AI Visual Inspection Implementation 2026: PoC, RFP and Acceptance Testing for Thailand Factories

A successful AI visual inspection implementation in a Thailand factory is not defined by an impressive model demonstration. It begins by agreeing which defect must be judged, on which product and line condition, within what response time, and who will decide an ambiguous case. This guide gives plant managers, quality teams, production engineers and IT/OT owners a complete procurement path—from PoC design and RFP requirements to FAT, SAT, TCO and production monitoring. It deliberately avoids invented accuracy, ROI and price claims. The objective is to give each factory a measurable basis for its own go/no-go decision.

Start the AI visual inspection implementation with the quality decision

An inspection model produces a score or a candidate defect. The factory, however, needs a disposition: pass the part, reject it, isolate it or send it for review. Therefore, write the project around the quality decision rather than around “AI.”

Break the present inspection into product, material, colour, finish and size; defect type, location and severity; line speed and available imaging time; the OK/NG/REVIEW route; and the records that connect image, lot, recipe and final disposition. A vague request to “find scratches” can produce a good-looking demo but no defensible acceptance test. A precise defect and action definition allows the team to compare rules-based vision, normal-only anomaly detection, supervised learning and human inspection—and to retain a non-AI solution where it is easier to maintain.

NIST’s smart-manufacturing AI/ML roadmap, published on 3 July 2026, identifies industrial-data complexity, data management, heterogeneous sensor/control integration, and trustworthy, explainable and reliable operation as deployment challenges. Evaluating only an image classifier therefore does not evaluate the manufacturing system.

Screen the right task for automated visual inspection

Do not automate every inspection at once. A strong first candidate has understood quality impact, controllable imaging conditions and traceable outcomes.

An internal defect that does not appear in an image cannot be found by a camera. A fine scratch on a reflective surface may be primarily an illumination problem. Ask whether the whole inspection cell can observe the defect reproducibly, not merely whether AI can classify a selected photograph.

Camera, lighting, lens and fixture form the foundation

AI does not magically repair unstable input. If the same good part looks different across shifts, oil conditions or fixture maintenance, the input distribution has changed.

Camera and lens

Match field of view, working distance, depth of field, part speed, exposure and transfer time to the smallest relevant feature. More pixels may reveal detail but also increase processing time, storage, lighting and lens demands. Select the optical chain from the repeatable visibility of the defect, not from the largest camera specification.

Lighting and enclosure

Compare ring, bar, dome, coaxial, backlight and dark-field arrangements according to the surface and defect. Where ambient light changes, add an enclosure as well as controlled illumination. Record the recipe and acceptable range. LED ageing, contamination, cable replacement and unauthorised setting changes belong in the maintenance plan.

Fixture and trigger

Control pose and vibration, and reproduce capture timing through sensors and the PLC. A model’s tolerance to movement is not a reason to accept unlimited variation. Removing physical variation first reduces data, validation and maintenance complexity.

ISO/TR 14997-2:2022 is a technical report for surface imperfections on optical elements, not a universal standard for every manufactured product. Its attention to fidelity, repeatability and reproducibility is nevertheless a useful analogy: repeat-capture the same part, remove and re-mount it, and repeat after maintenance or on another shift. These tests are an engineering interpretation, not a claim that the report mandates them for all AI inspection.

Choose between normal-only learning and supervised learning

No single method fits every visual-inspection task.

Normal-only learning does not eliminate the need for defective samples. Defects are still needed for the PoC and acceptance test, because “the model reacts” is not the same as “the response supports the quality decision.”

MVTec AD 2 contains eight industrial anomaly-detection scenarios and more than 8,000 high-resolution images. It provides defect-free training/validation images and test images containing normal and anomalous cases under lighting conditions not necessarily present in training. It separates a labelled public test part from a private-ground-truth part. This does not prove factory performance, and its licence does not permit commercial use. It does illustrate two sound test ideas: include lighting variation and keep final evaluation data separate from tuning data.

Design the PoC to decide go, revise or stop

A PoC should reveal production-transfer conditions and residual risks, not maximise a headline score.

1. State scope and exclusions

List products, defects, lines, shifts and equipment states. Explicitly list what is not inspected. If the underside is not imaged, underside defects remain out of scope.

2. Split and seal the data

Separate training, threshold-tuning and evaluation sets. Do not use the sealed evaluation set to adjust the model. Record product, lot, date, shift, machine settings, lighting and defect class, and prevent near-duplicate images of one item from crossing splits.

