Choosing Sampling or Full Inspection for Industrial AI: Lot Acceptance Risk, Inspection Errors and Disposition Cost

Separate part screening from lot disposition

Part screening asks whether an individual product should proceed. Lot acceptance sampling asks whether a defined lot should be accepted, isolated or inspected further under an agreed rule. NIST identifies lot disposition, rather than estimation of lot quality, as the main purpose of acceptance sampling.[1] A sample of one hundred conforming parts is therefore input to a sampling decision; it does not turn every remaining part into an individually inspected and approved product.

AI can change the cost and speed of obtaining visual evidence for each part, but it does not automatically change quality accountability. Ten thousand images may still omit an inaccessible surface, and individual model scores may share a failure under reflective conditions. Conversely, an established supplier's low-risk appearance checks may already have effective sampling and rejection handling. Additional full-inspection equipment need not create enough value to cover maintenance. This is a conditional investment judgment, not a claim about any technology's measured performance.

The existing industrial visual anomaly detection acceptance method addresses misses and review workload at a fixed threshold. Here the question moves up to lot policy. The same model might screen every part, assist interpretation of sampled parts, or provide risk alerts only. These uses require different evidence, costs and release responsibilities. Procurement documents should name the intended use rather than simply require an unspecified AI inspection accuracy.

Freeze the lot, sampling frame and disposition rules first

We recommend defining inspection lots through consistent part numbers, quality criteria and traceable production conditions. Retain lot identifiers, counts, production intervals, material or supplier lots, stations and shifts. A shipment is not necessarily an appropriate inspection lot. Mixing several origins in one truck and calculating one defect proportion may conceal a concentrated problem. If stratification is needed, design the sampling and acceptance rules for each stratum before observing results, rather than choosing favorable groups afterwards.

A sampling frame is an explicit inventory of eligible objects. Every part must map to a location or identifier and have an opportunity to be selected. A reproducible implementation freezes the object list, records the selection algorithm and seed, and preserves selected identifiers for physical retrieval. Correct random numbers do not establish correct sampling: taking accessible top-layer boxes instead of selected objects breaks the connection between the statistical model and the actual procedure. Record replacements, missing objects and duplicate selections and handle them under predetermined rules.

Also distinguish nonconforming units from individual defects. A part with both a scratch and a stain normally counts once in a plan based on nonconforming units. Three model bounding boxes are not automatically three defective parts. Quality owners should define defect classes, decision boundaries, dispute adjudication and missing-image handling. Different severities may require different policies; numerous normal examples of minor appearance issues must not dilute evidence about critical defects.

Finally, specify what rejection entails: supplier return, complete screening, rework followed by reinspection, or exclusion from the next operation. The workload calculation below assumes rejected lots receive full screening. If only returns are possible, estimate return logistics, replacement procurement and lead-time effects instead. Without isolation space and a responsible disposition owner, an attractive acceptance curve cannot deliver a reliable process.

Worked example: what evidence does a 125-part sample provide?

Everything in this example is constructed. No customer records or production experiments were used. Each lot contains N=10,000 parts and is checked for one defined, noncritical appearance nonconformity. A candidate single-sample plan randomly selects n=125 parts. Accept when the observed nonconforming count d is at most c=2; otherwise isolate the lot and inspect the remainder. Initially assume perfect inspection and conditions suitable for a binomial approximation. Neither 125 nor 2 is presented as a requirement or recommendation taken from a mandatory standard.

Let p denote the incoming true nonconforming proportion. Acceptance probability is Pa(p)=Σ[k=0 through 2] C(125,k)×p^k×(1-p)^(125-k), where C is the combination coefficient. NIST uses this cumulative binomial probability to explain the operating characteristic, or OC, curve, with a lot sufficiently large relative to the sample as an approximation assumption.[2] For a finite lot with a fixed known number of nonconforming units, a hypergeometric model is an alternative; the binomial approximation is not exact for every lot size.[4]

The calculated acceptance probabilities are approximately 97.47% at p=0.5%, 86.93% at p=1%, 54.25% at p=2%, and 4.77% at p=5%. The same plan can reject relatively good lots and accept relatively poor ones. Each point means what would happen under an assumed lot quality p. It does not mean that, after a sample passes, there is a 97.47% probability that this particular lot satisfies a quality requirement.

