Inspection Equipment Suppliers 2026 — Decide This Before You Compare

“We can’t find people for visual inspection.” “Two inspectors look at the same part and reach different conclusions.” “A defect escaped to the customer and turned into a complaint.” Many plants start comparing inspection equipment manufacturers from exactly this position. But in the projects we support at Japanese-affiliated factories in Thailand, what we notice again and again is that a set of internal decisions is still unresolved before anyone starts comparing machine specifications. What counts as a defect? 100% inspection or sampling inspection? How many false rejects are acceptable? Collect quotations while those questions are open, and the proposals that come back simply cannot be compared. This article sets out what to settle before the comparison begins.

What this article covers, and what it does not

Articles about inspection automation tend to fall into one of two shapes: a technical tour of the equipment, or a broad “AI is changing inspection” narrative. Yet what actually derails real projects is neither the technology nor the narrative. In our experience it is almost always one thing — the buyer has not finished defining “defect.” Sign a contract with that gap still open, and acceptance testing turns into an argument with the equipment supplier over whether a given part is good or bad.

Here is the scope of this article.

If you want the wider investment framework for automation first, Labour-Saving Automation: Cases and Cost-Effectiveness sets out how to approach automation when labour supply is the constraint. This article goes deeper into one particular process — inspection — which is the hardest part of the plant to define precisely.

One note before we start: unless stated otherwise, every monetary figure in this article is a published benchmark from Japan. These are not numbers you can transfer directly to a Thai plant, and we return to that point later.

Why inspection automation enquiries are increasing

Labour supply and the investment climate in Thailand

Let’s start with the facts behind the trend. If you justify an investment on a vague sense that “there’s a labour shortage, so let’s automate,” you will not be able to build a defensible number for the capital request.

In Thailand, people aged 60 and over now exceed 21% of the population, and labour supply is tight. Rising labour costs, competition from neighbouring ASEAN countries, currency movements and shifting trade conditions are all pushing manufacturers to re-examine their cost structures (source: KAIZEN, “Thailand Manufacturing Trends 2026”).

On wages, Thailand’s 2026 minimum wage runs from THB 337 to THB 400 per day depending on province, with the THB 400 ceiling unchanged from 2025 (source: Thai Law Online). Taken alone, an unchanged ceiling reads as “labour cost is flat” — but what actually bites in day-to-day operations is not the minimum wage itself. It is the total cost of recruiting, training and retaining people who can perform visual inspection, and that never shows up in the minimum wage figure.

• Recruitment: monotonous, eye-straining work tends to attract few applicants
• Training: it takes time to bring someone up to speed on limit samples
• Retention: people you have trained leave for other companies
• Variation: judgements shift when the person changes, and customers notice
• Peak demand: you cannot scale inspectors alone during a ramp-up, so the process becomes the bottleneck
• Records: defect details stay on paper, and compiling them costs effort

There is a second pattern worth noting: the shape that Thai manufacturing is settling into. Rather than full automation, the mainstream is a hybrid of human skill plus partial machine support, with robot arms, automated transfer and semi-automated inspection technologies being built into lines. More than 40% of large Thai manufacturers are reported to have adopted at least one Industry 4.0 technology (source: KAIZEN).

That “hybrid is the norm” point matters for inspection projects too. Set unmanned operation as the goal from day one and the requirements balloon immediately. Assuming people remain, and then deciding which parts of their burden to hand to a machine, tends to produce a far more realistic specification.

Production is not strong

You also need to look at the current production environment before making an investment case. Thailand’s Manufacturing Production Index (MPI) for May 2026 was down 0.8% year on year, and automotive production came in at 123,276 units, down 11.4% (source: Bangkok Shuho).

In other words, a payback plan built on rising production volume is a hard sell in the current environment. “To handle increased output” is a justification that will be challenged both in local approval and at head office. Reasons that hold up even when volume is flat — eliminating judgement variation, reducing escapes to customers, automating records, redeploying inspectors to other processes — are more likely to get through.

The machine vision market

The supply side is worth a look as well. The industrial machine vision market is forecast to grow from USD 15.83 billion in 2025 to USD 23.63 billion in 2030, a CAGR of 8.3% (source: MarketsandMarkets).

A growing market means two things at once: more options, and more variation in the quality of proposals. When the number of players rises, approaching them without firm requirements produces a chorus of “yes, we can do that.” Which is exactly why the buyer needs its own decision criteria.

Edge technology is moving too. As of 2026, mainstream edge AI devices include NVIDIA Jetson AGX Orin, Hailo-15 and Qualcomm RB6. They allow real-time machine vision inspection with no network latency, and product images do not need to be sent outside the plant (source: koromo). At overseas sites, whether product images may leave the country for a cloud service sometimes becomes a live issue with customers, so it is worth confirming early in the specification stage.

