AI Visual Inspection in ASEAN Factories: Cost, ROI and Pitfalls

Keeping people stationed on a manual inspection bench is getting harder every year at plants in Thailand and Vietnam. Wages climb, trained inspectors do not stay, and the pass-fail standard often lives in one senior operator’s head at the parent plant rather than in any document, which is why AI visual inspection keeps coming up. It is also when teams discover that most published guidance assumes a domestic factory in a high-wage country, with the group engineering team a short drive away. This article works through the decisions a multinational manufacturer with ASEAN sites actually faces: how to break the cost apart, how to read accuracy claims, who retrains the model after go-live, and where Thailand’s BOI incentives fit into the schedule.

Why the cost of manual inspection behaves differently at ASEAN sites

“Labour is cheap there, so manual inspection is fine for now” is still a common verdict from a parent company or regional HQ. Anyone standing next to the line sees that this assumption is eroding from several directions at once, and eroding in a shape that does not match a domestic plant in a high-wage country.

Three reasons the cheap-labour argument breaks down

The first reason is wages. Thailand’s statutory minimum wage is set by province, and as of July 2026 it ranges from 337 to 400 baht per day (Thailand-specific). That “per day” matters more than anything else in this paragraph. Read as an hourly figure, it throws a labour-cost-per-hour model off by an order of magnitude, and payback periods calculated on it become meaningless. This is one of the most common arithmetic accidents in overseas capex proposals. Note also that the rate has not been unified nationally as of July 2026 (Thailand-specific), so two plants belonging to the same group can sit on different baselines depending on which province their industrial estate is in. When a regional HQ builds one consolidated model for several sites, that spread quietly disappears and distorts the result.

Beyond that, the people actually doing visual inspection are rarely paid the statutory minimum. Any process that requires trained judgement is recruited above the floor rate, and on top of the base you carry overtime premiums, social security, shuttle buses, meal subsidies and dormitory support. In Vietnam, manufacturing wages have been reported as continuing to rise at 7 to 10 percent per year (Vietnam-specific, Intralink). A cost line growing at 7 percent roughly doubles in ten years. The payback model you build on today’s labour rate is a different document five years from now.

The second reason is that you cannot recruit and cannot retain. Every time a new plant opens in the same industrial estate, inspectors move. Visual inspection is seated, repetitive work, so it tends to be the first process to lose people when a better offer appears nearby. Replacements need a training period before they produce the same verdicts, and judgement is unstable in the meantime. In other words, your inspection quality is coupled to the local recruitment market. From a quality assurance point of view that is an uncomfortable dependency to carry.

The third reason is that the skill sits in individuals. At ASEAN sites, the final borderline calls are often made by a quality assurance expatriate on assignment, or by one or two long-serving local leaders. When that person rotates home or resigns, the standard shifts, and the customer says quality has changed. That pattern is a signal that a standardisation task is still outstanding — a task that exists independently of any automation decision.

The real cost of manual inspection, itemised

If the only saving you put on the table is the payroll of the inspectors, most business cases fail. Take inventory of the following instead, and the picture changes.

• Direct labour: number of inspectors x working hours x fully loaded hourly cost
• Training: hours until a new hire works unsupervised, calibration sessions where inspectors align their judgement, recurring every year in proportion to turnover
• Re-inspection and 100% sorting: emergency sorting when a lot becomes suspect, weekend shifts, borrowed headcount
• Customer complaint handling: root cause investigation, report writing, review meetings, travel, price concessions and compensation
• Shipment holds and expedited freight: switching to air freight to protect a delivery date
• Over-rejection: good product scrapped by inspectors who judge too strictly

Expedited freight is the line item that behaves very differently overseas. A delay a domestic plant absorbs with a next-day truck becomes an air freight switch when the shipment leaves Thailand or Vietnam for Japan, Europe or North America. The delta from a single switch varies widely by weight and route, but it often lands in the same order of magnitude as a month of inspection payroll. Pull the actual figures from your own air freight records; this is a line item to verify internally rather than from published statistics. Whether you are allowed to count that avoided cost as a benefit changes how easily the proposal clears the parent company’s approval process.

Start from loss, not from cost

The recommended order of calculation is to start with the loss you are already absorbing, not with the price of equipment. The logic is simple.

