Acceptance Sampling Plans for Medical Devices: AQL, Z1.4 & ISO 2859

Every medical device manufacturer inspects — incoming components, in-process subassemblies, finished units, and sterilization parameters. The decision that almost every quality engineer has to defend to an FDA investigator or a Notified Body auditor is deceptively simple: how many units did you inspect, and on what basis did you decide that number? Acceptance sampling is the formal statistical answer, and it is one of the most frequently cited areas in FDA Form 483 observations and warning letters because companies either skip the rationale entirely or copy a sample size with no documented justification.

This guide explains acceptance sampling for medical device manufacturing: the standards behind it (ANSI/ASQ Z1.4 and ISO 2859-1, descended from MIL-STD-105E), how the Acceptance Quality Limit (AQL) actually works, the switching-rule system that ties lots together, zero-acceptance (c=0) plans, and how all of this maps to the FDA and ISO 13485 requirement that sampling plans be written and based on a valid statistical rationale.

The Regulatory Requirement: Valid Statistical Rationale

FDA's quality system rules require sampling plans to be defensible. Historically this sat in the Quality System Regulation (QSR) at 21 CFR 820.250, which stated that "[s]ampling plans, when used, shall be written and based on a valid statistical rationale," and that manufacturers must "ensure that sampling methods are adequate for their intended use." With the Quality Management System Regulation (QMSR) — finalized January 31, 2024 and effective February 2, 2026 — FDA incorporates ISO 13485:2016 by reference. The statistical-techniques obligation is preserved through ISO 13485 clause 8.1 (monitoring, measurement, analysis, and improvement, which explicitly requires determining "applicable methods, including statistical techniques, and the extent of their use") and clause 8.4 (analysis of data), alongside design verification (7.3.8) and validation (7.3.7) requirements.

The key point FDA has reiterated in its own statistical-techniques training: FDA does not and cannot prescribe a specific sampling plan. The requirement is that whatever plan you use is statistically valid and that the rationale is documented. An investigator asked one in-vitro diagnostics manufacturer why five media plates were sampled during a leak-related recall investigation; the company had "no statistical rationale for sampling five media plates" and no work instructions on statistical techniques — a textbook 820.250(b) finding. The defect was not the number five; it was the absence of a documented rationale.

The same logic applies under ISO 13485 and during MDSAP audits: a number with no stated statistical basis is a finding waiting to happen.

What Acceptance Sampling Is — and What It Is Not

Acceptance sampling is a decision rule applied to a lot (a defined batch of product). You draw a random sample of size n, inspect it, compare the number of nonconformities (or nonconforming units) found against acceptance/rejection numbers (Ac/Re), and sentence the lot — accept or reject — accordingly. It is a screen, not a control: it does not improve quality, it sorts good lots from bad ones on a probabilistic basis.

Two properties make this practical:

1. It is cheaper than 100% inspection, especially for destructive or high-volume tests (e.g., peel strength of sterile barrier, bioburden, functional performance).
2. It carries quantified risk — you know the probability of accepting a bad lot and rejecting a good one, expressed through the operating-characteristic (OC) curve.

Acceptance sampling is not a substitute for process validation (IQ/OQ/PQ), statistical process control (SPC), or capable processes. For a validated, highly capable process, sampling confirms state-of-control; for an unvalidated or marginal process, no sampling plan rescues you. FDA expects sampling to sit on top of a controlled process, not in place of one.

The Standards: Z1.4, ISO 2859-1, and the MIL-STD-105 Lineage

Three documents form one family:

ANSI/ASQ Z1.4 and ISO 2859-1 are functionally interchangeable for attributes (pass/fail) inspection. They specify single, double, and multiple sampling plans, indexed by lot size and AQL, with a switching-rule system governing normal, tightened, and reduced inspection. The dominance is near-total in practice: an estimated ~99% of U.S. medical device companies that perform incoming inspection use some version of ANSI Z1.4, and FDA has explicitly endorsed the standard in its Medical Device Quality Systems Manual, which states that sampling plans "should be developed by qualified mathematicians or statisticians or be taken from established standards such as ANSI Z1.4."

Related but distinct standards:

• ANSI/ASQ Z1.9 / ISO 3951 — sampling by variables (measured continuous data, e.g., a dimension in mm), used when measurement is more informative and economical than pass/fail.
• Codex STAN 233 — food-sector variant.
• 21 CFR 800.20 — an FDA-specific attributes sampling plan for medical gloves, one of the rare cases where a plan is prescribed in regulation rather than left to the manufacturer.

The Acceptance Quality Limit (AQL)

The AQL is the worst tolerable process-average quality level, expressed as a percent defective (or nonconformities per hundred units). It is the quality level at which the plan accepts lots with high probability — conventionally ~95%, corresponding to a producer's risk (α) of ~5%.