At the same time, the limits of human visual inspection have been re-examined. Even experienced inspectors cannot fully prevent escapes, and the main causes are fatigue, declining concentration and individual differences in judgement criteria (source: OptiMax). One caution here: this does not mean “install a machine and escapes disappear.” Machines have their own escapes and their own false rejects. Human limits and machine limits are simply different in kind.

Five things to settle internally before choosing an inspection equipment manufacturer

This is the core of the article. In our experience, almost every inspection automation project that stalls does so not because the wrong equipment was selected, but because something was left undecided internally before the order was placed. There are five such items.

These five are not decisions the equipment manufacturer can make for you. A supplier can decide how to achieve something; it cannot decide what you call a defect. Hand this over wholesale, and the supplier will propose the definition that is easiest for it to implement. Whether that definition matches what your customer demands is an entirely separate question.

Translating accept/reject criteria from “words” into numbers and images

This is where the most disputes arise. Plenty of factories have inspection standards that read “no scratches,” “no contamination,” “no burrs.” Hand that wording to an equipment manufacturer as-is and trouble downstream is likely.

The reason is simple: the word “scratch” specifies nothing to a machine. A vision system perceives differences in light reflection, differences in brightness, changes in shape — all as numbers. If there is no physical quantity corresponding to “scratch,” there is no basis for setting a threshold.

Here is what tends to happen when the criteria go out as words.

So how do you get there? The approach we recommend on site runs in the following order, and it requires no special tooling.

• Gather a year’s worth of defective parts, plus physical parts returned by customers
• Standardise the names used for defects across the plant (many factories have three names for the same phenomenon)
• For each name, select a limit sample — the physical part that sits at the boundary of acceptable
• Photograph the limit samples under conditions close to how the system will actually image them
• Measure length, area, contrast difference and so on directly on those images
• Fix the number that says “acceptable up to this value,” using QA judgement and customer requirements
• Replace the standard with one that pairs numbers with images

Ideally this work finishes before you request quotations. In practice, though, you often discover only after photographing a limit sample that “we can’t capture this without changing the lighting.” In that case you firm up the criteria while running Stage 1 (sample evaluation, described later) together with equipment manufacturers. The important thing is to recognise that ownership of the criteria-setting work sits with you, not the supplier.

Keeping limit samples as image data pays off considerably later. When you replace equipment, add a product model, or roll out to another site, whether that image asset exists changes the up-front effort noticeably. Always confirm at contract stage who owns the images the supplier captures and whether you may reuse them (this also appears in the contract checklist below).

100% inspection automation, or sampling?

The next decision is inspection coverage, and it drives the investment figure heavily.

We receive a lot of enquiries that begin “we want to automate 100% inspection.” But going to 100% binds the equipment to takt time. Imaging, judgement and ejection all have to complete within the production cycle, and that constraint dictates both configuration and price. With sampling inspection, the time constraint loosens, which can allow more thorough inspection or more expensive measurement techniques.

The right order of reasoning starts with what the customer actually requires. Separate the items where 100% inspection is explicitly written into drawings or the quality agreement from the items your plant inspects 100% “just to be safe.” The latter may have room to move to sampling.

• Quality agreements and drawing notes: which items carry a 100% inspection requirement
• Past corrective action reports: is there a lingering promise to “perform 100% inspection”
• Process FMEA: where does detection depend on inspection
• Defect occurrence rate and pattern: sporadic, or clustered by lot
• Cost of detection downstream or at the customer: what is the loss if it escapes

The fourth point is easy to overlook. If your defects tend to cluster within lots, sampling inspection may well catch them. Defects that appear sporadically from an unknown cause will not be caught by sampling. Which pattern applies to you can be determined from past records. Skip that check and default to “100%, to be safe,” and you will over-invest.

Which do you accept: false rejects or escapes?

This is the fork in the road that determines whether you can write acceptance criteria at all. It is also the item least often discussed internally.

An inspection system’s judgement has two kinds of error, by definition.

The critical point is that these two are a trade-off. Tighten the judgement and escapes fall while false rejects rise. Loosen it and false rejects fall while escapes rise. No threshold exists that drives both to zero — not as a practical matter, but as a matter of principle.

So the buyer must decide first: which one, and to what extent, will you tolerate? Write “detect 100% of defects and reject no good parts” into a specification without settling this, and no equipment manufacturer can agree to it. Even if one appears to agree, the arrangement will break down at acceptance.

