Trained inspectors miss up to a third of defects. That's why factories are handing quality control to cameras.

Trained human inspectors miss roughly 15 to 30 percent of defective parts and wrongly reject about a third of the good ones, a Sandia National Laboratories study found. The same product gets different verdicts depending on who is looking and which shift they are on. That inconsistency, not labor cost, is why computer-vision inspection is the AI quietly paying off on factory floors. But the number that actually decides whether it pays is not the detection-accuracy figure vendors put on the slide. It is the one they tend to bury.

The case against the human eye

This is not a knock on inspectors. The human visual system evolved to scan a landscape for movement, not to catch a 50-micron scratch on a metal surface flowing past at two parts a second. Under real production speed, attention drifts, shifts end, and standards wobble, which is how identical parts earn different grades. The rule-based machine-vision systems that came before AI did not fix it either; legacy automated optical inspection is prone to high false-call rates, flagging good parts as defective at rates that in some settings reach 20 to 80 percent. The result is the same cost on both ends: real defects escape to customers, and good parts get scrapped on a bad call.

AI vision changes the economics by being consistent in a way a human cannot. It inspects every part, every cycle, at full line speed, with peer-reviewed studies putting detection accuracy in the 95 to 99 percent range, and without the fatigue or drift that degrades a human inspector. Individual deployments are concrete: Intel has reported up to $2 million in annual savings from AI vision inspection.

The number on the slide, and the number that matters

Here is the metric vendors lead with: detection accuracy, the share of real defects caught. It is the impressive one, and it is not the one that decides your return. The number that decides it is the false-positive rate, because rejecting a good part costs money too. The math is unforgiving: a 2 percent false-positive rate at 1,000 units an hour throws away 20 good parts every hour, which can be more disruption than the system prevents. A system that catches 99 percent of defects and falsely rejects 5 percent of good output can destroy the savings it was bought to create.

So read both numbers together. The good systems pair high detection with a false-positive rate down around 2 percent or below, with some deployed systems near 0.1 percent, and the gap between vendors is widest exactly there. Accuracy is the headline. The false-positive rate is the invoice.

Why pilots die at the same step

The other place the money leaks is deployment. A model that scores beautifully on test images often stumbles in real production, where vibration, temperature swings, and product variations it never saw in training degrade it. This is why so many vision pilots stall between demo and line. The teams that get past it use the same method: a shadow deployment, running the AI alongside human inspectors with no authority to reject over a validation period, comparing every call, tuning thresholds, and only then handing it reject authority with humans reviewing the borderline cases. Start at one high-impact station, prove the return, and scale from a validated win rather than a company-wide rollout. Done that way, payback typically runs 12 to 18 months, and because even small yield gains compound across high-volume production, the savings can reach into the millions.

The altitude shift

Take one missed scratch and one false reject and follow each to the ledger. The scratch a tired inspector waves through on the night shift becomes a warranty claim, or a recall, downstream, where it costs orders of magnitude more than catching it would have. The good part a poorly tuned model rejects becomes scrap, and twenty of those an hour, every hour, becomes the reason the project gets cancelled at review. The technology sits between those two failure modes, and the win is not eliminating one at the expense of the other. It is consistency: holding the same standard on every part on every shift, which is the one thing the human eye, however skilled, cannot promise.

The question to ask the vendor

Before you look at the accuracy number, ask for the false-positive rate at your actual production volume and for your specific defect type, and treat a vendor who will not state it as the answer itself. Then insist on a shadow run on your line, not their test set, because real production is where the impressive demos go to break. Computer vision in manufacturing is one of the most genuinely deployed and profitable forms of AI there is. It earns that status one tuned, validated station at a time, on the metric the slide leaves off.