Two misconceptions to avoid:

• AQL is not a guarantee that every accepted lot contains ≤ AQL% defective units. It is a process-average benchmark tied to a probability. Individual lots above the AQL can still be accepted; that is the sampling risk you accept.
• AQL is not the customer-facing quality target. The consumer's risk is protected by the Lot Tolerance Percent Defective (LTPD), also called the Rejectable Quality Limit (RQL) — the poor quality level that the plan rejects with high probability (conventionally ~90%, i.e., consumer's risk β ≈ 10%). Between the AQL and the LTPD lies the Indifference Quality Level (IQL), where the lot has a 50/50 chance of acceptance.

Typical medical-device AQL selections by defect severity:

Critical defects are commonly set to AQL 0, which collapses to Ac = 0 / Re = 1 — finding even one critical nonconformity fails the lot.

How to Read the Tables: Code Letter → Sample Size → Ac/Re

The Z1.4 / ISO 2859-1 system is a two-table lookup.

Table I — Sample size code letter. Cross-reference the lot (batch) size with the inspection level to get a code letter (A through R, plus S-1 to S-4 for special levels).

Table II-A — Master table (normal inspection, single sampling). Cross-reference the code letter (sample size) with the AQL to read Ac (acceptance number) and Re (rejection number). Arrow cells instruct you to move to the next larger or smaller plan so the Ac/Re pair matches the AQL boundary.

Worked example: a 4,000-unit lot inspected at General Level II yields code letter L. At AQL 2.5 for major defects, Table II-A gives a sample size of 200 with Ac = 10 / Re = 11. If you instead set AQL 0.65, the arrow directs you up to a larger sample size to keep the Ac/Re consistent at that tighter limit.

Single, Double, and Multiple Sampling

Double and multiple plans share the same OC curve as the equivalent single plan — they offer the same protection with fewer units inspected on average, at the cost of more bookkeeping.

Switching Rules: Normal, Tightened, Reduced

Z1.4 and ISO 2859-1 are designed for a continuing stream of lots, and the switching system is the mechanism that makes the scheme more powerful than any single plan. It tightens scrutiny when a supplier's quality degrades and relaxes it (with cost savings) when quality is consistently excellent.

Tightened inspection lowers the acceptance numbers (Ac/Re) for the same AQL, reducing the consumer's risk. Discontinuation is deliberately severe — it signals that the supplier's process is no longer acceptable under the scheme and that the producer must demonstrate improvement before resuming.

In practice, many importers and some manufacturers do not apply the switching rules rigorously. As one ASQ expert noted, you can run a plan without switching rules, but "you do run the risk of not meeting the alpha risk in the end" — the plans were designed to be used as a system, and ignoring switching forfeits the statistical protection the scheme provides. The regulatory consequence is concrete: several companies have received FDA Form 483 observations specifically for not using the switching rules while still claiming conformance to Z1.4 — a problem avoidable either by applying the rules or by having a written, risk-based procedure documenting why a fixed normal-inspection plan is used for an isolated-lot or low-volume scenario.

Zero-Acceptance (c=0) Plans

The Squeglia Zero Acceptance Number Sampling Plans (published by ASQ) are widely adopted in medical device manufacturing. A c=0 plan sets Ac = 0 / Re = 1: any nonconforming unit in the sample fails the lot. The driver is partly statistical (tight consumer protection for safety-critical product) and partly legal — historical product-liability litigation made "accept on zero" the defensible posture for devices where a single escaped defect can cause harm.

c=0 plans are not indexed by AQL in the same way as Z1.4; they are indexed by LTPD (the lot tolerance percent defective at a chosen probability, typically 5% or 10%). For a given LTPD, the plan defines a sample size such that a lot at that defect level is rejected with the stated confidence. This makes them especially useful when you want to state, "this lot has at least 95% confidence of containing fewer than X% defective units."

In Z1.4/ISO 2859-1 tables, very tight AQLs (e.g., 0.065 or below) naturally collapse to Ac = 0 via the arrow rule, so the two systems converge at the stringent end.

Operating-Characteristic (OC) Curves and Risk

Every sampling plan has an OC curve: a plot of probability of lot acceptance vs. actual lot quality (percent defective). The curve tells you:

• Producer's risk (α) — probability of rejecting a good lot at the AQL (conventionally ~5%).
• Consumer's risk (β) — probability of accepting a bad lot at the LTPD/RQL (conventionally ~10%).
• The discrimination — how steep the curve is between AQL and LTPD. Steeper = better discrimination = larger sample size.

Reading the OC curve is the most direct way to defend a sampling plan's validity to an auditor: it shows the explicit risk trade-off you accepted. When a manufacturer selects a plan, the documented rationale should name the AQL, the LTPD/RQL being protected against, the associated α and β, and why those are appropriate to the device's risk (per ISO 14971) and intended use.