Annual expected escape loss = escape rate x annual shipped units x average cost of handling one escape

The last term can be estimated from your own complaint records over the past two or three years: investigation hours, sorting cost, freight, concessions. Those are your own figures, so you do not need an external statistic to defend them. Once that number exists, the question “how much is it worth to cut the miss rate” becomes a discussion about money rather than a discussion about feelings. Requirements definition for AI visual inspection should start here.

For the wider picture — which processes to automate first, and how inspection ranks against material handling — see our companion piece on factory automation in Thailand. Automating inspection alone, while the transfer before and after it stays manual, often hits a ceiling faster than expected.

What AI visual inspection can and cannot do today

The term AI visual inspection is too broad to support a decision on its own, because realistic difficulty varies enormously by defect type. Request quotations before sorting this out and you will receive numbers built on incompatible assumptions.

Realistic difficulty by defect type

The table below reflects general tendencies. Actual feasibility depends on material, surface finish and lighting, so your own samples always have to be tested. Still, it is a useful way to organise your thinking before an RFQ.

The bottom row is the important one. Where an experienced inspector “just knows”, the obstacle is less a technical limit of machine vision defect detection than the absence of a requirement specification. A standard that no human can express in words and images cannot be taught to a model. Turn that around and the implication is encouraging: if the project forces you to write the standard down, that document becomes a quality assurance asset regardless of what happens to the capex.

Rule-based machine vision versus deep learning

Confusing these two inflates both budget and schedule. The principle is straightforward.

• The appearance is stable and the criterion can be written numerically (area above so many square millimetres, brightness difference above a threshold, part present or absent) — rule-based automated optical inspection is sufficient. It is fast, cheap, explainable and highly repeatable.
• Good parts vary widely, the defect looks different every time, and no numeric rule can be written — this is where deep learning defect detection earns its place.
• Both are present, which describes most real lines — a hybrid, where rule-based logic handles pre-processing, positioning and unambiguous defects, and only the hard judgement is routed to the model.

In practice the hybrid is usually the realistic answer, and the reason is operational rather than technical. A local engineer can adjust a rule-based threshold; a model needs retraining. The narrower the slice you hand to AI, the wider the slice your site can manage on its own.

Whether a vendor can say “this step does not need AI” is a decent proxy for the quality of their proposal. A design that solves everything with a model may reflect confidence, or it may reflect an unwillingness to spend time decomposing requirements. Vendors who scope AI narrowly tend to produce lower total cost over the life of the system.

How to use market data without leaning on it

Across regions, AI in machine vision is already mainstream. Research published by Cognex in March 2026, covering more than 500 manufacturers, integrators and OEMs across North America, Europe and Asia (global, multi-region sample), found that 57 percent already use AI in machine vision and a further 30 percent plan to adopt it in the near future. Adoption by industry in the same study was 72 percent in logistics, 61 percent in electronics and 60 percent in automotive (global, multi-region sample).

On the market side, the global AI visual inspection system market was valued at 29.82 billion USD in 2025 and 36.84 billion USD in 2026, and is forecast to reach 85.24 billion USD by 2030, a compound annual growth rate of 23.3 percent (global, The Business Research Company, July 2026). North America is the largest region and Asia-Pacific the fastest growing.

You can put these figures in a proposal, but handle them carefully. “Competitors are doing it” is a weak basis for capex and is exactly the sentence a group IT or finance reviewer will attack. Keep market data as background. The spine of the decision should remain the expected escape loss and inspection cost you calculated from your own records in the previous section.

AI visual inspection cost: build the quotation yourself

A single lump-sum price cannot be compared across vendors. Split the scope into the six line items below and require every bidder to quote at the same granularity. It helps both commercial negotiation and technical evaluation.

