If you are introducing AI into an incoming inspection sampling plan, the first decision is not how many parts an algorithm wants to inspect. It is which approved plan applies to a specific supplier, part revision, characteristic and incoming lot. In a Thai factory, purchasing, warehouse and QA records often live in different systems. AI can assemble evidence and flag exceptions, while ISO 2859-1:2026 or the contractually agreed scheme remains the authority for sampling and lot disposition. This guide explains how to buy, pilot and accept such a system.
Define the incoming lot before choosing a sampling plan
Incoming inspection determines whether a received lot can be used, must remain on hold, or needs a formal nonconformance decision. Goods receipt is different from QA release. A warehouse may correctly count containers while a critical dimension, material grade or certificate remains unverified. Establish the physical and digital boundary between receipt, inspection, hold and release before adding a model. Our receiving and incoming inspection system guide discusses that transaction boundary in more detail.
The lot ID must follow the contract, drawing and supplier manufacturing history. Items arriving on the same day may come from different production lines, heats or material batches; splitting one homogeneous population into many artificial lots can also distort inspection history. Record supplier and manufacturing site, part and revision, source lot, box or pallet IDs, quantity and arrival ID. An AI reader may propose a value from a label, but QA resolves ambiguous identity. If a lot is split or combined, retain parent-child links and define which inspection results remain valid. Never copy a pass result onto both new lots without checking the sampling population.
A usable plan includes the inspection characteristics, standard and edition, inspection level, AQL index where applicable, lot size, plan type, random selection procedure, sample size, acceptance and rejection criteria, switching state, exception route and approver. If any required input is missing, a plausible sample count generated by AI is still unusable. Hold the lot and resolve the missing input instead of copying the last receipt’s plan.
ISO 2859-1:2026, AQL and the approved baseline
ISO describes ISO 2859-1:2026 as a third-edition system of AQL-indexed sampling schemes for lot-by-lot inspection by attributes. Its public summary identifies single, double and multiple sampling, updated application guidance and skip-lot procedures. The public summary is not the complete standard or a lookup table. The factory must use the licensed applicable edition and its approved procedure, together with the customer contract, to calculate the actual sample size and acceptance/rejection numbers. A customer-mandated older edition or a separate company specification does not automatically change when ISO publishes a revision.
ASQ describes ANSI/ASQ Z1.4 as an attribute-sampling scheme with normal, tightened and reduced plans and switching rules for a continuing stream of lots. Z1.9 addresses variables sampling. These are separate documents; do not assume table values or conditions are interchangeable with ISO 2859-1:2026. Store the standard, edition, contract and plan revision used for each decision.
AQL is an index within a scheme, not a promise that every accepted lot contains less than a stated percentage of defective units. NIST’s operating-characteristic discussion explains how the probability of accepting a lot depends on its actual defect fraction and the chosen sample size and acceptance number. Good lots can be rejected and poor lots can be accepted with nonzero probability. Producer and customer risk must therefore be agreed for the characteristic and use case. A screen saying “AQL 1.0 means up to 1% defects are fine” would mislead an inspector.
Imagine a 1,000-piece arrival. Quantity and an AQL label alone do not establish its sample size: the applicable inspection level, plan type, switching state, characteristic and edition are still needed. This example deliberately gives no sample or Ac/Re number. The production system should retrieve those values from the approved plan and show the rule version alongside the observed result. Double and multiple plans need their own stage logic. Critical, safety or regulated characteristics may require contract-specific checks or 100% inspection even when a general attribute scheme applies elsewhere. A certificate alone does not prove the physical lot, heat number and drawing revision match.

Use supplier history for risk triage, not automatic lot release
AI can rank evidence that a QA reviewer should examine: formal incoming nonconformances, supplier corrective actions, manufacturing site changes, late change notices, missing certificates and unusual inspection records. Separate supplier, site, part, revision, characteristic and time period. One supplier’s history across unrelated plants or products is a poor proxy for the arriving part. A new part has little history, and a familiar part may become unfamiliar after a process transfer or drawing change. The display should say “insufficient data” or “changed conditions” rather than treating missing data as zero risk.