Suppose internal discussions designate 0.5% as a relatively good condition and 5% as unacceptable, with acceptance probability respectively required to be at least 95% and at most 10%. Under these illustrative requirements and the binomial approximation, the candidate meets both endpoints. Quality owners must still review the intermediate region and critical-defect policy. NIST's producer-risk and consumer-risk definitions help distinguish the two losses.[3] Showing the customer only poor-lot rejection performance conceals the consequences of interrupting acceptable supply.

With two nonconforming units observed, the frozen example plan accepts the lot; with three, it isolates it. A tight delivery deadline does not justify repeatedly drawing new samples until a favorable result appears. Double or sequential sampling can have planned stages, but the continue, accept and reject rules must be specified in advance and the prescribed observations combined.[3] Ad hoc sampling until a pass changes the risk originally promised.

Inspection after rejection may cost more than the initial sample

Continue with the same example. Assume inspection requires 0.5 minute per part, and the remaining 9,875 parts of every rejected lot are screened without rechecking the original sample. Average total inspection is ATI=125+(1-Pa)×9,875. This separates routine sampling from the extra work triggered by rejection.[2] Minitab's official guidance likewise recommends considering acceptance probability and total inspection together instead of judging a plan solely by its sample size.[5]

Using the calculated probabilities, p=0.5% gives about 375 inspections and 3.12 person-hours per lot on average. At p=1%, these become about 1,416 and 11.80 hours; at p=2%, about 4,643 and 38.69 hours; and at p=5%, about 9,529 and 79.41 hours. Direct manual inspection of all ten thousand parts requires about 83.33 person-hours. These figures include inspection operations only, excluding movement, isolation, rework and review administration.

This explains why budgeting for only a little over one hundred inspections per lot can be misleading. As incoming quality worsens, more lots enter full screening. Conversely, an average of 3.12 hours does not imply that every lot finishes within that time. Once a particular lot is rejected, it needs nearly a full lot's inspection effort. A delivery contract allowing only two hours of waiting may be infeasible despite low long-run average cost.

A decision sheet should include normal workload, maximum rejected-lot workload, parallel station capacity and isolation storage. If two inspectors each provide eight effective hours, they plainly cannot complete approximately 83.33 person-hours of screening in one day. This is capacity arithmetic, not an assumption that people can remain continuously productive. Responses include reserving review shifts, arranging supplier disposition or narrowing acceptance scope, rather than quietly increasing the acceptance number when a queue forms.

Compare three feasible paths beyond equipment price

The first path retains random sampling and the existing rejection process. It fits stable quality definitions, traceable lots, acceptable residual risk and genuine rejection-handling capacity. Investment focuses on sample retrieval, consistent interpretation and isolation. Initial equipment demands can be modest, but uninspected products retain risk and rejected lots create workload variation. If the present system satisfies the constraints, retaining it is a real baseline; automation should not be assumed necessary for every project.

The second path screens each part with AI and confirms flagged cases. Assume the same ten-thousand-part lot has p=0.5%, model sensitivity is 98%, and the false-positive rate among conforming parts is 0.5%. Expected counts are 49 detected nonconforming parts, one missed part, and 49.75 false alarms: about 98.75 reviews. At two minutes each, that is approximately 3.29 person-hours. All model parameters are hypothetical; this calculation does not establish that AI is cheaper than sampling.

The 3.29 hours excludes loading, image acquisition, computing equipment, identity matching, rejection hardware, maintenance and checks of unflagged output. Moreover, it describes a screened collection of parts, not the earlier lot acceptance probability. Compare against one delivery objective first: allowed residual defects, conditions requiring suspension and whether each part needs evidence. Without a common objective, an ROI comparison of the two hour totals treats different deliverables as equivalent.