The six line items

1. Imaging (camera, lens, lighting): resolution, field of view, depth of field, illumination method (coaxial, dome, low-angle, backlight). This is derived backwards from the required resolvable feature size and deserves the most technical scrutiny. Cutting resolution to save money cannot be recovered by any model downstream.
2. Handling, fixtures and mechanics: whatever brings the part to the imaging position in the same attitude at the same speed every cycle. This is frequently the largest single item in the total. High-mix low-volume lines raise the difficulty sharply because of changeover.
3. AI software and licensing: initial licence plus annual maintenance. Whether pricing is per line, per camera or per site changes long-run cost dramatically. If you intend to roll out, confirm the price of the second and third unit before you sign the first.
4. System integration: requirements definition, optical design, mechanical design, PLC integration, upper-system integration, installation and witnessed acceptance. At overseas sites, add travel, accommodation and the cost of maintaining a local response capability.
5. Training data creation: capturing images and labelling where the defect is. Fix who does it, against which language version of the standard, and how many hours it takes, before signing. Leave this vague and it becomes a dispute.
6. Maintenance, retraining and spare parts: LED degradation, dirty optics, loose lenses, PC refresh, plus retraining whenever a new variant arrives. Judge on a five-year total or you will judge wrongly.

Cost ranges published by domestic vendors in Japan generally present the scope as AI software plus camera plus lighting plus PC plus integration fee (Japan-specific, NTT Communications, 2025). Note that this composition usually excludes item 2, handling and fixtures, which is unavoidable if the machine has to run in a real line. Copying a domestic reference range from any single country into an overseas budget will leave you short.

Which line items grow, and which shrink, at an overseas site

The message of that table is that the biggest cost variable at an overseas site is not the price of the hardware but who can respond locally. Two identical configurations from two vendors — one with resident engineers in-country, one dispatching from abroad — diverge substantially over five years. Selecting on the initial quotation alone hides that divergence until it is too late to change.

Compare on a five-year total cost of ownership

A simple template every bidder fills in works well.

• Initial cost (sum of items 1 to 5)
• Annual licence and maintenance x 5 years
• Expected number of retraining events x cost per event x 5 years
• Spare parts
• Expected annual downtime hours x loss per hour

On the savings side, place the fully loaded inspector payroll x 5 years, the reduction in expected escape loss, and the reduction in expedited freight. In this format, a proposal with a higher initial price but a lighter operating burden is evaluated fairly. A comparison table built on initial price alone systematically favours proposals that conceal operating load.

One honest caveat: there is very little reliable public statistical data on the cost of automating visual inspection in Thailand or Vietnam (no reliable public data for Thailand or Vietnam). Avoid generalised “market rate” answers; the only way to know is to price the six items with named suppliers. Conversely, if a proposal asserts a market rate without a source, it is worth asking where that number came from.

Three strategies when you do not have enough defect samples

Nearly every project hits this wall, and the awkward part is that it is worst at the sites with the best quality control. A low defect rate means few NG samples to learn from — a paradox widely noted in the field (Nsight). Understanding it up front changes how the project is structured.

Strategy 1: accumulate — build the capture routine before you buy anything

The least glamorous option and the most reliable. The key point is to start capturing before the investment decision is made.

• Place a simple imaging station beside the existing manual inspection bench
• Photograph every NG part before it is scrapped or reworked, without exception
• Record date, lot, machine number, defect type and the inspector’s identity alongside the image
• Capture good parts under identical conditions — understanding the variation among good parts is arguably the most valuable output

Watch the imaging conditions. Images captured under lighting, camera and angle conditions different from the production system lose much of their training value. Ideally you would fix the optical concept before selecting a vendor, which is rarely possible in practice. A workable compromise is to capture under several lighting conditions, and at minimum to record exactly what conditions were used.

A side benefit is that the actual defect profile becomes visible. Six months of capture often reveals that customer complaints trace back to a single defect type. If so, the requirement for AI visual inspection may be far narrower — and cheaper — than originally assumed.

Strategy 2: generate — create defects deliberately, augment data

• Deliberately damage real parts: press a dent with a jig, omit a component, place foreign matter. Effective for reproducible defect types.
• Image augmentation: rotation, flipping, brightness and contrast shifts, added noise. Helps suppress overfitting.
• Synthetic images: composite a defect region onto a good-part image, or generate defect images with a generative model.

Generation has limits. An artificially pressed dent and a dent produced by the actual process can look subtly different, and a model trained only on generated data failing on the real line is a well-known failure mode. Treat generated data as supplementary; collecting a genuine set of real NG samples remains unavoidable. It is also unwise to write a contract that assumes generated data can substitute for real samples.

Strategy 3: design around it — good-part learning and anomaly detection

If NG samples will not accumulate, train on good parts only and flag anything that does not look like a good part. This is usually called anomaly detection or good-part learning.