Use denominators. Ten defects in ten inspected lots and ten defects in ten thousand inspected lots mean different things. If inspection is reduced, detected defects may fall simply because fewer units are examined. Keep receipt volume, inspected volume, inspection policy and discovery point beside the defect count. Distinguish a submitted 8D report from verified corrective-action effectiveness. Link every AI summary to the source record, approval status and affected lot.
Candidate features include the same part and characteristic’s confirmed nonconformances, current switching state, approved supplier changes, certificate mismatches and missing inspection results. Avoid accidental proxy variables such as driver name, inspector identity or weekday as a supposed measure of supplier product quality. If a model displays a probability, document its training population, calibration and false alarm and miss rates for the population where it will be used. A traceable statement such as “corrective-action effectiveness is still open for this part” is often more useful than an unexplained score.
Keep the rules engine and AI assistance separate
The rules engine selects the approved plan from contract, part, characteristic, lot size and switching history. AI retrieves and summarizes supplier evidence and suggests a reason for QA review. A high risk signal can propose an additional check or hold for an authorized person to approve. A low risk signal cannot quietly reduce mandatory sampling. Neither signal changes AQL, sample size, acceptance criteria or switching state by itself. An AI model update should not modify the approved sampling procedure; a procedure update should not depend on model retraining.
| Step | AI may assist with | Final authority |
|---|---|---|
| Read labels and certificates | Extract candidate IDs and show mismatches | Receiving and QA personnel |
| Find a plan | Prepare search inputs | Approved rules engine and plan master |
| Summarize supplier history | Link evidence and flag review items | QA checks source facts |
| Propose more inspection or hold | Explain reason and affected lots | Authorized QA approver |
| Change AQL, Ac/Re or switching status | No silent change | Contract and formal change approval |
| Release, reject or dispose of a lot | Organize evidence | Designated quality authority |
Missing inputs must produce a hold or review state. Examples include a label revision that differs from the drawing, an unregistered manufacturing site, cartons that do not match receipt quantity or an ambiguous contract reference. AI may suggest a candidate value, but record the source and human confirmation. The formal plan and optional AI warning should appear in separate areas of the screen. A temporary extra-inspection decision needs its own approval trail and affected-lot scope. This fits into the broader quality management system for overseas plants.

Preserve AQL switching rules and skip-lot eligibility
Normal, tightened and reduced inspection states are not toggles driven by each lot’s AI score. An approved scheme tracks a continuing sequence of lots and specified conditions. ISO’s public summary says its 2026 edition includes switching rules and skip-lot procedures, but the exact eligibility and transitions must be taken from the applicable standard and contract. ASQ separately describes switching within Z1.4. A private rule such as “halve the samples after one good lot” should not replace either scheme.
Define the history key: contract, supplier manufacturing site, part or approved homogeneous family, characteristic and standard edition. Mixing unrelated product histories may let a good product hide a bad one. Record lot order, inspection date, result, resubmitted lot, skipped lot, transition rule and approver. Decide how missing sequence numbers, returns and late-discovered defects affect the history. If AI detects contradictory records, pause the transition for QA rather than filling gaps to make a dashboard look complete.
Skip-lot sampling requires eligibility, selection, ongoing surveillance and revocation. The mere existence of a provision in the new ISO edition does not authorize its use with every customer. Specify exclusions for critical characteristics and certificates. When a plan revision takes effect, record the effective time and first eligible lot. Keep the plan snapshot on an issued inspection order; do not retrospectively recalculate an inspected lot just because a master table changed. If a wrong edition was applied, search the affected lots and let QA decide containment, reinspection and customer communication.