Once the policy is set, you can write acceptance criteria along these lines. The actual numbers come from your own requirements and agreement with the supplier, so what follows is the framework, not the values.

• Composition of the evaluation sample set (N good parts, N per defect type, N boundary samples)
• Who provides the samples and when the set is frozen (adding samples later is a reliable source of disputes)
• Escape tolerance (how many of the specified defect samples may pass)
• False reject tolerance (how many good samples may be rejected)
• How judgement stability is verified (agreement rate when the same samples are run repeatedly)
• Re-verification under environmental variation (external light, temperature, contaminated surfaces)
• Remedies if the above are not met (adjustment period, cost responsibility, final disposition)

Of these, “responsibility for providing samples” and “re-verification under environmental variation” matter especially at Thai plants. Add samples after the fact and the supplier will say it was never told. And because factory conditions vary more than in Japan, it is safer to write verification under real environmental variation into acceptance, rather than testing the machine in isolation.

Handling and recording after a reject decision

This is the single biggest reason projects end with “we installed the machine but headcount didn’t drop.”

After the system displays “NG,” where does that part go? Who checks it? How is it recorded? Install equipment without designing this, and the shop floor will, almost without exception, create a routine of “let’s have someone look, just to be safe.” At that moment, the labour saving evaporates.

Take particular care over the fifth row, “how re-check results are treated.” Whether you permit a person to override an NG judgement and ship the part goes to the heart of quality assurance. If you do permit it, document who holds that authority and how it is recorded — otherwise it is an easy finding in a customer audit.

The record design from row seven onwards actually strengthens the investment case too. Once defect details are compiled automatically, you gain material for process improvement. Inspection equipment is not a machine that reduces defects, but it can be a machine that captures defect data. That is a usable argument in a capital request. For designing the loop that feeds data back into the process, Making Use of Factory Automation Consulting in 2026 covers cross-process design thinking as well.

Scope of product models, and the scope humans keep

The last item. The more high-mix, low-volume your plant is, the more this swings the quotation.

• Models in initial scope: how many models, in descending order of volume
• Models added later: are the cost and lead time per additional model included in the quotation
• Changeover effort: how many minutes, and who is able to do it
• Common fixtures and transfer: does each model need a dedicated jig
• Selection of inspection items: the line between what the machine takes and what people keep
• Items the machine cannot handle: feel, sound, smell, functional testing
• Redesign of the remaining manual process: does the layout still work for the people who stay

Always confirm “the cost of adding a model” at quotation stage. Compare on initial installation cost alone and you may pick a configuration that levies a heavy charge every time a model is added. With deep learning configurations in particular, adding a model means retraining, which can mean starting again from data collection.

“Items the machine cannot handle” matters too. A visual inspection system can only judge what is optically observable. Roughness detected by touch, abnormal noise during assembly, stiffness in a fit — these cannot be judged from images, or require a different method entirely. Declaring them out of scope and leaving them with people is not a retreat. On the contrary, naming what is out of scope sharpens the specification.

Visual inspection and machine vision approaches, and where each fits

Once the five internal items are settled, the conversation moves to method. Only now do we reach “rule-based or deep learning.” We often see this considered in the reverse order, but method is chosen after the objective is fixed.

When rule-based (conventional machine vision) fits

Start from this premise: most inspection items can be handled rule-based. Before concluding that deep learning is required, we recommend checking whether the problem can be solved with rules. The reasons are start-up speed, explainable judgement, and a wider range of adjustment your own team can perform locally.

• The defect appears consistently (missing parts, wrong orientation, presence or absence of a hole)
• It can be defined numerically (exceeding a specified dimension, area threshold, colour difference)
• Imaging conditions can be fixed (a fixture locates the part, external light can be blocked)
• The basis for judgement must be presented (items reported to the customer as measured values)
• Defect samples cannot be gathered (the defect rate is extremely low and physical examples barely exist)
• The product changes often (frequent design changes, and you want to avoid repeated retraining)

The fifth point is especially important. In a process with a low defect rate, there simply are not enough defect images to train on. When “let’s use AI” comes up, the first thing to check is how many defect samples you physically have.

When deep learning (AI visual inspection) fits

That said, there are judgements that rules cannot fully express.

• The defect has no fixed shape (irregular scratches, mottled contamination, casting skin anomalies)
• Normal variation is wide (wood grain, textiles, cast surfaces — the base pattern itself varies)
• Writing thresholds keeps generating exceptions (rules swell to dozens and become unmaintainable)
• Experienced staff say “I can just see it” (unarticulated, but their judgement is consistent)
• Training data can be prepared (enough real or reproduced defects can be collected)

The last condition is where reality bites. What to do when defect samples cannot be gathered is the question we are asked most often in AI visual inspection projects. Here is how we lay out the options.