The advantage is that almost no NG samples are required, and good parts are abundant, so start-up is fast. The disadvantages are that defects cannot be classified by type, and false rejects tend to be high. Where good parts vary widely, normal variation gets flagged as abnormal.

The pragmatic pattern is therefore two-stage. Start with anomaly detection and route anything suspicious to a human. NG samples accumulate through operation, and as defect types become identifiable, migrate them one by one to a classification model. This combines a fast start with an acceptable end-state accuracy.

Choosing among the three

One rule applies to all three: validation data must consist of real NG parts. Using generated data for training is acceptable. Using it for acceptance testing is meaningless.

Writing down a standard that lives in one person’s head

This is the topic domestic guidance covers least and the one that stalls the most overseas projects.

When the limit samples exist only at the parent plant

Most manufacturers maintain limit samples — physical parts agreed with the customer that mark the boundary between accept and reject. At overseas sites, those limit samples frequently exist only at the parent plant, or the local copy differs subtly from the original, or it has aged and changed colour. In many cases the ultimate authority is not a physical sample at all: the inspection standard lives in one senior operator’s head at the parent plant, and everything downstream is an approximation of that person’s judgement.

Introducing AI visual inspection means redefining that boundary in numbers and images. Put differently, automating on top of an ambiguous limit sample can leave you worse off than before, in a state where nobody can explain whether the model’s verdict matches the customer’s standard.

Four steps to transplant a subjective standard

1. Make the current judgement visible. Have several inspectors independently judge the same 20 to 30 parts. Separate the parts everyone agrees on from the parts that split opinion. A high split rate means no standard exists yet.
2. Verbalise why the split happened. Was it scratch length, scratch depth, or position (cosmetic surface versus non-cosmetic)? Only here do the axes that can be quantified appear.
3. Quantify and write the standard. Length above so many millimetres, area above so many square millimetres, a drawing-based definition of which surfaces are cosmetic, and the rule for combining multiple defects. Always include photographs. A text-only standard does not get used.
4. Agree it with the customer. Skip this and you will eventually be told your standard is not theirs. Changes that tighten judgement — which is what a rise in false rejects amounts to — also need internal agreement, because they hit yield directly.

How to build a multilingual standard document

At Thai and Vietnamese sites, prepare the standard in the group working language, the local language (Thai or Vietnamese) and, where needed, English. What matters is structure, not translation elegance.

• Lead with photographs rather than prose. Show an accept example, a reject example, and always a borderline example.
• Separate pages by judgement axis (scratch, stain, chip, printing).
• Express numbers graphically, using photographs with a scale in frame.
• Eliminate catch-all clauses such as “any other item with significantly poor appearance”. That is not a standard; it is delegation of the decision.

Worth noting: this document has value even if the automation never happens. It shortens inspector training and reduces the quality swing caused by turnover. If the capex slips a year, this workstream is still worth completing.

Treat the grey zone as a third category

Real judgement has clear accept, clear reject, and an ambiguous middle. Design the operation with that middle explicitly routed to a human. Forcing a binary decision turns threshold setting into an internal political argument. With three categories, the target becomes operational: keep the share falling into the middle band below an agreed percentage.

Reading accuracy numbers on a 100% inspection line

When a proposal says 99 percent accuracy, what should you check? The answer largely determines whether the project succeeds.

“99 percent accurate” says almost nothing

Consider a process with a 1 percent defect rate (illustrative example). A model that labels everything as good achieves 99 percent accuracy while catching zero defects. What you need are two different numbers.

• Miss rate: defective parts judged good. This flows straight to the customer and should approach zero.
• False reject rate: good parts judged defective. This hits yield and re-inspection workload.

Moving the threshold trades one against the other. Tighten to reduce misses and false rejects increase. Both cannot be zero simultaneously.

Threshold policy generally falls into three positions.

• Tight, where anything suspicious is rejected: low miss rate, high false reject rate. Appropriate where escape cost is extreme — safety-critical parts, medical, food. Requires more people for secondary check.
• Standard: both rates moderate. The usual landing point for general components with a defined secondary check headcount.
• Loose: high miss rate, low false reject rate. Only defensible when escapes have limited impact, or when a downstream process or the customer’s incoming inspection would also catch them.