Make physical sample selection and findings traceable
Even a correct table fails if inspectors always take items from the easiest top carton. Define how samples are distributed across pallets, boxes and production times. If a mobile terminal selects positions, store the sampling population, randomization record, inaccessible-box treatment and reason for replacing a selected unit. Confirm the instructed carton can actually be reached on the shop floor.
Link each measurement to units, gauge, calibration, drawing revision and relevant conditions. For attribute inspection, distinguish number of nonconforming units from number of nonconformities; a unit with multiple defects can be counted differently under different approved plans. Variables data cannot simply be inserted into an attribute acceptance table. AI may classify images or readings, but boundary samples require the approved visual limit and a human decision where the procedure says so.
Fail, hold, reinspection, sorting, supplier return and concession are separate states. A failed acceptance decision must not automatically release usable ERP stock. Sorting or rework after the result is a later disposition workflow; repeated convenient resampling to obtain a pass undermines the plan. State how a resubmitted lot is identified and handled under the governing procedure.

Architecture and access for a Thai factory
A practical configuration brings purchase and expected receipt data from ERP, verifies the physical lot in WMS, and manages plan selection and results in QMS or an inspection application. AI reads the allowed certificates, change notices and historical records, then puts evidence links on the inspection screen. A purchase order alone is not enough as a key. Connect supplier ID, manufacturing site, part and revision, supplier lot, receipt ID, characteristic ID and plan version. Preserve the supplier’s original lot text even if you add a normalized value.
Define states such as received, awaiting inspection, inspecting, on hold, accepted for use and rejected pending disposition. Decide exactly which state ERP/WMS may allocate to production. When a release message fails, keep the system and physical lot on hold. Separate inspector entry from QA authorization. Night-shift inspection may be complete while release waits for an approver. Role permissions should separate plan-master edits, inspection, result correction, temporary extra-inspection approval, switching approval, lot disposition and AI model changes. Delegate authority with a named period and reason.
The audit trail should reconstruct receipt, plan selection, physical sample selection, measurement, AI alert, human review, disposition and inventory release. Retain actor, time, source record, plan edition, model version, evidence shown and approval reason. A later model must not overwrite what an operator actually saw. Thai, Japanese and English interfaces should map the same part, unit, status and critical characteristic to stable IDs; translations must not change acceptance meaning. For documents sent to an AI provider, specify allowed data, retention, access, processing location and whether provider training is permitted.
PoC, RFP, FAT and SAT acceptance criteria
Start a pilot with one part family and a few suppliers. First digitize the existing approved plan and prove that lot IDs and inspection findings connect. Then use AI only for cross-checking certificates and summarizing histories. Test whether it finds exceptions that humans missed and whether false alerts obstruct receiving. Because the rules engine can run without AI, model outages can be tested without stopping formal inspection.
NIST’s AI Risk Management Framework structures governance around mapping, measuring and managing AI risks. Its Measure playbook calls for metrics that match important deployment risks and documentation of what cannot be measured. For this application, evaluate wrong lot matching, incorrect rule edition, critical-characteristic omission, certificate mix-ups, translation errors and drift after model changes, not only model score. Agree on thresholds against the actual parts and customer requirements before the pilot.
An RFP should identify the standard edition and licensing, plan-master owner, precedence when several plans apply, switching history, approval rights, offline behavior, ERP/WMS interfaces, audit export and AI data boundary. Ask vendors to reproduce the selected plan and evidence from the same lot input. Clarify who updates the tables and who approves a contract change. A proposal that says only “AI automatically optimizes sample count” leaves the contractual obligation unresolved.
| Test | Action | Expected evidence |
|---|---|---|
| Ordinary receipt | Use a registered part and complete lot data | Approved plan edition and sample-selection record appear |
| Revision mismatch | Change label versus drawing | Hold for QA; no guessed release |
| Critical characteristic | Set AI score to low risk | Mandatory inspection remains |
| Supplier issue | Add confirmed nonconformance and open action | AI links source; QA can approve extra check |
| Switching boundary | Supply an approved qualifying history | Correct state and rule edition; no silent change |
| Interface failure | Interrupt ERP/WMS release message | Physical lot and inventory both remain on hold |
| Model outage | Stop or roll back AI | Formal plan and human decisions continue |
FAT tests plan-table import, lot boundaries, permissions, state transitions and audit export in the vendor environment. SAT repeats the cases with real Thai-site labels, terminals, gauges, network, hold locations and shift roles. In both, demonstrate that an operator can reject a bad AI suggestion. Faster inspection is not a successful acceptance test if incorrect releases increase.