3. Define false rejects and misses separately

A false reject sends a good unit to NG or REVIEW and consumes yield or reinspection capacity. A miss passes a defective unit and creates escape risk. A single “accuracy” value hides this trade-off. Report counts and denominators by defect class and severity. Where rare critical defects lack enough evidence, mark them “not yet validated” and retain additional testing or human inspection.

4. Create a third REVIEW route

Route uncertain scores, bad images, unknown products and system faults to REVIEW. Measure review volume, delay, adjudicator and final decision. Human review reduces pressure to force ambiguous cases into OK or NG.

5. Reproduce line conditions

Include speed, vibration, ambient light, dust, oil, changeover, cleaning, maintenance and network delay. Anything not reproducible in the PoC becomes a named SAT item, not an invisible assumption.

Build a defensible PoC scorecard

No vendor or article can set universal pass values. Quality and manufacturing must derive them from customer requirements, defect consequences, current inspection and review capacity.

A percentage without its denominator is insufficient. An aggregate can hide one failing product. Deliver raw images, processed images, model version, threshold, decision reason and PLC response so results can be re-analysed.

Put complete requirements in the RFP

Structure the RFP around the cell, data, integration, operation and acceptance—not “one AI camera.”

Business and quality

• Defect definitions, severity and OK/NG/REVIEW disposition
• Products, future variants, recipe switching and permissions
• Acceptance method for misses, false calls, capture failures and response time
• Audit, image retention, search and deletion needs

Optical, mechanical and controls

• Rationale for camera, lens, lighting, enclosure, fixture and sensors
• Field of view, working distance, permitted pose error, speed and vibration
• PLC signals, reject mechanism, fail-safe behaviour and safety boundary
• Cleaning, verification, replacement recovery and local spare availability

AI and data

• Method, data provenance, labelling procedure and independent evaluation set
• Versioning and approval for model, threshold, recipe and software
• Handling of uncertainty and unknown input; REVIEW and manual override
• Data ownership, retraining rights, transfer, encryption and access logs
• Drift triggers, retraining approval and rollback

Systems and service

• Edge hardware, network, PLC and MES/QMS interfaces
• Throughput, latency, error handling, clock synchronisation and offline mode
• Backup, recovery, cybersecurity and remote support
• Thai/English training, manuals, service hours and local response
• Responsibility for FAT, SAT, stability run, defects and change control

NIST’s AI for Manufacturing project examines fitness-for-purpose, interoperability, human-AI teaming, interpretability and traceability. Its integration metrics include throughput, latency, error rates, semantic correctness and scalability. An RFP should similarly test the system around the model.

Separate FAT, SAT and final acceptance

FAT verifies the supplied configuration before shipment or at the supplier’s site. SAT verifies the installed cell under actual plant conditions.

Acceptance records should show untested items, conditional acceptance, correction dates, retest method and the person accepting residual risk. If a threshold is changed to pass, re-evaluate both misses and false calls. Define which tests repeat after a model, dataset, lighting or recipe change.

Estimate cost and TCO without invented prices

Initial costs

Generic prices are misleading because optics, line modification, integration, data and service differ. Break the quotation into site survey and PoC; camera, lens, lighting, enclosure, fixture and controls; edge computer, storage and network; software, PLC and MES/QMS integration; data collection, labelling, development and validation; and installation, FAT, SAT, training and documentation.

Recurring costs

Recurring costs include support subscriptions, hardware replacement, storage and backup, monitoring, adjudication, re-labelling, retraining, revalidation, product additions and remaining human review.

Use one period and explicit assumptions:

TCO = initial cost + recurring cost over the period + internal labour + expected cost of downtime/change risk − residual value

Net benefit = change in reinspection/sorting + change in escape/scrap/rework losses + change in downtime + change in audit/trace labour − new operating burden

Every ROI input should have an assumption, evidence source, boundary and sensitivity. Do not infer escape savings from a short trial when critical defects are rare. Use the PoC for technical feasibility, then update cost and benefit after a controlled stability period.

Thailand BOI publishes information concerning automation and technology investment, but this article does not determine eligibility. Promoted activities, eligible expenditure, application timing and effective periods can change. Confirm the current position directly with BOI and qualified advisers before commitment, and do not build the business case on an unconfirmed incentive.

Monitor data drift and equipment degradation in production

A model accepted at launch can change in effect as products, materials, processes, cameras, lighting and work practices change. Monitor this broadly as drift, while separating model issues from dirty lights, a loose lens, fixture wear, upstream process shifts or new products.

Display OK, NG, REVIEW, capture and communication failures; adjudicated misses and false calls; score distributions and unknown inputs; response time and queue; available camera/lighting health; and model, threshold and recipe history. Break results down by product, defect, shift, station and version.