To be candid, we do not recommend pushing ahead with deep learning when you have only a handful of defect samples. A classifier built on insufficient training data will behave unpredictably the first time it meets an unexpected appearance on the real line — and you cannot predict which way it will fail. In that situation, installing a “capture-only” unit to build an image library, or breaking inspection down and starting with the parts that rules can solve, tends to get you there faster in the end.

Whether judgement stays inside the machine or images go outside is another design fork. As noted above, edge AI devices allow real-time inspection with no network latency and remove the need to transmit product images externally (source: koromo). Where a customer’s information security requirements are strict, that configuration can be a precondition rather than an option.

Treat automated dimensional inspection as a separate problem

We are often asked to combine visual and dimensional inspection in one machine. It is not technically impossible, but what has to be managed is completely different, so we recommend treating them as separate evaluations.

Automating dimensional inspection does not end when the machine is bought. You have to fold it into your metrology management system, including the calibration schedule, control of reference standards, retention of calibration records, and named responsible personnel. Skip that and you are stuck the moment a customer auditor asks to see the calibration record for that instrument.

• Is resolution adequate against tolerance (the ratio of measurement variation to the tolerance band)
• Calibration frequency and method (who does it, with what, on what cycle)
• Who performs calibration (in-house, the equipment maker, or an external calibration body)
• Can calibration be done within Thailand (shipping to Japan means long downtime)
• Record retention (where calibration certificates are kept and who manages them)
• Measurement system validation (the plan if a customer requires it)
• Temperature environment (are there temperature swings in the plant that affect accuracy)

“Can calibration be done within Thailand” is easily overlooked at overseas sites. If the configuration assumes shipping the instrument to Japan each time, you need an alternative for that period. At quotation stage, we recommend confirming whether maintenance and calibration can be completed locally.

Dividing roles between poka-yoke devices and inspection equipment

Finally, a point about sequence. Whenever we are consulted about inspection equipment, there is one question we always ask at least once: “Could that defect simply not be produced in the first place?”

In sequence, poka-yoke comes first. There are three reasons. First, it is usually cheap, so the barrier to starting is low. Second, if defect occurrence itself falls, both the performance demanded of the inspection system and the downstream re-checking effort fall with it. Third, the process of designing poka-yoke organises your understanding of how defects arise — which in turn clarifies the requirements for the inspection system.

If most of your defects fall into the first four rows, we recommend looking at poka-yoke before a visual inspection system. If material-origin or handling-origin defects dominate, inspection equipment is the right answer. Classifying your own defects against this table is one of the things you can do in-house before contacting any supplier.

How to read costs and payback

Cost breakdown (Japanese benchmarks, and the caveats when applying them to Thailand)

Now to cost. First, the published benchmarks from Japan. For AI visual inspection, a PoC (trial validation) is put at JPY 3–5 million, and full deployment at JPY 10 million and above (sources: Nsight, koromo).

Example breakdowns are published as well, and we cite them directly. These amounts do not transfer to a Thai plant as-is, but they are useful for understanding how the cost divides into line items.

(sources: Nsight / koromo)

Several things can be read out of this breakdown.

• The hardware share is lower than most people expect. In the single-line example, camera plus PC is JPY 1.1 million — a little over 20% of the JPY 4.6 million total
• AI development is the largest item. JPY 2 million in the single-line example and JPY 3.5 million in the multi-line example; the biggest line in both
• Installation is not negligible. JPY 1 million and JPY 2 million respectively are booked for it
• Maintenance and model additions appear as separate items. Compare on initial cost alone and you will miss them

Now the caveats when translating this to a Thai plant.

The bottom two rows are items that do not appear in any published breakdown. Jigs and transfer mechanisms can cost more than the inspection machine itself. Retrofitting an existing line in particular requires a mechanism to bring the product to the imaging position in a stable orientation, and that is where the engineering hours go. When comparing quotations, make absolutely clear which side owns this scope.

Labour cost savings alone will not carry the case

Build the payback calculation purely on labour cost reduction and, in most cases, it will not stand up. Three reasons.

• Re-checking staff remain: if people inspect the false rejects, the headcount saving shrinks
• The “saved” people just move to another process: actual payroll spend does not fall, which is hard to explain internally
• The wage level differs from Japan: at Thai wage levels, the denominator of the saving is smaller

On the third point, the fact is that Thailand’s 2026 minimum wage runs from THB 337 to THB 400 per day by province, with the THB 400 ceiling unchanged from 2025 (source: Thai Law Online). On that basis, payback calculations presented in Japanese articles rarely hold up in Thailand. Copy a Japanese case study straight into an internal document and finance will question the assumptions.

