Sampling Inspection and AQL —— A Complete Guide from Pass Rate Determination to Sampling Plan Design

In the daily work of quality management in manufacturing, sampling inspection is one of the most common yet easily misunderstood processes. Many quality engineers understand AQL (Acceptable Quality Level) as "inspect a few, check them, and release if they pass," but they often overlook the rigorous statistical foundation and the rich design space behind sampling inspection. This article will start from the basic principles of sampling inspection, systematically explain the methods for determining AQL, the logic for selecting sampling plans, and common pitfalls in implementation, helping quality managers fundamentally understand the core question of "how many to sample and how to judge them."

1. Basic Logic of Sampling Inspection: Why Not Inspect Everything?

Before discussing sampling plans, the first question to answer is: why not inspect all products? Full inspection (100% inspection) seems the safest, but in actual production, it faces three insurmountable obstacles. First, when inspection is destructive, full inspection means all products will be damaged, which is clearly unacceptable. For example, the deployment test for car airbags and the tensile strength test for materials can only be inferred through sampling. Second, for large-scale production, the cost of full inspection is prohibitively high. A production line that produces 100,000 electronic components per day would require inspection labor and equipment investment far exceeding the cost of the products themselves if each item were inspected individually. Third, and the most easily overlooked point, is that humans are not machines. Long-term repetitive full inspection work can lead to inspection fatigue, increasing the rate of missed inspections. Studies have shown that in visual inspection tasks, the detection rate of full inspection is often only 80%~85%, whereas properly designed sampling inspection combined with statistical process control can provide more stable quality assurance.

Therefore, the core goal of sampling inspection is not to "find every defective product," but to make a statistically confident judgment about the quality level of the entire batch with the minimum inspection cost.

2. The Concept of AQL: What is "Acceptable Quality Level"?

AQL (Acceptable Quality Level) is the most central concept in sampling inspection, but it is also the most misunderstood. Many practitioners interpret AQL as "the allowed defect rate," believing that as long as the batch defect rate does not exceed AQL, the batch is considered acceptable. This understanding is not entirely accurate.

According to the international standard ISO 2859 (corresponding to the national standard GB/T 2828.1), AQL is defined as "the poorest level of process average quality that is still considered acceptable in a series of consecutive submitted batches." In simpler terms, AQL is a process quality standard, not a single batch judgment standard. When the supplier's process average defect rate is equal to or better than AQL, the sampling plan will accept the batch with a high probability (usually around 95%). When the process average defect rate is worse than AQL, the rejection probability will rise sharply.

From a statistical perspective, AQL corresponds to the baseline point of the producer's risk (α). In a typical attribute sampling plan, when the actual batch defect rate equals AQL, the probability of the batch being accepted is approximately 0.95—meaning that even if the supplier's quality level just meets AQL, about 5% of the batches will still be "misjudged" and rejected, which is the risk borne by the producer.

3. The Three Elements of a Sampling Plan: Sample Size, Acceptance Number, and Rejection Number

A complete attribute sampling plan is uniquely determined by three parameters: sample size n, acceptance number Ac, and rejection number Re. For example, the plan (125, 3, 4) means randomly drawing 125 samples from the batch. If the number of defective items ≤ 3, the batch is accepted; if the number of defective items ≥ 4, the batch is rejected.

A key understanding here is that the sample size is not directly proportional to the batch size. A common mistake for beginners is "a batch of 10,000 pieces, 5% is 500 pieces," but statistics tell us that the sample size is determined by the required quality resolution (i.e., the shape of the OC curve), not the batch size. When a batch is sufficiently large (e.g., more than 10 times the sample size), the batch size itself has almost no impact on the OC curve. This is why in the ISO 2859 sample size code table, the sample sizes for batches of 3201～10000 and 10001～35000 can be exactly the same—the sample size is only related to the code and inspection level, not proportional to the batch size.

The rejection number Re is usually equal to Ac + 1, meaning that once the number of defective items exceeds the acceptance number, the batch is deemed nonconforming. However, in double or multiple sampling plans, the judgment rules are more complex, allowing for a second sample to be drawn if the first sample results are inconclusive.