Define action after an alert: stop or continue production, switch to human inspection, quarantine from the last known-good verification, call maintenance, or roll back. Automatic retraining is not the first response; it may learn an equipment fault. Diagnose the cause, approve the data, retrain under change control and revalidate against a sealed set.

The NIST AI RMF Core organises work into four iterative functions—Govern, Map, Measure and Manage—and calls for monitoring, override, incident response, recovery and change management. In factory terms: assign owners and rules, define the use and risk, measure under deployment-like conditions, then manage the live system. It is guidance, not a Thailand-specific legal requirement.

Keep accountable human judgment

Automating visual inspection changes what people inspect; it does not erase responsibility. The system can perform repeatable screening while people handle exceptions, critical defects, root-cause analysis and customer decisions.

A REVIEW screen should show the original image, indicated location, reason, product and lot. Record who accepted, rejected or held the unit and why. Treat overrides as evidence for improving thresholds and definitions, not as operator failure. For high-consequence shipment decisions, decide explicitly whether AI-only release is allowed. Maintain a safe fallback to manual inspection and train staff on scope limits, signs of bad imaging, override criteria and incident reporting.

Assign project responsibilities

Name one accountable approver. “Quality and IT will discuss” is not ownership. Apply request, impact assessment, testing, approval, release and rollback to model, threshold, lighting and product changes.

Common failure patterns

Running the PoC only on favourable samples

Register production variation, seal condition-balanced evaluation data, and manage the diversity of good units separately from the shortage of important defect samples.

Accepting the system on aggregate accuracy alone

Separate class-level misses from false calls, retain denominators and sample provenance, and state which conditions remain unvalidated.

Mixing model faults with equipment faults

Retain original images and equipment state, and code capture, lighting, alignment, inference and communication failures separately.

Counting only the camera and AI licence

Include fixtures, controls, data, labelling, integration, validation, training, storage, monitoring and product additions in TCO.

Launching without a production owner

Assign ownership for KPIs, alerts, retraining approval, supplier escalation and manual fallback before production starts.

FAQ: AI image inspection procurement

What is AI visual inspection?

It detects or classifies candidates such as scratches, contamination, chips or assembly errors from images. A production solution also includes optics, fixtures, controls, rejection, storage and human review.

Can image inspection AI completely replace manual inspection?

It depends on scope and conditions. Stable, visible and well-defined repetitive judgments are candidates. Unknown anomalies, subjective criteria and critical shipment decisions often still need REVIEW and a manual fallback.

What accuracy target should defect-detection AI have?

There is no universal value. The factory sets class- and severity-specific miss and false-call limits from customer requirements, escape consequences and review capacity. Always state denominator, conditions and unvalidated scope.

How should AI visual inspection costs be compared?

Compare PoC, optics, mechanics, controls, integration, data, software, FAT/SAT and training, then include support, storage, monitoring, retraining, new products and REVIEW labour in TCO.

Does normal-only learning remove the need for defective samples?

No. Defect samples remain necessary for PoC and acceptance. A rare critical defect without enough evidence should remain “not validated,” with further tests or human inspection.

How is data drift handled?

Monitor by product, shift, station and model version, diagnose equipment/process/data/model causes, retrain with approved data and revalidate against a sealed set.

Related guidance

For combining image results with equipment and process signals, see AI anomaly detection for factories. For turning inspection records into quality improvement, see AI analysis of quality inspection data.

Conclusion: work backwards from acceptance

An AI visual inspection implementation in Thailand should define defects, disposition, imaging, misses, false calls, REVIEW, data, integration, acceptance and monitoring before choosing a model. The PoC supports an investment decision; the RFP contracts for evidence and responsibility; FAT and SAT establish fitness under supplied and actual conditions. Compare TCO, monitor drift and equipment ageing, and preserve accountable override and fallback. That is a more durable route to automation than delegating final responsibility to a score.

If you would like support defining the target station, PoC scorecard, RFP or FAT/SAT plan, contact TOMAS TECH. You can start while the camera and AI method are still undecided; the discussion can begin with the actual part, equipment and quality workflow at your Thailand facility.

References

• NIST, 2026 Roadmap on Artificial Intelligence and Machine Learning for Smart Manufacturing
• MVTec, MVTec AD 2
• ISO, ISO/TR 14997-2:2022
• NIST AIRC, AI RMF Core
• NIST, Artificial Intelligence (AI) for Manufacturing
• Thailand BOI, Automation information