Tying Sampling to Risk Management

A statistically valid plan is not automatically the right plan. The selection must be risk-based, consistent with ISO 14971 and the QMSR/ISO 13485 expectation that statistical methods be proportionate to risk:

• Higher-risk characteristics (e.g., sterility assurance, dose accuracy, leak integrity of a fluid path) warrant tighter AQLs (0.065–1.0), Level III, and often c=0 acceptance.
• Lower-risk characteristics (cosmetic, non-functional attributes) can use looser AQLs (2.5–4.0) and Level I/II to control inspection cost.
• Destructive tests often justify Special levels (S-1 to S-4) with documented acceptance of the larger sampling risk.

The risk management file should connect each inspected characteristic to its residual risk, so the sampling plan is traceable to the risk control decision — exactly what an investigator checks.

Practical Implementation Checklist

• Rationale documented. For every sampling plan, state the lot definition, characteristic inspected, inspection level, AQL (and LTPD protected), and the statistical basis (cite Z1.4, ISO 2859-1, Squeglia c=0, or a calculated OC plan).
• Risk linkage. AQL severity tied to ISO 14971 residual risk for each characteristic.
• Random sampling procedure. Method for drawing a representative random sample defined and followed.
• Switching rules applied (or a documented, risk-based decision to use fixed normal inspection with justification).
• Critical defects at Ac = 0. Critical characteristics accept on zero by default.
• Plan reviewed when changes occur — process changes, supplier changes, CAPA outcomes, or shifting defect data trigger plan review (an explicit QSR/ISO 13485 expectation).
• Records retained. Sample size, defects found, lot disposition, and any re-inspection documented per DHR/QMS records requirements.

Common Nonconformities and Warning-Letter Themes

Recurring FDA findings in this area:

1. No documented rationale for the sample size — a number chosen by tradition ("we always test five") with no statistical basis.
2. Sample size not tied to risk — identical plans for critical and minor characteristics.
3. Switching rules ignored while still claiming conformance to Z1.4 / ISO 2859-1.
4. Re-inspection without a valid resubmission rule — re-testing a rejected lot until it passes ("testing into compliance") with no documented statistical basis.
5. Plan not reviewed after changes — supplier, process, or CAPA changes that warranted a tighter plan, with no review record.

Each of these is an 820.250(b) / ISO 13485 8.1 finding and is avoidable with a one-page rationale per plan.

Key Takeaways

• FDA and ISO 13485 require sampling plans to be written and based on a valid statistical rationale — the specific plan is your choice; the justification is not optional.
• ANSI/ASQ Z1.4 and ISO 2859-1 are the dominant attributes-sampling standards (MIL-STD-105E lineage); Z1.4 is FDA-recognized (FR 5-62).
• AQL is the worst-tolerable process average; LTPD/RQL is what protects the consumer. An OC curve quantifies both risks.
• The switching-rule system (normal/tightened/reduced/discontinue) is what makes the scheme statistically powerful over a stream of lots — use it or document why not.
• c=0 (Squeglia) plans are the medical-device norm for safety-critical characteristics, indexing on LTPD for defensible consumer protection.
• Tie every plan to ISO 14971 risk and review it whenever the process, supplier, or risk profile changes.

Sources

• 21 CFR 820.250, Statistical techniques (Quality System Regulation); FDA, Statistical Techniques (CDRH presentation, fda.gov/media/160131).
• FDA, Quality Management System Regulation final rule (89 FR 7496, February 2, 2024); QMSR effective February 2, 2026, incorporating ISO 13485:2016 by reference.
• ISO 13485:2016, Medical devices — Quality management systems, clauses 7.3.7, 7.3.8, 8.1, 8.4.
• ANSI/ASQ Z1.4-2003 (R2018), Sampling Procedures and Tables for Inspection by Attributes; FDA Recognized Consensus Standard, FR Recognition Number 5-62.
• ISO 2859-1, Sampling procedures for inspection by attributes — Part 1: Sampling schemes indexed by AQL (clause 9, switching rules).
• ANSI/ASQ Z1.9 / ISO 3951, sampling procedures for inspection by variables.
• Nicholas L. Squeglia, Zero Acceptance Number Sampling Plans (ASQ Quality Press).
• ASQ Ask the Standards Experts, Zero Acceptance Number Sampling Plans and the FDA (2017); Z1.4 2008: AQL, Nonconformities, and Defects Explained (2012); Sampling Plan Review? (switching rules and annual review).
• FDA, Medical Device Quality Systems Manual (endorsement of ANSI Z1.4 for sampling plans).
• IEEE Santa Clara Valley Reliability Society, Confidence and Reliability Based Sampling Plans in the Medical Device Industry (~99% industry adoption of ANSI Z1.4).
• Taylor Enterprises (variation.com), Acceptance Sampling Update (AQL/LTPD documentation; Form 483s for not using switching rules).
• Quality Magazine, Brief History of ANSI/ASQ Z1.4 (Dodge–Romig and MIL-STD-105 lineage).
• ISO 14971:2019, Application of risk management to medical devices.
• ECA Academy / GMP-Compliance, Statistical Methods are also important for Medical Devices (21 CFR 820.250 commentary).
• Critical Manufacturing, AQL in Medical Device Manufacturing and the role of MES.