4. How to Choose the AQL Value: Balancing Product Quality and Cost

Determining the AQL is the most critical decision-making step in the design of a sampling inspection plan, directly affecting the strictness and cost of the inspection. The selection of AQL values requires a comprehensive consideration of the following factors:

First, the importance of the product and the functional safety level. For key characteristics that affect personal safety (such as critical dimensions in car braking systems or sterility indicators in medical devices), AQL should be set very strictly, typically at 0.01～0.065. For general functional characteristics, AQL can be set at 0.1～0.65. For non-functional characteristics such as appearance, AQL can be relaxed to 1.0～6.5.

Second, the historical quality performance of the supplier. When the supplier's process capability is stable and the defect rate has been consistently below the target AQL, the AQL can be appropriately relaxed. Conversely, for new suppliers or those with unstable historical performance, a stricter AQL should be set and combined with tightened inspection.

Third, the balance between inspection cost and the consequences of defects. The stricter the AQL, the larger the required sample size, and the higher the inspection cost. However, a too lenient AQL can lead to defective products entering the production line, causing rework, production stoppages, or even customer complaints. From a quality economics perspective, the optimal AQL point is the quality level that minimizes the sum of "inspection cost + defect loss."

During the product development stage, AQL is often determined by the design department based on product characteristics and customer requirements; in the mass production stage, the quality department should regularly review the applicability of AQL and dynamically adjust it based on supplier performance and customer feedback.

5. Three Modes of Inspection Levels: Normal, Tightened, and Reduced

ISO 2859 specifies three inspection levels: normal inspection, tightened inspection, and reduced inspection. The switching between these three levels forms the "dynamic adjustment mechanism" of sampling inspection.

Normal inspection is the default state, suitable for suppliers with stable quality levels. When multiple consecutive batches are rejected, or the process average defect rate significantly deteriorates, the inspection should switch to tightened inspection. Tightened inspection increases the sample size or lowers the acceptance number to raise the rejection probability, putting pressure on the supplier to improve quality. Interestingly, the rule for switching from normal to tightened is very clear: if 2 out of 5 consecutive batches are rejected (or more strictly, 2 out of 5 batches are judged nonconforming under normal inspection), the inspection should immediately switch to tightened.

Reduced inspection is applicable when the process quality has been consistently stable and significantly better than AQL. The sample size for reduced inspection is usually 40%~60% of the normal inspection, effectively reducing inspection costs. However, reduced inspection has strict conditions: at least 10 consecutive batches must be accepted, and the process average defect rate must be significantly below AQL, with the production process in a statistically controlled state. If a batch is rejected during reduced inspection or if production anomalies occur, normal inspection should be immediately resumed.

This "normal—tightened—reduced" transfer rule is not just an inspection strategy but also a quality signaling mechanism—suppliers can perceive the customer's evaluation of their quality level in real-time through changes in inspection strictness.

6. Types of Sampling Plans: Single, Double, and Multiple Sampling

From the perspective of operational complexity, attribute sampling plans have three main types. Single sampling plans are the simplest, where a single sample is drawn and a direct judgment is made, suitable for scenarios with low inspection costs or short inspection cycles. Double sampling plans allow for a second sample to be drawn if the first sample results are inconclusive, making a joint judgment. The advantage of double sampling is that the average sample size is smaller than single sampling under the same quality resolution, suitable for scenarios with high inspection costs or destructive testing. Multiple sampling plans further extend this idea, allowing up to seven samples to be drawn, with an even smaller average sample size but significantly increased management complexity.

For example, with AQL=1.0, inspection level II, and a batch size of 10,000 pieces: the single sampling plan is (200, 5, 6), meaning 200 pieces are drawn each time; the double sampling plan is the first sample (125, 2, 5) and the second sample (125, cumulative 250, 6, 7), meaning that in most cases, only 125 pieces need to be drawn to make a judgment, and a second sample is only required if the number of defective items in the first sample falls into the "gray area" (3～4 pieces).