So how should it be built? When we help structure this, we break the benefit into the following items and ask the client to fill in their own numbers.

Rows two and three are the strongest numbers you have, if the historical record exists. At plants that have actually gone through a complaint response or a full sorting exercise, the cost of a single event can approach the price of the equipment. If those records do not exist, start keeping them now. On measuring labour-saving effects specifically, Labour-Saving Automation: Cases and Cost-Effectiveness also covers the approach.

If you factor in BOI incentives

For investment in Thailand, incentives from the BOI (Thailand Board of Investment) can affect the calculation. This needs care, though.

First, the facts. BOI Announcement 4/2569, published in the Royal Gazette on 31 March 2026, promotes the use of automation and robotic systems to raise production efficiency in the automotive industry. In addition to import duty exemption on machinery, it grants a 50% reduction in corporate income tax for three years on the amount invested in automation and robotic systems (excluding land and working capital). Applications are accepted until the end of 2027 (source: Tilleke & Gibbins).

Separately, on 15 January 2026 the BOI overhauled its investment promotion measures, strengthening tax incentives and widening eligibility in areas including advanced manufacturing, automation, EVs and high-value-added R&D (source: Alvarez & Marsal). Investment activity has been brisk: BOI investment promotion applications in 2025 numbered 3,370 cases worth THB 1.876 trillion, up 67% year on year (source: Nation Thailand).

When bringing incentives into your evaluation, we recommend this sequence.

• Confirm whether your business and investment fall within scope (this varies by industry, investment content, location and timing of application)
• Confirm the application deadline (applications under Announcement 4/2569 are accepted until the end of 2027)
• Confirm which portion of the investment is eligible (land and working capital are stated as excluded)
• Model both with and without incentives (a plan that assumes approval collapses if approval is denied)
• Confirm the documents and lead time required (the order of equipment ordering versus application timing can become an issue)
• Confirm directly with the relevant authority or the BOI (do not decide on published information alone)

Eligibility varies by industry, investment content, location and application timing, so please confirm directly with the relevant authority or with the BOI. The description in this article is a summary of publicly available information and is not a guarantee of applicability to any individual case.

Internally, we recommend putting both the with-incentive and the without-incentive case side by side in the capital request. Base an investment decision on incentives and then discover you are ineligible, and the whole plan has to be rebuilt. Conversely, if the project stands up without incentives, any incentive becomes upside.

Choosing an inspection equipment manufacturer: seven criteria

With the five internal items settled, the method direction visible and the cost structure understood, you can finally start comparing suppliers. These are the seven criteria we use.

Criterion 1 is, in our view, the first real signal about who you are dealing with. A supplier who looks at your standard and says “you haven’t defined what a scratch is” will be easier to work with later. If everything you hand over comes back as “we can accommodate that,” the requirements work may simply have been deferred, not done.

Strengths and weaknesses by type of provider

Providers of inspection equipment come in several types. Even when they are all grouped under the same label of “automation equipment manufacturer,” their areas of strength differ considerably. This is not about naming companies — it is a general tendency by type.

No single type is the right answer. That said, we would advise avoiding a structure where it is unclear who ultimately owns responsibility for the inspection result. If image processing sits with company A, transfer with company B and installation with company C, then when false rejects appear, everything stalls in the search for the cause. Establish a single point of contact before contracting.

Items to confirm in the quotation and contract

Here is a practical checklist of what to confirm in quotations and contracts.

We would single out row seven (threshold change authority), row nine (language) and row eleven (maintenance). Equipment lacking all three tends to stop at Thai plants. If you cannot adjust thresholds, cannot read the screen and have no one in-country who can fix it, the shop floor will revert to visual inspection the moment something goes wrong.

Information to give every bidder on identical terms

The biggest reason competitive quotations fail to compare is that each bidder received different information. Show company A the limit samples, give company B only the drawings and brief company C verbally, and the proposals come back built on different assumptions.

Here is what should go to every bidder identically.

The last row is unglamorous but effective. Dictate how the quotation is broken down and you can put every supplier’s numbers side by side. Leave it open and company A quotes a lump sum while company B provides a detailed breakdown, and comparison costs you hours. Using the breakdown example above (camera, processing PC, development, installation, maintenance, model additions) as a template is one practical option.

It is equally important to give an identical sample set to every bidder. Record the count and types you handed over, and do not add to it afterwards. The moment you do, the comparison is broken.

Failure patterns that are especially common at Thai plants

Much of the above applies equally in Japan, but Thai plants have failure patterns rooted in local circumstances. Here is what we have actually seen.