After launch, monitor plan misapplication, lot-ID mismatch, waiting time, approval latency, AI suggestions accepted or rejected, and defects found downstream. Interpret defect counts with inspection volume and detection stage; fewer findings after sampling is reduced do not prove improved quality. Look for bias by supplier site and part revision, and route new parts, process changes and returns after long inactivity to human review. Approve each improvement through change control with a rollback condition.
Common implementation failures and cost boundaries
Part numbers in purchasing, customer drawings and supplier labels may differ. Maintain an approved cross-reference with effective dates; an AI match based on similar text is only a proposal. Measure unreadable labels and manual-entry corrections, because an unreadable carton may belong to a different lot. Decide whether an offline terminal forces a hold or permits a controlled paper record. On reconnection, link the original record, actual inspection time and later entry time without duplicating a result.
If a gauge is later found overdue for calibration, search every lot inspected with that gauge in the affected period. Retain its ID and calibration state at inspection time. The budget must include plan-master cleanup, rights to digitize and maintain standard tables, ERP/WMS interfaces, labels, terminals, gauges, network coverage, training and audit retention. Separate implementation from recurring costs. Measure time spent finding a plan, exception detection, prevented premature releases and ability to explain a lot later; do not promise a blanket reduction in mandatory sample counts.
Deploy in three stages: audit existing plans and master data without AI; test physical and digital lot states with real shifts and receipts; then add AI evidence checks using past cases. This helps identify whether a failure came from master data, shop-floor execution or the model. Test out-of-order events: a supplier change notice may arrive during inspection, or a complaint about a previous lot may arrive after release. Preserve the original inspection order, create a new review event and let QA decide scope. Keep test data separate from production switching history, and reconcile the last legacy lot with the first new-system lot. If a terminal fails mid-inspection, restore which samples were taken and which results were saved without assigning them to another lot. Supplier feedback should use confirmed nonconformance evidence and the applicable drawing and plan, not an unexplained AI score.
Frequently asked questions
Can AI set an incoming inspection sampling plan by itself?
It can organize supplier evidence and propose an exception. The formal plan comes from the contract, applicable edition and approved rules. A change needs QA authorization and an audit trail.
Does the same AQL mean the same sample size for every part?
No. Lot size, inspection level, plan type, switching state, characteristic and contract all matter. Use the governing table, not the AQL number alone.
Can a low supplier risk score justify skipping inspection?
No. Reduced or skip-lot inspection is available only when the governing scheme and customer permit it and its eligibility, surveillance and revocation rules are met. Critical characteristics may retain separate checks.
Does 100% inspection eliminate the need for sampling governance?
Full inspection still needs controlled gauges, detection capability and records. Some destructive or slow tests cannot be applied to every unit. Specify the method by characteristic.
Should we automatically migrate to ISO 2859-1:2026?
Publication does not amend a customer contract. Confirm the edition, effective lot, table updates, training and acceptance tests with QA and the customer.
Conclusion
Start with clear lot and characteristic definitions, the approved plan, switching history, physical selection and release authority. AI then helps QA see supplier changes and missing evidence with source links. Authorized people approve plan changes and lot disposition, while the system preserves the applicable rule, model version and reason. That is a procurement scope that can be tested on real receipts.
If your Thai factory is mapping its current incoming-inspection workflow and deciding where AI assistance belongs, contact TOMAS TECH. A pilot on one part family and its real receipts can establish a practical system boundary before solution selection.