Inspection

Designing Inspection Sampling Plans That Actually Catch Defects

Inspect 10 percent of the lot is not a plan. How to size samples from detection probability, cut inspection load, and stop escapes at the same time.

Sampling decides money on both sides of the ledger. Inspect too much and you pay forever: a 100 percent visual check on a 40,000 unit monthly part at 20 seconds per piece is 222 labor hours, roughly $7,300 a month at loaded rates. Inspect too little and you own the escape: one defective lot reaching an automotive customer can trigger sorting fees, a chargeback of $10,000 to $50,000, and a controlled-shipping status that costs more than a year of inspection. The sampling plan is where you set that trade deliberately, with arithmetic, instead of inheriting whatever sample size somebody wrote on a router in 2011.

The core math is probability of detection. If a 1,000 piece lot contains 2 percent defectives and you sample 50 pieces, the chance of catching at least one is 1 minus 0.98 to the 50th power, about 64 percent, meaning one bad lot in three sails through clean. Push the sample to 150 and detection rises to about 95 percent. The Inspection Sampling calculator runs the detection probability and the total inspection load so you can price the tradeoff per part number. The counterintuitive rule: required sample size barely depends on lot size, so percentage rules like inspect 10 percent are statistically illiterate on both big and small lots.

Practical anchors: ANSI/ASQ Z1.4 general level II puts a 1,000 piece lot at a sample of 80; at an AQL of 1.0 that plan accepts on 2 defectives and rejects on 3. Typical AQLs run 0.65 to 1.0 for major defects and 2.5 for minor cosmetic ones; safety-critical characteristics do not get AQLs, they get capability requirements or 100 percent automated verification. Detection reality check: catching a 0.1 percent defect rate with 90 percent confidence needs a sample near 2,300, which no manual plan sustains. Below roughly 1,000 PPM, sampling stops being a control and becomes a ritual; switch to process control or automated inspection.

The levers: first, switch from lot sampling to source control; SPC with Cpk above 1.33 on the generating process outperforms any receiving-dock sample and costs less after the first quarter. Second, use switching rules: tightened inspection after 2 rejected lots in 5, reduced after 10 clean lots, which cuts steady-state inspection load 40 to 60 percent on good suppliers. Third, stratify the sample: pull pieces across the run, first-off, mid-shift, and after every adjustment, because defects cluster around changeovers and a random-from-the-box sample misses time-based problems. Fourth, automate high-volume checks; a $60,000 vision system replacing 1.5 inspectors pays back in 12 to 18 months.

Failure modes: the deadliest is sampling as theater, pulling 5 pieces from the top of every box regardless of lot size, which detects a 1 percent defect rate about 5 percent of the time. Second, AQL misread as a guarantee; an AQL 1.0 plan will accept lots at 1 percent defective most of the time by design, so never promise a customer that 1.0 means zero. Third, inspector drift: visual escape rates run 10 to 30 percent even for trained inspectors and worsen after about 30 minutes of continuous inspection, so rotate tasks hourly. Fourth, no reaction plan; a rejected lot with no containment procedure is just a delayed shipment.

Cadence: daily, log lots inspected, lots rejected, and defects found by code; a rejected lot triggers containment within the shift, bracketing back to the last known good lot. Weekly, review detection data by supplier and part family, and apply switching rules mechanically, not by mood. Monthly, compute inspection hours and cost per part family next to the escapes that beat the plan; any family with zero rejects for 6 months is a candidate for reduced inspection or a capability-based skip-lot plan. Annually, re-derive sample sizes from actual defect history; most plants find 20 to 30 percent of inspection load protects against defects that no longer occur.

World-class looks like inspection spending falling while escapes fall too, because detection moved upstream into capable processes and automated checks. Concretely: receiving inspection on under 10 percent of part numbers, the rest on certified supplier status with audited capability; manual in-process sampling only where Cpk sits below 1.33; escape rates to customers under 25 PPM; inspection cost under 1 percent of sales, against 2 to 4 percent at inspection-heavy plants. The mindset shift is the real marker: sampling is a temporary tax paid while the process gets fixed, and every sampling plan carries an expiration question, namely what would let us stop.

Published 2026-07-02.