How Many Items Do I Sample From an Incoming Lot, and How Many Defects Do I Allow — Inspection Cost vs Customer Complaint Trade-off?

Editor’s note

This problem is a statistical quality control problem; not equipment-side maintenance (#009) but the statistical-sample acceptance/rejection of incoming items. On the floor it is known as “incoming quality control”, “lot inspection” or “quality sampling”. In the academic literature it is the Acceptance Sampling Plans family, sitting within the broader Statistical Quality Control; Dodge and Romig’s 1959 Sampling Inspection Tables is the foundational table source, Schilling and Neubauer’s 2017 Acceptance Sampling in Quality Control is the modern reference, Wald’s 1947 Sequential Analysis is the foundation of dynamic decision theory, and Montgomery’s 2019 Introduction to Statistical Quality Control is the standard textbook for the OC curve and process-quality integration. The international standards ISO 2859 (general acceptance sampling procedures) and ANSI/ASQ Z1.4 (attribute plans, switching rules) are the backbone of industry practice.

Distinction from #009 (Preventive Maintenance): Planned maintenance before equipment fails — an on-machine decision. Acceptance Sampling is a statistical-sample accept/reject decision on incoming parts/lots — a quality-inspection statistical decision. The two problems sit side by side in the same sector (manufacturing) but on different layers: PM keeps the machine running, AS balances the producer-consumer risk on incoming parts.

Most-skipped point in the sector: the OC curve (Operating Characteristic curve). The practitioner’s intuition “take the AQL 1% plan” does not count the plan’s actual protection profile (consumer risk + producer risk + AOQL). The OC curve is the function that maps the true lot defect rate p to the plan’s acceptance probability Pa(p); at p=0 Pa=1 (a perfect lot is always accepted), at p=1 Pa=0 (a fully defective lot is always rejected), and an S-curve in between. The numerical basis of plan design: AQL = p_α (Pa = 1−α, typically 95% acceptance), LQL = p_β (Pa = β, typically 10% acceptance). The academic approach draws the OC curve for every plan and compares it to the expected quality; the practitioner’s intuition only looks at the (n, c) pair, no curve is produced, the protection profile remains a non-numerical claim. Schilling and Neubauer (2017) and Montgomery (2019) treat the OC curve as the central object in any textbook treatment; in vendor selection, the question “demo: under an AQL 0.65% plan, what is the acceptance probability for p=3%” tests whether OC curve support is real.

Second skipped point: switching rules and supplier-based plan calibration. ISO 2859 and ANSI Z1.4 do not give a single plan but a three-plan-one-system: normal (start, average compliance), tightened (auto-tighten if 2 of the last 5 lots were rejected), reduced (loosen if all of the last 10 lots passed + the supplier’s process is stable). Supplier-based switching avoids unnecessary cost on reliable suppliers + auto-raises protection on worsening suppliers. The practitioner’s “use the same plan everywhere” intuition skips the switching-rule machinery → total inspection cost stays 30-50% high, quality trend goes untracked.

Third skipped point: the lot-homogeneity assumption. The single-sampling formula assumes the lot is single-process, single-shift output (hypergeometric or approximate binomial distribution). In practice, suppliers mix lots from different shifts, machines and raw-material batches into one “lot” delivered to the customer; the defect distribution is then not concentrated but segmented. The OC curve gives biased results; the actual consumer risk under the plan is higher than computed. Fix: a “single-process, single-shift lot” clause in the supplier contract, or move to stratified sampling (segmented sampling), or integrated variables sampling with process capability indices (Cp, Cpk).

A step-by-step path for an SMB

Stage 1 — Measure first, plan later. At least 12 months of lot history: for each supplier-part combination the lot size, sample size (n), defects found, accept/reject decision, post-acceptance defects observed in the field (including customer complaints). Defective-part cost line items: hourly inspection cost, post-acceptance defect cost (return + contract penalty + production downtime), rejection cost (re-inspection + supplier negotiation + delivery lag). Supplier-performance inventory: accept/reject ratio in the last 50 lots, AOQ trend, Pareto of defect types. Quality target contract: AQL, LQL (written in contract), producer/consumer risks (α, β).