The third path retains independent random acceptance while using AI for line-wide alerts or interpretation support. It can accumulate evidence about operating conditions, but pays for two workflows. High-scoring samples can support diagnosis and improvement; they must not quietly replace random acceptance samples. Otherwise the acceptance evidence becomes biased toward suspicious parts and the original curve no longer describes the plan. Replacing human interpretation with model judgments also requires re-evaluating inspection error and its effect on lot risk.

A procurement comparison should use a common period and include equipment depreciation or rental, integration, review labor, maintenance downtime, rejection handling and delivery losses. Keep quality risk visible as a constraint. If credible monetary defect-loss estimates are unavailable, do not invent a precise return. First compare feasible options meeting the same risk conditions, then document unpriced risks, assumptions and responsible owners.

Two failures that can overturn the decision

The first is a distorted sampling frame or retrieval process. Suppose defects concentrate in the final box after a material change, but sampling covers only the first half of production. A calculation for 125 randomly selected parts from the entire lot is inapplicable. Evidence should connect selected and physically retrieved positions, production-time coverage and box identifiers. Rebuild the frame, consider splitting lots or planned stratification, and repeat physical traceability checks. Adding a test to the random-number program alone does not fix this failure.

The second is inspection error. If true defect proportion p, sensitivity s and false-positive probability f are fixed, an independent-classification model gives an observed nonconforming probability q=p×s+(1-p)×f. Substituting q into the sampling acceptance formula yields acceptance under imperfect judgments, rather than ideal inspection risk. This is our derivation from the law of total probability, not an institution's assurance about field performance.

For example, p=5%, s=80% and f=0 imply q=4%. A 125-part sample permitting at most two flagged parts then accepts with probability approximately 11.96%, versus 4.77% under perfect interpretation. It also exceeds the example's chosen 10% ceiling. Faster inspection can therefore allow more bad lots through. If reflections cause correlated misses across parts, independence fails too, and even this adjusted calculation is insufficient.

Reinspect parts with dependable adjudication evidence across shifts, lighting and inspectors. Separate misses, false alarms and disputes, preserving original evidence. Remediation may involve additional views, calibrated criteria, station changes or another inspection method. Afterwards, re-estimate the relevant errors and recalculate the lot curve. Increasing the sample size of an exterior visual inspection cannot automatically solve an unobservable internal-crack problem.

A review package for proceeding, narrowing scope or deferring

A reusable review sheet should contain seven concrete items: traceable lot definition; counting by unit or defect; actual retrieval of random samples; acceptable risk endpoints and intermediate behavior; evidence supporting inspection error; separate procedures for rejected lots and missing results; and owners of normal and worst-case workload. Each item needs an evidence owner, version and unresolved issues. Mark unsupported items as unconfirmed.

Proceed when recalculated sampling risk meets the business's agreed limits within a defined scope, inspection errors do not breach those limits, retrieval is traceable, and a rejection exercise fits capacity and deadlines. Narrow scope when some part numbers or conditions lack evidence but others are supported. Defer when there is no usable frame, critical defects are unobservable, rejected lots cannot be isolated and handled, or all evidence comes from selected demonstrations.

After deployment, monitor lot acceptance, rejection handling time, reinspection disagreements and sampling deviations separately. An improving acceptance rate might reflect process improvement, biased retrieval or failing interpretation. It is insufficient by itself for rewarding a system. Changes to models, cameras, materials or lot rules should trigger reassessment of affected evidence rather than automatic reuse of an old curve and commitment.

This durable decision framework uses statistical and product documentation readable as of November 10, 2026; it does not assess any factory's quality system. The example assumes binomial approximation, fixed inspection errors and agreed disposition. Arithmetic checks are not production experiments. Industrial AI value should be demonstrated through risk and complete cost against the same delivery objective, with final release tied to explicit lot records and accountable owners.