Where you sit is a management decision, not a technical one. Decide before go-live who owns the threshold. If production loosens it unilaterally, escapes rise; if quality tightens it unilaterally, the savings evaporate.

Practical design for a line running 100% inspection

Where takt time is fixed, there is effectively one workable architecture: the model performs the primary judgement, and humans perform a secondary check only on parts judged NG or grey.

The consequence is that the headcount calculation changes shape. Inspectors move from viewing every part to viewing only what the system flags.

Hours saved = original inspection hours x (1 minus the share routed to secondary check)

So the higher the share routed to a secondary check, the more of the saving disappears. At a 20 percent secondary check share the workload falls by 80 percent rather than 100 percent; push that share to 50 percent and half the expected saving is gone. “We tightened the threshold chasing zero misses and then could not reduce headcount” is among the most common outcomes.

Requirements definition should therefore fix two numbers rather than an accuracy percentage.

• The maximum acceptable miss rate, derived backwards from the escape loss calculated earlier
• The maximum acceptable secondary-check share, derived backwards from the headcount target

Whether those two can be met is the pass-fail criterion for the PoC.

What to measure at acceptance, and how

Design the measurement before contract signature.

• Build the evaluation set from parts not used in training. Obvious, and frequently violated.
• Include real NG parts. Evaluation on generated data is invalid.
• Measure at real line takt, in continuous operation. Offline still-image evaluation is not sufficient.
• Run the evaluation for at least several days to a week, to capture lot-to-lot variation and shift-to-shift lighting differences.
• Include environmental variation: day and night, rainy and dry season, air conditioning behaviour.

The last point rarely appears in guidance written for temperate-climate domestic plants. In factories with windows or with shutters that open, ambient light changes through the day. Between dry and rainy season, in-plant humidity and condensation change, which can affect lens and part surface condition (Thailand-specific and Vietnam-specific climate factor). A PoC passed on a dry-season day shift and failing on a rainy-season night shift is avoidable, but only by designing the evaluation window deliberately.

From PoC to production: phases, contracts and acceptance

Four phases and their exit criteria

Phase 0, problem definition (roughly 1 to 2 months). Select the target process, calculate the escape loss, measure current judgement variance, prepare the standard document. Exit criterion: the target defect types, the acceptable miss rate and the acceptable secondary-check share are documented.

Phase 1, imaging feasibility (roughly 1 to 3 months). Bring real samples, or image on site, and confirm the defect appears in the image at all. If it does not, the problem is optical, not algorithmic. Exit criterion: images in which the target defect is visible to a human.

Phase 2, PoC (roughly 2 to 4 months). Collect training data, build the model, measure. Exit criterion: the two numbers above are met. If they are not, either narrow the scope, revisit the optics, or stop.

Phase 3, build and ramp-up (roughly 3 to 6 months). Handling and fixture design and fabrication, PLC integration, upper-system integration, installation, trial running, transition to production. Exit criterion: passing continuous-operation evaluation on the real line.

These durations are indicative and swing widely with the difficulty of the part and the defect. Plan on roughly a year end to end, and treat “approve this year, realise the benefit this year” as unrealistic. Where a BOI application or a capital plan is involved, share this timeline early.

PoC traps

• Handing over only the easy samples. Selecting flattering parts out of courtesy to the vendor produces a PoC that passes and a system that fails. Hand over the borderline cases.
• Starting without a defined success criterion. A PoC begun as “let’s see how it goes” ends as “it worked reasonably well”, which cannot support a decision.
• Running the PoC in an environment unlike production. Images captured in a vendor lab are not images captured in your plant. Do it on site wherever possible.
• Squeezing the PoC price too hard. A free or nominally priced PoC gives the vendor no budget for effort, and the quality of your decision inputs falls accordingly.

What contracts and acceptance criteria must say

New variant handling deserves particular attention at overseas sites. Production allocation at a subsidiary plant is frequently changed by the parent company or regional HQ. If the cost of adding a variant is undefined at signature, every model change or capacity shift reopens the investment decision.

Thailand BOI incentives and investment timing

For capital investment in Thailand, whether BOI (Thailand Board of Investment) incentives apply feeds directly into the business case. This area attracts a lot of misunderstanding, so it is worth stating precisely.