Stage 2 — Build the knowledge capital. Supplier-based quality trend (CUSUM or simple control chart). OC curve for current plans (a binomial approximation is enough to build intuition). AOQL calculation: upper bound on average outgoing quality after the accept-reject balance. ASN calculation: inspection-cost savings from moving single → double / sequential. Switching-rule thresholds: tightened trigger from the last 5/10 lots, reduced trigger from the last 5/10 lots.

Stage 3 — Pilot. 12-20 weeks. In a sub-set (e.g. the 3 highest-volume suppliers and 5 part types) run a systematic plan: AQL and LQL in the contract, switching rules active, OC curve drawn for every plan, AOQ trend reported weekly. Decisions stay with the quality team, the system recommends. Success criterion in writing, beforehand: total inspection cost down 15% minimum, customer complaints down 20%, AOQ variance down 30%.

Stage 4 — Rollout. 6-12 months to scale across all supplier-part combinations. Double / sequential sampling is switched on (for high-volume, stable suppliers). Bayesian-economic plans become a later option (per-lot economic decision). Quarterly quality committee: AOQ trend, switching-rule distribution, supplier performance score, economic impact report.

Risks — what can go wrong

1. Supplier quality drift (drifting AQL). The supplier starts at the contracted AQL (0.65%), the process drifts over time (new operator, material change, machine wear); the drift is slow, so even with a reduced switching state it goes unnoticed → AOQ steadily degrades, customer complaints suddenly spike. Fix: CUSUM or EWMA control chart for supplier-based trend tracking, automatic switching-rule trigger, quarterly supplier process capability (Cp, Cpk) report.

2. Reject-rate “rewarding” misuse. Some operations score the quality team on “number of lots rejected” — the team develops a bias toward unnecessary rejection, producer risk α inflates, the supplier relationship deteriorates, re-inspection cost balloons. Fix: score the quality team not on AOQ + customer-complaint cost but on a “correct accept + correct reject” ratio + economic impact.

3. Lot-homogeneity assumption breaks. If the supplier mixes shifts/machines/raw-material batches into one “lot”, the OC curve gives biased results; the actual consumer risk under the plan is higher than computed. Fix: a “single-process, single-shift lot” clause in the contract, move to a stratified sampling model, or integrated control with variables sampling (process capability indices).

4. Single-vendor lock-in — QMS. Without a contract clause for “annual standard-format export (CSV, JSON, Parquet) of plan decision history, OC curve parameters, switching-rule change log, supplier performance data, AOQ trend”, leaving the system means losing the operation’s quality memory — 3-7 years of supplier process calibration, switching-rule history, and customer-complaint-defect linkage lock at one supplier.

Solution method — technical view

Objective function choice:

• Objective 1 — Minimum total inspection cost: Classic plan design; minimise ASN subject to α and β constraints.
• Objective 2 — AOQL (Average Outgoing Quality Limit) bound: Dodge-Romig’s original approach — guarantee the upper bound on average outgoing quality after acceptance.
• Objective 3 — Minimum economic total cost (Bayesian-economic): Minimise the expected sum of inspection + acceptance + rejection costs under a prior distribution.
• Objective 4 — Minimum customer-complaint risk: Tight protection through the consumer risk β, looser producer risk α.

Multi-objective: weighted sum (inspection + AOQL + customer complaints), hierarchical (regulatory AQL first, then economic balance), or Pareto front (quality team choice).

Plan variants — pick by the field:

• Single sampling (Dodge-Romig 1959): Single stage, simplest. Low-mid volume, simple attribute inspection.
• Double sampling (Dodge-Romig 1959): Two stages, ASN 20-40% lower. Mid-high volume, stable supplier.
• Multiple / sequential: multi-stage or item-by-item decision. High volume, fast-decision setting.
• ISO 2859 / ANSI Z1.4 switching plans: Industry standard approach, supplier-based calibration via switching rules.
• Variables sampling: Integrated with process capability indices (Cp, Cpk), using continuous measurement — same protection with smaller samples.
• Bayesian-economic plans (Pearn-Wu 2007): Prior distribution + per-lot economic decision. High-value lots.
• Rectifying inspection: Rejected lots are 100%-tested and accepted — guarantees post-acceptance AOQ.

Academic references

Listed in the sources block of the page frontmatter. Dodge and Romig (1959) foundational tables; Wald (1947) sequential analysis; Schilling and Neubauer (2017) modern standard reference; Pearn and Wu (2007) Bayesian-economic plan; Montgomery (2019) SQC textbook.