Row three, staff turnover, has to be treated as a real risk at Thai plants. Avoid an operating design that depends on one person. We recommend targeting at least two capable people for each of: operating the equipment, adjusting thresholds, daily checks, and simple troubleshooting. Alongside that, keeping procedures in a form you can update yourself — videos, Thai-language work standards — means operation survives staff changes.

Row five, sharing limit samples, matters just as much. If the physical limit samples approved by head office in Japan exist only in Japan, local judgement will be made according to local interpretation. It is worth establishing, at the same time as the equipment goes in, a rule to share them as images across both sites and to reflect every update in both places.

On row nine, we will be honest: this is a hard problem. When head office decides the model, all the local site can do is present the conditions as numbers. Takt time, number of models, external light conditions, the maintenance structure, the languages that can be supported — present these as facts, and head office has something to work with. “It doesn’t suit us here” as a subjective statement will not change the plan. The same dynamic shows up across automation generally, and Collaborative Robots: Costs and Implementation Approach covers similar ground.

How to stage the rollout

Finally, the stages of the approach. Rather than jumping straight into 100% inspection automation, we recommend working in steps.

There are three reasons to stage it. First, Stages 0 and 1 expose almost every gap in your internal criteria. Actually photographing samples produces discoveries such as “this defect doesn’t show under the current lighting” or “inspectors were judging this item differently from each other.” Second, Stage 2 lets you experience the real false reject level. Hearing a number and watching good parts get rejected on your own floor land very differently. Third, it limits the loss if things go wrong.

The urge to skip Stage 2 and go straight to Stage 3 is understandable, but line integration is hard to undo. If you can secure a window to trial it offline, it is safer to take it.

Here is what tends to happen when stages are skipped.

• Skip Stage 0: you argue with the supplier at acceptance over whether a part is defective
• Skip Stage 1: after contracting, you discover “this defect cannot be captured”
• Skip Stage 2: false rejects surge right after go-live and the line stops
• Under-designed Stage 3: no labour saving appears and the investment cannot be justified
• Under-prepared Stage 4: you repeat the same failure at the next site

Frequently asked questions

How much does a visual inspection system cost?

The published benchmarks from Japan put an AI visual inspection PoC at JPY 3–5 million and full deployment at JPY 10 million and above. Example breakdowns give a single-line configuration at approximately JPY 4.6 million (high-resolution camera JPY 600,000, processing PC JPY 500,000, AI development JPY 2 million, installation JPY 1 million, one year of maintenance JPY 500,000) and a multi-line full deployment at approximately JPY 10 million (camera subsystem JPY 2 million, server JPY 1.5 million, AI development JPY 3.5 million, installation JPY 2 million, maintenance and model additions JPY 1 million) (sources: Nsight, koromo).

However, these are Japanese benchmarks and do not transfer directly to a Thai plant. Hardware is largely imported and exposed to duties and exchange rates, while installation and development labour costs and the maintenance structure all differ. These breakdowns may also exclude jigs and transfer mechanisms — and in a retrofit to an existing line, that portion can account for a large share. Actual figures vary with the models in scope, the inspection items, takt time and the installation environment, so the reliable approach is to firm up your internal standard first and then take quotations from several suppliers.

How realistic is 100% inspection automation?

It is realistic if you narrow the models and inspection items. Try to automate every item on every model at once, and the requirements swell and the investment jumps with them.

A practical approach is to start with the highest-volume models and the inspection items that are easiest to define numerically. Items such as missing parts, part orientation and the presence or absence of a hole tend to come up relatively easily, while irregular scratches and subtle colour differences are considerably harder. And some items — surface feel, abnormal noise — cannot be judged from images at all as a matter of principle. Leaving those explicitly with people makes the equipment specification clearer.

Also, separate the items where the customer requires 100% inspection from the items your plant inspects 100% just to be safe. The latter may be convertible to sampling inspection, which changes the scale of the investment.

What is the difference between a machine vision inspection system and AI visual inspection?

The judgement mechanism differs. A conventional machine vision inspection system (rule-based) judges against numerical thresholds on dimension, area and brightness. AI visual inspection (deep learning) judges from patterns learned across training images.

As a rule of thumb: if the defect appears consistently and can be defined numerically, use rule-based; if it is irregular and hard to quantify, consider deep learning. But deep learning needs training image data. In a process with a low defect rate and only a handful of defect samples, you will run straight into the problem of insufficient training data.

They also differ in how easily a judgement can be explained. Where you must present the basis of a judgement to a customer, rule-based tends to be easier to explain. If you use deep learning, decide before implementation how you will present that basis. In practice, configurations that combine both — assigning the better-suited method to each item — are also used.