Investment activity in the first half of 2026

Investment applications in Thailand in H1 2026 were reported at 43.6 billion USD across 1,299 projects, up 37 percent year on year (Thailand-specific, Thailand Business News / Asia News Network, July 2026). Within that, applications under the BOI Smart and Sustainable Industry measure accounted for 132 projects worth 507.6 million USD; machinery, automation and robotics accounted for 82 projects worth 387.4 million USD; and electrical and electronics accounted for 179 projects worth 3.56 billion USD (all Thailand-specific, same sources). Approved projects overall are expected to create more than 82,000 jobs (Thailand-specific).

For context, large AI data centre investments are what lift the headline figure. The machinery, automation and robotics bracket, which is where an inspection system investment sits, is a small share of the total. Still, the trend supports a reading that institutional backing for automation investment is continuing.

What Smart and Sustainable Industry actually grants

This is the most misquoted point in the area. Corporate income tax exemption under this measure is fundamentally 50 percent of the investment amount (Thailand-specific). It reaches 100 percent only where automation or robotics is introduced into the production line and at least 30 percent of the value of the machinery being upgraded is sourced from Thailand’s domestic automation industry (Thailand-specific). Any statement that BOI makes automation investment 100 percent tax exempt unconditionally is incorrect, and should not be copied into an internal paper.

The 30 percent local-content condition has practical consequences for how you configure an inspection cell. Cameras and AI software are usually imported, whereas the mechanical portion — conveyors, fixtures, frames, control panels — can often be sourced from integrators or machine builders inside Thailand. Whether you meet the condition therefore depends on the combination of system configuration and supplier selection, not on the category of the project alone.

BOI programme terms and eligibility criteria are revised over time. Application categories, eligible machinery definitions, required documentation and the method of proving local content must be checked against current official sources, with a specialist familiar with BOI application practice. The description here outlines the general structure of the scheme and does not guarantee eligibility for any specific case. Note also that this section is Thailand-specific and does not describe incentives available in Vietnam or elsewhere in ASEAN.

Do not miss the application window

Where BOI incentives are used, applications generally must be filed before equipment is ordered and imported. The common failure is that the PoC goes well, the production system is ordered on the strength of that momentum, and BOI is considered afterwards, too late.

The practical countermeasure is to confirm BOI eligibility and the required lead time during Phase 0 or Phase 1. Starting the enquiry after PoC results are in pushes the whole investment schedule backwards. Knowing the conditions early also lets you keep local-content share in mind while the configuration is still on paper.

Designing the operating model: who retrains the model?

This is the least discussed phase and the one where failures concentrate. The system runs well for six months, then a material lot changes, the lighting degrades, a new variant arrives, accuracy drops, nobody can fix it, and everyone goes back to 100% manual inspection. That is the classic ending.

Assume there is no image-AI specialist on site

Few plants in Thailand or Vietnam have a resident engineer who can work with vision models. Recruiting one means competing with IT employers in Bangkok or Ho Chi Minh City, and a factory salary band rarely wins that competition (Thailand-specific and Vietnam-specific labour market condition).

So begin the design from the assumption that no specialist is on site, then choose one of the following, or a combination.

• Choose a no-code or low-code retraining tool: the local quality assurance staff add images and press a button. You sacrifice depth, but the operation keeps running.
• Contract remote maintenance with the vendor: the site sends images, the vendor updates the model. Write the no-travel-required arrangement into the contract rather than assuming it.
• Centralise at the parent company or a regional HQ: with several sites, keep model management in one place and let each site focus on operation. This is the efficient shape when you intend to roll out.

The Cognex 2026 study is suggestive here. Among respondents with three or more years of AI vision experience, 86.1 percent said rollout to multiple sites was easy, against 75.3 percent of those with less experience, a gap of 10.9 points; 81.2 percent said development and deployment were fast, against 72.1 percent, a gap of 9.1 points (all global, multi-region sample). The same research reports that early adoption is motivated by accuracy — detecting fine and complex defects — while the weight given to ease of use, meaning whether the shop floor can run it unaided, rises with years of operation.

That is not a statement about how to choose your first cell. It is a statement about whether you can stand up your second and third yourself. If you do not intend to stop at one line at one site, put “can our own people operate it” into the selection criteria from the start.

What to monitor, and the early signs of drift

Model drift does not arrive overnight; it creeps. Build daily or weekly monitoring into the system at installation.