The core consideration in choosing which plan to use in practice is the balance between the management cost of the inspection operation and the average sample size. For automated online testing, single sampling is more convenient; for laboratory testing or outsourced testing, the average sample size advantage of double and multiple sampling plans is more prominent.

7. OC Curve: Understanding the Discrimination Power of Sampling Plans

Each sampling plan has its unique operating characteristic curve (OC curve), which describes the functional relationship between the actual batch defect rate and the probability of batch acceptance. The OC curve is the most powerful tool for evaluating and selecting sampling plans.

An ideal sampling plan should have an OC curve that is close to 1 (high probability of acceptance) at AQL and close to 0 (high probability of rejection) at LTPD, meaning the curve is steep and the transition zone is narrow. However, in reality, sampling plans are limited by sample size, and the OC curve always has a "gray area"—where the defect rate is between AQL and LTPD, the acceptance probability is neither high nor low, and the judgment results have significant uncertainty.

Understanding the OC curve has three layers of significance for quality managers. First, it helps you intuitively assess the discrimination power of the sampling plan during design—the steeper the curve, the stronger the discrimination power. Second, it reveals the marginal benefits of increasing the sample size: increasing the sample size from 50 to 100 significantly improves discrimination power; but increasing it from 200 to 400 shows a markedly reduced improvement. Third, it establishes a "quantitative game" consensus between suppliers and customers—both parties can negotiate AQL and LTPD based on the OC curve rather than subjective judgment.

In actual work, using the standard tables in ISO 2859 or online calculators to draw OC curves is a basic skill that every quality engineer should master.

8. Common Misunderstandings and Practical Suggestions

In the years of guiding manufacturing companies to implement sampling inspection, the following five misunderstandings are the most common.

Misunderstanding One: Not sampling small batches. Many companies believe that batches smaller than 100 pieces do not need to be sampled or only symbolically sample a few pieces. In fact, ISO 2859 has clear sampling plans for small batches, and batches of 2～8 pieces also have corresponding sample size codes. The key is not the batch size but the economic assessment of the inspection—if the inspection cost is too high, a zero-defect sampling plan (c=0 plan) or a relaxation strategy based on historical data can be considered.

Misunderstanding Two: AQL is set and never changed. Quality is dynamic, and AQL should also be dynamic. When suppliers continuously improve and process capability significantly increases, adhering to the original strict standards not only increases costs but also damages the supplier-customer relationship. It is recommended to conduct an AQL review every quarter or every six months, dynamically adjusting based on historical inspection data and supplier performance evaluation results.

Misunderstanding Three: Using the batch allowable defect rate (LTPD) instead of AQL. LTPD is the baseline point of the consumer's risk (β), indicating the defect rate at which the sampling plan should reject the batch with a high probability. Confusing AQL and LTPD can lead to the loss of direction in sampling plan design. Simply put, AQL is the "standard for good batches," and LTPD is the "standard for bad batches," and the gap between the two determines the discrimination power of the sampling plan.

Misunderstanding Four: Non-random sampling. This is the most basic and easiest mistake to make—quality inspectors tend to pick "easy to handle" or "visually good" samples rather than true random samples. This directly undermines the statistical foundation of sampling inspection, making all judgment results lose their confidence. The solution is to use random number tables or computer-generated random position codes and regularly verify whether the quality inspector's sampling operations comply with the principle of randomness.

Misunderstanding Five: Ignoring the zero-defect sampling plan. In high-risk fields such as the automotive industry (IATF 16949) and medical devices, due to the pursuit of zero defects, c=0 sampling plans are becoming increasingly popular. In a c=0 plan, the acceptance number Ac = 0, meaning no defective items are allowed in the samples. This plan appears extremely strict, but in actual application, it requires selecting the corresponding sample size for different AQL levels to avoid insufficient discrimination power due to a small sample size.

The core of sampling inspection is statistical inference, not a substitute for full inspection.

Knowledge code: 6.2.2

Version: v20260705

Author: Quality Think Tank Quality Think Tank is dedicated to providing systematic professional knowledge, methodologies, and practical tools for quality management practitioners, helping companies continuously improve their quality capabilities.