SME AI Adoption 2026 | Why 68.3% Admin and Only 34.9% Factory

On paper, SME AI adoption really is moving. Break the numbers down by department, though, and the picture changes: AI has reached 68.3% of general affairs and administration functions but only 34.9% of manufacturing and production. Meanwhile, the issue that owners name most often as a serious management problem is the labour shortage in direct departments, at 39.7%, against 15.3% for indirect departments. Where AI has landed and where it hurts are inverted. This article traces why that gap opens, using two primary surveys, and works through the cost and payback numbers a Japanese-owned manufacturing subsidiary in ASEAN needs before deciding what to fund next.

SME AI adoption is not “behind” — it has not reached where it hurts

Most articles about AI in smaller manufacturers open the same way: adoption is only around 20%, and four out of five companies are still not using it. They then close with the same prescription: start with generative AI, apply for a grant. Neither statement is wrong, but the framing drops the point that matters most. The problem is not the low adoption rate. It is the direction the adopted AI is pointing.

According to the Organization for Small & Medium Enterprises and Regional Innovation, Japan (SMRJ) and its “Fact-finding Survey on the Use of AI and Related Technologies by SMEs” (March 2026), AI adoption by business function runs at 68.3% for general affairs and administration, 60.3% for sales, distribution and service, and 58.5% for management and planning — against 34.9% for manufacturing and production. That is a gap of more than 33 points between the back office and the shop floor.

Over roughly the same window, Shoko Chukin Bank ran its “Survey on the Use of Generative AI by SMEs (January 2026 survey)”, which asked owners to rank their management priorities. The item most often flagged as a serious management issue was addressing labour shortage and improving efficiency in direct departments — manufacturing, sales and transport — at 39.7%. The same item for indirect departments came in at 15.3%, near the bottom of the eight-item list.

Line the two surveys up and the shape is unmistakable. The area that owners feel most acutely is the direct department, and that is where AI adoption is lowest. The area that hurts least is the indirect department, and that is where AI has landed most thickly. AI has gone where it is easy to install, not where it is effective.

The practical difference between those two readings is large. If SMEs are simply “behind”, time fixes it: awareness spreads, prices fall, case studies accumulate, and the laggards catch up on their own. If AI has “not reached” the shop floor, time fixes nothing. When the reason is structural, the structure has to change first — otherwise the manufacturing and production adoption rate will look much the same a decade from now.

Here is the conclusion up front. The reason AI does not reach the shop floor is neither a knowledge gap nor a budget gap. It is that there is no data in front of the AI. And a company without data cannot, even in principle, calculate the return on an AI investment. That is why the single largest barrier remains “no concrete use case in sight” at 35.6%. The rest of this article takes that structure apart and lays the available moves out in three layers, in order. The full roadmap for generative AI is covered separately in our generative AI implementation guide; this article narrows to one question — how to make AI reach the floor.

Where you stand, in numbers — two primary surveys, one conclusion

As a foundation for the argument, it helps to put two very different primary surveys side by side. Both were run between autumn 2025 and early 2026. Their populations and question designs differ, but they point the same way.

Overlay the two and a structure appears that neither shows on its own. Shoko Chukin Bank digs into how generative AI is actually used; SMRJ captures which type of AI has landed in which department. One gives you the use cases, the other the distribution.

The 20.4% AI adoption rate splits sharply by department

The SMRJ survey puts overall AI adoption at 20.4% — companies that have deployed AI across the business plus those using it in some operations. A further 18.6% are evaluating it, which gives 39.0% leaning towards adoption.

The Shoko Chukin Bank breakdown for generative AI is finer grained. Deployment of a company-specific model or AI agent stands at 0.9%, company-wide deployment of a packaged service at 4.1%, and deployment by a single department or a subset of members at 11.2% — 16.3% in total at company level, a figure that does not tie exactly to the components because of rounding. Against that, 14.9% report no company deployment but actively encourage individual use of generative AI, 64.9% report no company deployment and leave usage entirely to the individual, 1.4% prohibit or restrict it, and 2.5% fall into other categories. Company deployment plus encouraged individual use — the survey’s “active user” segment — comes to roughly 30%.

The two surveys cover different ground, so the figures are not directly comparable. What they share is the shape: fewer than two companies in ten have deployed AI at company level. The 64.9% who leave it to the individual is unique to the Shoko Chukin Bank question set, and it indicates how much usage is happening outside any company framework. So far, none of this is news. The next breakdown is where it gets interesting.

The top three cluster in a band between 58% and 68%. Manufacturing and production sits alone at 34.9%. This is not a story about manufacturers being slow adopters. It is a distribution inside the same companies: AI has reached the indirect departments and has not reached the direct ones.