Which should come first, poka-yoke devices or inspection equipment?

As a general rule, poka-yoke comes first. A poka-yoke (mistake-proofing) device exists to prevent defects from being made; inspection equipment exists to prevent defects from getting out. If defect occurrence itself falls, both the performance demanded of the inspection system and the hours spent re-checking false rejects fall with it.

To decide, look at the breakdown of your own defects. If human-error defects dominate — missing parts, wrong orientation, skipped sequence steps, missed fasteners — poka-yoke is likely to address them. If your defects are mainly appearance defects originating in the material, or scratches occurring during transfer, occurrence is hard to stop and inspection equipment is the answer.

On cost, poka-yoke can usually be started smaller. Some items can be handled with a single sensor and a change to the sequence control. And because designing poka-yoke clarifies how defects arise, the requirements for any subsequent inspection system become clearer too.

If we cannot gather defect samples, should we give up on AI visual inspection?

You do not have to give up, but we would not recommend proceeding with deep learning as things stand. A classifier built on insufficient training data behaves unpredictably when it meets an unexpected appearance on the real line.

The options include: considering an approach that trains only on good-part images and treats outliers as anomalies; artificially reproducing boundary-level defects and photographing them; starting a retention routine now for scrap and customer returns; or installing a unit that makes no judgement and simply accumulates images with lot information. Alternatively, you can break inspection into items and automate only the quantifiable parts with rules.

Frankly, when the conditions are not in place, “defer implementation for now” is often the correct decision. Even then, you can start today on the preparation — beginning to retain defective parts, and digitising limit samples as images.

Why did headcount not fall after we installed inspection equipment?

In most cases, because the flow after an NG judgement was never designed. If it is not settled who checks the part after the machine outputs “NG,” where it goes, and how it is recorded, the shop floor will create a routine of “let someone look, just to be safe.” If false rejects are frequent, that re-checking alone consumes a fixed amount of labour.

The remedy is to design the post-NG handling before implementation. The physical ejection method, the temporary storage location, whether re-checking is required and by whom, how the re-check result is treated (may a part a person judges good be shipped?), feedback to the equipment, and the content and destination of records — settle these before you place the order.

A second cause is contracting without deciding the acceptable false reject level. A specification of “reject no good parts and detect 100% of defects” cannot hold as a matter of principle. You have to decide first which side you tolerate and how far, and write it into the acceptance criteria.

Can we evaluate automated dimensional inspection together with visual inspection?

Sometimes one machine can do both, but we recommend evaluating them separately. Visual inspection outputs an OK/NG judgement, referenced against limit samples and agreed thresholds. Dimensional inspection outputs measured values, referenced against drawing tolerances and measurement uncertainty.

When you automate dimensional inspection, after installation you must fold it into your metrology management system: the calibration schedule, control of reference standards, retention of calibration records, and named responsible personnel. A customer auditor may well ask to see the calibration records.

The point to check most carefully at an overseas site is whether calibration can be completed within Thailand. If the configuration requires shipping the instrument to Japan each time, you need an alternative for that period. Confirm where maintenance and calibration will be performed at quotation stage.

Do our internal criteria have to be finalised before we approach equipment manufacturers?

Ideally yes, but in reality there is a lot you only learn once you photograph the limit samples. Discoveries such as “this defect doesn’t show under the current lighting” or “inspectors were judging this item differently” only emerge from actual imaging.

So a practical approach is to progress Stage 0 (standardising defect names, organising limit samples, deciding the false reject policy) as far as you can, then firm up the criteria while running the Stage 1 sample evaluation together with equipment manufacturers.

That said, hold on to the understanding that ownership of the criteria-setting work sits with you. A supplier can decide how to achieve something; it cannot decide what you call a defect. Hand that over wholesale and you will be offered the definition that is easiest for the supplier to implement. Whether it matches what your customer demands is another matter entirely.

Can we include BOI incentives in the payback calculation?

You may, but we recommend modelling both with and without incentives. A plan predicated on an incentive has to be rebuilt the moment you learn you are ineligible.

As a matter of fact, BOI Announcement 4/2569, published in the Royal Gazette on 31 March 2026, promotes the use of automation and robotic systems to raise production efficiency in the automotive industry, granting — in addition to import duty exemption on machinery — a 50% reduction in corporate income tax for three years on the amount invested in automation and robotic systems (excluding land and working capital). Applications are accepted until the end of 2027 (source: Tilleke & Gibbins). On 15 January 2026 the BOI also overhauled its investment promotion measures, strengthening tax incentives and widening eligibility in areas including advanced manufacturing and automation (source: Alvarez & Marsal).

