Almost every conversation about AI inspection accuracy starts with "what accuracy does the model get?" That question assumes there is one number you can compare between vendors. In practice, an inspection system makes two completely different kinds of mistake, and the cost of each can differ by an order of magnitude — so the real answer starts not with a number, but with a question: which mistake hurts more?
Two kinds of error, two different costs
A false accept is when the system lets a defective product through as OK. A false reject is the opposite — the system rejects a good product as NOK. Both lower "accuracy" in the same way, but their business consequences are usually not symmetric at all. A false accept on a critical defect can mean a complaint, a whole-batch return, or a safety issue at the customer. A false reject of a good part is usually the cost of re-inspection or lost throughput — painful, but rarely catastrophic.
That is why two systems with identical "overall accuracy" can be worth completely different things to your process, if one mostly makes the cheaper mistake and the other mostly makes the expensive one.
Why there is no single universal threshold
We do not use one decision threshold for every project. Acceptance criteria are agreed before the pilot, based on the risk and economics of the specific process — there is no single accuracy value that is correct everywhere. A false accept on a critical defect is weighed differently than a cosmetic deviation with no functional impact.
That directly shapes how we train the model in AIVQC Trainer: acceptance criteria — the target balance between false accepts and false rejects — are agreed before training, not fitted to the result afterwards. Adjusting the threshold after training is not gaming the system; it is a deliberate decision about which way the model should err, if it has to err at all.
Same mechanism, different weight depending on the defect
For cracks and damage, the false-reject rate matters most — a model that confuses natural material texture with a scratch rejects good products and generates unnecessary sorting cost. For flash and deformation, where the defect is inherently irregular, the false-accept rate matters more — whether the model lets borderline cases near the tolerance threshold through. This is not a universal rule; it is a consequence of what the specific defect looks like and what each type of mistake costs in that particular process.
That is why, before a pilot, we ask not only "will you detect this defect" but also: what costs more — a complaint over a missed defect, or the cost of rejecting and re-inspecting a good part? The answer decides which way to shift the threshold.
What the qualification report shows
That is why a pilot report never stops at a single "accuracy" number. The quality section shows recall, precision, and — separately — the false-accept and false-reject rates, measured on an independent test set the model never saw during training. A reference validation result is not a guarantee for other products or process conditions — every inspection case gets its own measurement, because both the defect and the cost of a mistake differ.
Questions worth asking before setting a threshold
- What costs more in your process: letting a defect through, or rejecting a good part?
- Does the defect you want to detect affect safety or product function, or is it cosmetic?
- Is there another check further down the process that would catch a false accept — or is AI inspection the last line of defence?
- How much does re-inspection and sorting cost if the threshold is set more conservatively?
- Should different variants of the same defect — borderline and obvious — be treated the same way, or differently?
If you are not sure which way the threshold should be shifted for your case, that is exactly what the initial consultation before a pilot resolves — based on the real cost of both types of error in your process, not a single universal accuracy number.