82.6% of the AI in use is generative AI; shop-floor AI runs around one in ten

Why does manufacturing and production get left behind? Look at which types of AI have actually been deployed and the answer starts to surface.

These are SMRJ figures (n=322, multiple responses). Generative AI dominates at 82.6%. The types that create value on a factory floor — image recognition at 11.2%, demand forecasting at 8.1%, anomaly detection at 4.3% — all sit at or below roughly one in ten.

Stated objectives (n=331) put operational efficiency and shorter task times first at 87.0%, followed by quality improvement at 32.3%, addressing labour shortage at 31.7%, creating added value at 24.5%, reducing errors and mistakes at 24.2%, growing revenue and winning customers at 22.1%, and cost reduction at 14.8%. On realised effects, operational efficiency and shorter task times again leads at 83.2% — objective and outcome match neatly. On time savings, companies are getting exactly what they aimed for.

The Shoko Chukin Bank survey then tells you what that time saving consists of. Among active users, the leading application is drafting and summarising emails, reports and meeting minutes at 74.0%, followed by desk-work support and automation at 48.7%, strategy and planning support at 30.0%, design generation at 29.3%, research at 20.2%, application development at 16.9%, and training and internal education at 9.9%. Support and automation of shop-floor and face-to-face work comes in at 9.5%.

74.0% against 9.5%. That contrast is the same phenomenon as 68.3% against 34.9%, seen from the application side.

The place that hurts and the place AI has reached are inverted

Three numbers, side by side.

The higher the priority, the thinner the adoption. The lower the priority, the thicker. The correlation runs backwards from end to end.

That inversion also shows up in how results feel. In the Shoko Chukin Bank overall assessment, taken from active users only, 73.7% give generative AI some form of positive evaluation, yet only 31.7% say they feel a positive effect on the business as a whole, and 19.3% report no effect so far. More than seven in ten say it is useful; only three in ten say it moves the business. Given that generative AI is cutting time in the indirect departments, that is exactly what you would expect. The time it is cutting belongs to the 15.3% side of the priority list.

The same survey also shows that companies with a formal deployment report a positive business effect more than 10 points above those relying on individual sign-ups. Read alongside the 64.9% who leave it to the individual, the implication is that a large share of companies have not yet reached the starting line for usage that moves the business at all.

Why it does not reach the shop floor — data has to exist before AI does

This is the core of the argument. Why does only 9.5% of generative AI usage touch shop-floor work? The answer has nothing to do with technical difficulty or price.

Generative AI is the only AI that runs without any of your own data

Generative AI took an overwhelming 82.6% share not because it performs better than the alternatives, but because its preparation cost is zero.

Generative AI arrives pre-trained on general text and images. What the user supplies is a prompt and, at most, a handful of reference documents. It works on the day the account is created. No historical data to collect, no labels to attach, no model to train.

No other class of AI has that property. The flip side is that generative AI handles general language processing, not judgements specific to your company. Which is precisely why usage concentrates on emails, reports and minutes at 74.0%. If a tool can only be applied to work that needs none of your own data, that is the work it gets applied to.

The 82.6% figure is not a ranking of technical merit. It is the reflection of a single fact: exactly one type of AI runs with zero preparation.

Image recognition, anomaly detection and demand forecasting do not start without your own data

Every AI that creates value on the shop floor presupposes company-specific data.

The order of adoption — 82.6%, then 11.2%, then 8.1%, then 4.3% — is not a ranking by technical difficulty. It is a ranking by how hard the required data is to obtain. Anomaly detection sits last because acquiring a continuous signal stream off physical equipment takes more physical effort than any of the other three.

This is where most shop-floor AI projects stop. A plant explores visual inspection AI, talks to a vendor, and is told to prepare several hundred to several thousand images of good and defective parts under identical shooting conditions. As covered in our article on AI visual inspection in factories, that image-collection step takes weeks to months — and it rarely appears as a line item on the AI project quotation. The same thing happens with predictive maintenance. Fitting vibration sensors goes smoothly enough, but the model cannot classify anything until several months of normal-state baseline data have accumulated.

In short, shop-floor AI does not arrive because of an AI problem. It fails to arrive because of the state of the records sitting in front of it. A plant running on paper and spreadsheets has nothing to feed the model.

Lay the moves out in three layers, in order

Given all of the above, the available moves separate into three layers. The split is determined by two things: whether company-specific data is required, and what kind of return the layer produces.