• NG rate trend: investigate any move up or down from the usual band
• Share of secondary checks where the model said NG and the human said good: indicates rising false rejects
• Customer complaints: the worst possible indicator of rising misses; you do not want to learn this way
• Image brightness and contrast statistics: detects lighting degradation and dirty optics before accuracy falls
• Product mix changes: log new variant introductions and material changes

The last point is a procedural matter. When material, moulding conditions, tooling or surface treatment change, does the information reach whoever owns the inspection system? If that link is broken, unexplained accuracy loss follows. Put the inspection system owner into the change control (4M change management) workflow.

Assign the roles before go-live

The table is only an example. What matters is that the retraining row has a name in it. Projects that go live with that cell empty stop within a year with high probability.

For the wider question of how to embed AI in factory operations, our practical guide to generative AI implementation covers related ground. The technologies differ, but the principle that the design must be something the site can run without outside help is the same.

Turning inspection results into data: MES and traceability

If the entire return on AI visual inspection has to come from headcount reduction, most cases will not clear. The value of the inspection result existing as data belongs in the evaluation.

Design so the results are not discarded

In manual inspection, the judgement happens inside a person’s head and the record retained is a pass-fail flag plus, at best, a defect category. With AI quality inspection in manufacturing, every part leaves an image and a basis for the verdict. That is a meaningful difference.

• Analyse defect occurrence by time, machine, lot and operator
• Detect the moment a defect rate starts to rise and feed it back to the process
• When a complaint arrives, retrieve images for the specific lot
• Prove objectively, during an audit, that inspection was performed

The third and fourth points carry unusual weight at overseas subsidiaries. When a customer or the parent company asks for an explanation of a quality problem, “the inspector checked it” and “here are images of every part in that lot” lead to very different conversations.

Decide the linking key first

To make inspection data usable, decide at design time what the image and verdict are linked to.

• Product serial number, the strongest option where unit-level control exists
• Lot number or work order number
• Machine number, mould number, cavity number
• Date, time, shift, operator ID
• Material lot number

Unit-level identification is the hardest to retrofit. Without serialisation, a defective unit cannot be traced back to an individual item. If lot control already works, starting at lot granularity is perfectly practical.

For how the plant should hold data and collect production results, our overview of production management system implementation for factories in Thailand is a useful companion. Whether the inspection system stands alone or integrates with production management changes what has to be built.

In food and pharmaceutical operations, where lot traceback requirements are strict, inspection images become part of the traceability record; the expected level of rigour is discussed in our article on food factory traceability systems. If you are evaluating RFID as the identification method for individual items, see also our note on RFID in factory inventory management and cost. Inspection data reaches full value only once you can say which physical item it belongs to.

Design the feedback loop into the process

Ideally, results feed back into the process. If defects concentrate in one cavity, stop that mould. If defects rise in a particular time band, look for correlation with temperature or humidity. Operations like these become possible only when the data is structured.

In Vietnam, investment in robotics, inspection systems and data-driven quality management is reported as advancing, led by electronics and electrical manufacturers (Vietnam-specific, Intralink). The same source frames robotics, machine vision, MES and IIoT as contributing jointly to downtime reduction, quality stability and visibility. Treating visual inspection automation not as a standalone machine but as the entry point to the plant’s data platform improves the quality of the investment decision.

Frequently asked questions

How much does AI visual inspection cost?

Reference ranges published by domestic vendors in Japan typically present the scope as AI software, camera, lighting, PC and integration fee (Japan-specific, NTT Communications, 2025), which is a useful starting structure. However, that composition usually excludes handling, fixtures and mechanics, which a real production cell always needs. There is also very little reliable public statistical data on visual inspection automation cost in Thailand or Vietnam (no reliable public data for Thailand or Vietnam). Rather than chasing a market rate, split the scope into the six line items described above, obtain quotations at identical granularity from several suppliers, and compare on a five-year total.

Can we replace manual visual inspection entirely?

A plan that assumes total elimination is not recommended today. The realistic architecture is a primary judgement by the system with human secondary check on NG and grey verdicts only. Even so, headcount falls substantially if the secondary-check share is low. Because a high false reject rate cancels the saving, write both a miss rate ceiling and a secondary-check share ceiling into the requirements. Detecting the vague sense that something is wrong, and reacting to unforeseen defect modes, should be assumed to remain human work for the time being.

