How do you calculate AI inspection sample size?

Multiply validation runs by parts per run for the planned total, scale by system availability for the realistic sample, then split by conforming proportion. The defaults give 1,000 planned parts and 704 conforming after losses.

Why account for system availability in sample size?

If the cell is only available 88% of the time, you lose parts to downtime, 120 in the example, so planning on the raw 1,000 overstates what you actually inspect. Availability turns the plan into a realistic count.

How many defective parts do I need to validate detection?

Enough to estimate detection rate per defect type with confidence, often dozens per type. The example yields 176 defective parts, which must be checked against how many defect classes you need to cover.

What if I don't have enough defective parts in the sample?

Raise the number of runs, seed the sample with known defective parts, or lower the conforming proportion to reflect a deliberately defect-rich validation set. A naturally clean line rarely produces enough natural defects.

What's a realistic conforming proportion to use?

Use your line's actual first-pass yield if validating on natural production, or a planned mix if you are seeding defects. The 80% default leaves a healthy 176 defective parts out of 704 conforming.

Does higher availability always mean a better validation?

It means fewer parts lost to downtime and a sample closer to plan, but it does not change the conforming/defective split. You still need enough defective examples regardless of uptime.

Machine Vision & Industrial Inspection AI calculator

AI Inspection Sample Size Calculator

AI Inspection Sample Size estimates how many parts a vision-system validation campaign will actually yield once you account for uptime and the conforming-versus-defective split. Quality and automation engineers use it when planning acceptance runs for a new AOI cell, because the headline plan (runs times parts per run) overstates what you really capture after downtime. Knowing the realistic conforming and defective counts up front lets you confirm you will have enough defective examples to validate detection, not just a pile of good parts. The calculator chains planned volume, availability, and conforming proportion into the numbers you can actually report on.

What this calculator does

Estimate the total number of parts needed to conduct a statistically meaningful AI inspection model validation, based on required validation runs, parts per run, system availability, and the expected proportion of non-defective parts in the sample.
Use it when planning an AI inspection model validation study and you need to know how many parts to schedule and what proportion of those parts should be conforming versus seeded defectives.
It computes total planned parts, the realistic sample after availability losses, and the conforming and defective parts within that sample.

Formula used

Total planned parts = validation runs x parts per run
Adjusted sample (for availability) = total planned parts x system availability
Conforming parts in sample = adjusted sample x conforming proportion

Inputs explained

Number of validation runs planned:
Parts inspected per validation run:
Vision system availability during validation:
Proportion of conforming parts in the sample:

How to use the result

Use it while scoping a vision validation or acceptance campaign, before committing line time, to confirm the run yields enough of each part type.
It assumes a constant conforming proportion and a flat availability factor; real runs cluster downtime and defects, so treat the splits as planning estimates.

Common questions

How do you calculate AI inspection sample size? Multiply validation runs by parts per run for the planned total, scale by system availability for the realistic sample, then split by conforming proportion. The defaults give 1,000 planned parts and 704 conforming after losses.
Why account for system availability in sample size? If the cell is only available 88% of the time, you lose parts to downtime, 120 in the example, so planning on the raw 1,000 overstates what you actually inspect. Availability turns the plan into a realistic count.
How many defective parts do I need to validate detection? Enough to estimate detection rate per defect type with confidence, often dozens per type. The example yields 176 defective parts, which must be checked against how many defect classes you need to cover.
What if I don't have enough defective parts in the sample? Raise the number of runs, seed the sample with known defective parts, or lower the conforming proportion to reflect a deliberately defect-rich validation set. A naturally clean line rarely produces enough natural defects.
What's a realistic conforming proportion to use? Use your line's actual first-pass yield if validating on natural production, or a planned mix if you are seeding defects. The 80% default leaves a healthy 176 defective parts out of 704 conforming.

Last reviewed 2026-05-12.