Vendor AI features are assessed through a standard questionnaire, and generative AI use is brought under an acceptable-use policy.
With accountability explicit, leadership can approve the vision system for production. Inspectors focus on the exceptions the AI flags rather than every part on the line — and the governance template built for this deployment is reused for each new plant-floor AI system.
What this makes possibleHigher inspection throughput, fewer escaped defects, and a stalled pilot turned into the company's model for adopting AI: discover it, govern it, then scale it. Later plant deployments follow a known path instead of reopening the same debate.