Can we start with very few defect samples?

Possibly, but the approach has to change. Anomaly detection, where the model learns good parts and flags anything that does not look like one, can be stood up with almost no NG samples. The trade-offs are that defect types cannot be classified and false rejects tend to rise. The paradox that well-controlled sites have low defect rates and therefore few NG samples is widely noted (Nsight), so this situation is not unusual. In practice, start with anomaly detection, accumulate NG samples through operation, and migrate identifiable defect types to a classification model as they emerge.

What accuracy percentage should we accept?

Do not judge on overall accuracy. On a process with a 1 percent defect rate (illustrative example), labelling everything good yields 99 percent accuracy. The two numbers that matter are the miss rate and the false reject rate, and each ceiling is derived from your own situation. The miss rate ceiling comes from the point at which escape loss per incident multiplied by expected shipment volume stays within tolerance. The false reject, or secondary-check share, ceiling comes from the headcount reduction you need. Whether those two can be written numerically into the acceptance clause largely determines the outcome of the project.

Rule-based machine vision or AI — which should we choose?

If the criterion can be written numerically, rule-based automated optical inspection is enough: faster, cheaper, explainable, and adjustable by a local engineer. AI earns its place where good parts vary widely, the defect looks different each time, and no numeric rule can be written. Most lines contain both, so a hybrid — rule-based positioning and clear-cut defects, model-based hard judgement — is the practical configuration. The narrower the AI scope, the more the site can operate unaided.

How long does it take from decision to production?

Across the four phases of problem definition, imaging feasibility, PoC and production build, plan on roughly a year (planning guidance, not a benchmark). It varies widely with the difficulty of the part and the defect, but “approve in this fiscal year and realise the benefit in this fiscal year” is unrealistic. If BOI incentives are in scope, the application generally has to precede the equipment order (Thailand-specific), so confirm the schedule early.

Can we operate it without an image-AI specialist on site?

Design on the assumption that you do not have one. There are three options: select a no-code tool where local QA staff add images and press a button to retrain; contract remote maintenance so the site only sends images; or centralise model management at the parent company or a regional HQ. In every case, name whoever performs retraining in the responsibility matrix, and fix the annual number of events and the method in the contract.

What happens when new product variants are added?

Each new variant needs image collection and retraining. Contracts that do not anticipate this become a source of friction. Specify the cost per added variant, the number of images required and the lead time at signature. Because production allocation at overseas subsidiaries is often changed by the parent company or regional HQ, this clause has an outsized effect on long-run cost.

Conclusion: get the order of decisions right

The most common failure in evaluating AI visual inspection is starting from a technology comparison. The order proposed in this article, restated:

1. Calculate the loss per escaped defect from your own records
2. Inventory the true cost of manual inspection, well beyond direct payroll
3. Narrow the target defect types and separate what needs AI from what does not
4. Verbalise the inspection standard and publish it as a multilingual, photo-led document
5. Fix the acceptable miss rate and acceptable secondary-check share as numbers
6. Obtain quotations split into the six line items and compare on a five-year total
7. Choose your defect-sample strategy: accumulate, generate, or design around it
8. Put the retraining owner and the monitoring mechanism into the contract
9. Confirm BOI eligibility and application timing before placing any order (Thailand-specific)
10. Decide the linking key for inspection data and design the connection to production management and traceability

Of these, item 4 (writing the standard down) and item 8 (the operating model) are rarely emphasised in guidance written for domestic plants, yet they are what decides success at ASEAN sites. Both also deliver quality assurance benefits on their own, even if the capex is deferred. Starting there is not a way of postponing the investment decision; it is a rational sequence.

The global market for AI visual inspection systems is forecast to grow at 23.3 percent annually and reach 85.24 billion USD by 2030, with Asia-Pacific the fastest-growing region (global, The Business Research Company, July 2026). A growing market, however, is not a reason for your plant to invest. The justification sits inside your own loss figures and cost structure.

If you are at the stage of working out where to start, or whether a given process is even a candidate, we are happy to talk it through — no quotation request required. From our work implementing production management and factory IT systems at plants across Thailand and ASEAN, we can help structure the questions in a way that reflects local conditions. You can reach us through the contact form whenever it is useful.