However, eligibility varies by industry, investment content, location and application timing. Do not decide on published information alone; confirm directly with the relevant authority or with the BOI. The sequence of equipment ordering versus application timing can also become an issue, so we recommend checking early in the investment planning process.

Summary

What trips up the comparison of inspection equipment manufacturers is not differences in machine performance. In our experience, most projects stall because the buyer has not finished deciding what it calls a defect. Hand over a standard that still reads “no scratches” and “no contamination,” and acceptance testing turns into an argument with the supplier over whether a given part is defective.

Five things to settle internally before the comparison: 1) translate accept/reject criteria into numbers and images; 2) decide between 100% inspection and sampling inspection; 3) decide whether false rejects or escapes are the lesser evil; 4) design the handling and recording of rejected parts; 5) fix the models in scope and the range humans keep. Items 3 and 4 are the ones most often missed. False rejects and escapes are a trade-off, and no threshold exists that drives both to zero. And without a designed post-NG flow, the shop floor will build a “let someone look, just to be safe” routine and the labour saving disappears.

Method selection comes after the objective is fixed. If the defect appears consistently and can be defined numerically, rule-based; if it is irregular and hard to quantify, deep learning — but the latter needs training image data. Proceeding with only a handful of defect samples is not something we recommend. Installing a capture-only unit to build an image library, breaking inspection into items and starting with what rules can solve, or deferring for now, can all get you there faster in the end.

Treat automated dimensional inspection as a separate problem from visual inspection. Because the output is a measured value, it has to be folded into your metrology management system — calibration, reference standards, traceability and record retention included. Whether calibration can be completed within Thailand is an item to confirm at quotation stage.

On sequence, poka-yoke comes first. Poka-yoke devices exist to prevent defects being made; inspection equipment exists to prevent defects getting out. If your defects originate in human error, reducing occurrence through mistake-proofing gives you a clearer view of both cost and effect.

On cost, use the Japanese benchmarks — PoC JPY 3–5 million, full deployment JPY 10 million and above (sources: Nsight, koromo) — to understand the structure, then rebuild the figures on the assumption that Thailand changes them through imports, local installation, the maintenance structure, language support and jigs and transfer mechanisms. A payback calculation resting on labour cost savings alone is hard to sustain given that Thailand’s 2026 minimum wage runs from THB 337 to THB 400 per day by province (source: Thai Law Online). Breaking the benefit into items such as fewer escapes to customers, avoided sorting work and automated record-keeping, and building the numbers from your own history, is far more persuasive. Model BOI incentives both with and without, and confirm eligibility directly with the relevant authority or the BOI.

What makes the difference at a Thai plant, specifically: Thai-language operator screens and manuals, maintenance that can be completed locally, training for more than one person (against staff turnover), limit samples shared as images, and a decided method for presenting the basis of a judgement in a customer audit. Install equipment without those five in place, and the shop floor will revert to visual inspection the first time something goes wrong.

For the approach, we recommend the sequence Stage 0 (definition) → Stage 1 (sample evaluation) → Stage 2 (offline trial) → Stage 3 (one-line integration) → Stage 4 (horizontal rollout). Stages 0 and 1 expose nearly every gap in your internal criteria, and Stage 2 gives you a real feel for the false reject level. Line integration is hard to undo, so secure a window to trial it beforehand if you can.

It is entirely fine to talk to us before you have chosen an equipment manufacturer — or even while you are still deciding whether to automate at all. “Our inspection standard is still written in words and we don’t know how to quantify it.” “We have almost no defect samples, but we’re being pitched AI.” “Head office looks set to choose the model, and we’re not sure it fits local conditions.” “We have three quotations but the assumptions differ, so we can’t compare them.” If any of that sounds familiar, we can start together with identifying what needs deciding internally and organising the information to hand to bidders. A conversation before any implementation decision is perfectly welcome — feel free to reach us through the contact form.

References

• Nsight (commentary on AI visual inspection cost breakdowns)
• koromo (commentary on AI visual inspection in manufacturing)
• OptiMax (commentary on product inspection automation and AI)
• MarketsandMarkets, “Industrial Machine Vision Market”
• KAIZEN, “Thailand Manufacturing Trends 2026”
• Thai Law Online, “Minimum Wage in Thailand”
• Bangkok Shuho (Manufacturing Production Index, May 2026)
• Tilleke & Gibbins, “Thailand Unveils New Incentives for Automotive and HEV/PHEV Manufacturing”
• Alvarez & Marsal, “Thailand’s Renewed BOI Incentives”
• Thailand Board of Investment (BOI) Automation
• Nation Thailand (BOI investment promotion applications in 2025)