Updated July 2026 · 7 min read · Adente Vision Engineering Team
What does over-rejection actually cost on a good line?
Over-rejection is the cost of throwing away good parts, and it is easy to miss because the parts look like legitimate scrap. When an inspector wrongly rejects a conforming part, you lose the material, the labor already invested in that part, and the throughput the part represented, all for a defect that was never there. On a line that is genuinely making good product, this is pure loss with nothing to show for it.
It stays hidden because a false reject and a real defect land in the same scrap bin. Nobody questions a rejected part unless someone measures how many rejects were actually good, so an inspector tuned too tight can quietly erase yield while looking like it is protecting quality. The only way to see it is to separate the two error types and cost each one. Adente Vision is an edge-AI visual inspection unit built by ADENTE Advanced Engineering Technologies, part of the Aden Group, sold through automation system integrators, and the point of this post is the number most inspection business cases skip: what your false rejects cost.
How does a high false-reject rate erode yield?
A high false-reject rate erodes yield directly, because yield is the share of good parts that make it through, and every good part rejected in error subtracts from it. Push an inspector's threshold tighter to make sure no defect escapes, and you catch more real defects, but you also reject more borderline-good parts, and those rejects come straight off your quality rate and your OEE.
The effect compounds through rework. Many false rejects are not scrapped outright; they are re-inspected, reworked or manually dispositioned, which spends operator time re-checking parts that were fine to begin with. That labor is a second cost stacked on the first, and it grows with the false-reject rate. A line can hit its escape target and still bleed money if it does so by over-rejecting, which is why a headline "we catch everything" claim deserves the follow-up question: at what false-reject rate.
Why is precision a direct lever on scrap?
Precision is the share of rejected parts that were genuinely defective, so it is the metric that maps one-to-one onto scrap waste: raise precision and a larger fraction of what you throw away was actually bad, which means less good product in the bin. Where recall protects your customer from escapes, precision protects your yield from false rejects, and a good inspector has to hold both at once rather than trading one away for the other.
That balance is the real work. An inspector that outputs a confidence score lets you set the operating point to your cost of quality rather than to a single fixed threshold, so borderline parts can be routed for a second look instead of being scrapped on a guess. As a reference for what holding both looks like, Adente Vision reached 99.65% F1 on a live cap-inspection line, and a high F1 implies the system was not buying its low escape rate with a wall of false rejects, because F1 falls if either error runs high. Your own figure depends on your parts and needs an application-specific measurement, but the principle holds: chase recall alone and precision, and your yield, pays for it.
Over-rejection versus under-rejection: where the cost lands
The two inspection errors send their costs to different places, and seeing them side by side is what keeps a line from fixing one by quietly worsening the other. Over-rejection (a false reject) destroys yield inside your plant; under-rejection (an escape) sends cost downstream to assembly, the field or the customer.
| Cost line | Over-rejection (false reject) | Under-rejection (escape) |
|---|---|---|
| Good parts scrapped | Material and invested labor lost on a conforming part | None |
| Rework and re-inspection | Operator time re-checking parts that were fine | None |
| Yield and OEE | Quality rate drops, good output falls | Masked until the escape is found downstream |
| Downstream failure | None | Cost multiplies per stage the defect travels |
| Customer PPM and trust | None | Escapes raise PPM and risk chargebacks or recall |
Scrap, rework and yield-loss figures vary widely by part and process and are an industry range, not an Adente number; fill the table with your own costs. The downstream multiplier reflects quality literature's rule-of-ten, that a defect costs roughly ten times more to fix at each stage it travels.
How do you balance false rejects against escapes?
You balance them by setting the operating point to the real cost of each error on your line, not to a default threshold. If an escape from your line triggers a recall and a false reject costs one scrapped part, the two are not equal, and the confidence threshold should reflect that asymmetry rather than sit at whatever value the system shipped with. The confidence score is what lets you tune this deliberately.
Route the grey zone instead of guessing it. Parts above the confidence threshold pass, parts clearly below are rejected, and borderline parts can go to a second check or a human review rather than being forced into a class. That keeps the escape rate low without paying for it in scrapped good parts, and it means the operating point is a business decision about your cost of quality, made with the numbers in front of you.
What should you measure before and after AI inspection?
Measure both error rates before and after, or you cannot prove the change paid off. Record the false-reject rate (or the scrap that turns out to be good on audit) and the escape rate (PPM to your customer) under your current method, then measure the same two after the new inspector is running. A drop in escapes with no rise in false rejects is a real gain; a drop in escapes bought with more false rejects is a cost shifted, not removed.
Keep the parts and conditions consistent across the before and after, so the comparison is fair, and separate false rejects from real defects in the scrap audit so the yield number is honest. For the two errors and their asymmetric cost, see the sibling post on the false-negative rate; for the method behind few-shot and anomaly detection, see the pillar guide on AI visual inspection; and to see where yield recovery sits across the tasks the unit runs, browse the real applications.