The classic SME mistake is to complete Layer 1 and then try to jump to Layer 3. The jump stalls because Layer 2 is missing. And the awkward part is that Layer 2 “is not AI”, so a budget request filed under the heading of AI adoption tends not to clear. That is the structural trap.

Layer 1, indirect departments — immediate effect, pays back on labour savings alone

Layer 1 is the layer most companies are already running. Emails, reports and meeting minutes, translation, document summarising, first drafts of internal paperwork. The 74.0% and 48.7% figures from the Shoko Chukin Bank survey both live here.

At a Japanese-owned manufacturing subsidiary in Thailand or elsewhere in ASEAN, this layer bites harder than it does in Japan. Three languages circulate daily — Japanese, the local language and English — and translation and summarising for head-office reporting is a permanent fixture. Taking a report written in Thai or Vietnamese by local staff, reshaping it into Japanese and sending it to head office consumes several to more than ten hours a week at many sites.

Layer 1 does come with two traps that quietly erase the benefit. The Shoko Chukin Bank list of post-deployment issues maps onto them directly: internal rules and policies not keeping pace at 25.1%, effects not measurable so results stay invisible at 21.3%, knowledge concentrating in one department or a few staff at 14.7%, and failure to embed the tool into business processes at 12.0%.

The first trap is the absence of rules. Concerns about information management and security are cited by 21.0%, and when people start using personal accounts with no policy in place, drawings and cost data eventually get pasted into a prompt. Setting a generative AI usage policy at the outset is not about restriction — it is about stating clearly what may be entered, so that usage can grow.

The second trap is how the target work is chosen. Handing every employee a chat window splits the workforce into those who use it and those who do not. Making the effect reliable means naming specific routine tasks to be replaced. Choosing something like Excel work automation — the aggregation, transcription and reformatting that recur in a fixed shape every month — leaves the saved time in a measurable form.

Layer 2, turning shop-floor records into data — not AI, and the layer that is missing

Layer 2 is the work of recording what happens on the floor in a form that can be aggregated afterwards. Concretely, three things.

• Electronic forms — replacing daily work reports, inspection records, checklists and defect reports with tablet or PC entry
• Production data capture — acquiring output, run time, downtime and changeover time per process, timestamped, automatically or semi-automatically
• Signal acquisition from equipment — reading current, run signals, counters and temperature from machines or PLCs and accumulating them as time series

The word AI appears nowhere in that list. Which is why the internal positioning of this layer floats free. It does not clear under the banner of an AI project, and as a plain efficiency investment its payback is long, as the cost section below shows. The result is a proposal with no natural home in the approval process, deferred year after year.

Without Layer 2, though, Layer 3 cannot begin at all. With no images of good and defective parts, a visual inspection model has nothing to learn from; with no timestamped downtime records, there is no way even to define what an anomaly detection model should be evaluated against. As long as electronic forms and a paperless factory are framed as “getting rid of paper”, the investment gets postponed every year. The reason the framing has to change is easiest to show with the numbers, in the next section.

Layer 3, shop-floor AI — it cannot run without Layer 2

Layer 3 is the layer that only starts working on top of accumulated data. Automated visual inspection, equipment anomaly detection, demand forecasting: the 11.2%, 4.3% and 8.1% categories from the SMRJ survey.

The benefit from Layer 3 is different in kind from Layers 1 and 2. Layers 1 and 2 deliver labour savings — people spend less time — and the value scales with the hourly labour rate. The core benefit of Layer 3 is avoided loss. Stopping defects from escaping, reducing unplanned equipment stoppages, cutting both stockouts and excess inventory. In every case, the benefit takes the form of a loss that would have occurred and no longer does.

That difference changes how the payback is calculated. Labour savings come out of headcount multiplied by hours multiplied by rate. Avoided loss cannot be calculated at all without knowing how much that loss currently amounts to. And the mechanism for knowing that amount is Layer 2. The logic closes on itself here.

A company without Layer 2 cannot decide whether to do Layer 3. It might be worth doing; it might not. The inputs for the decision simply do not exist.

Costing each layer and its payback (Thailand site, 120 employees)

Now to test the argument with money. What follows is a model calculation for a single Japanese-owned manufacturing subsidiary in Thailand. The figures shift from site to site, so the sensible use of this section is to substitute your own measured values and rework the arithmetic. That is why every formula and assumption is written out alongside the result.

How to set the shared assumptions

When these are swapped for your own numbers, the hourly rate comes from the following formula.

Hourly labour rate = monthly labour cost / operating days per month / actual working hours per day

Whether social insurance, housing allowance and bonus accruals are folded into the monthly labour cost moves the rate up appreciably. This article uses a conservative base-salary figure. Including the on-costs pushes the benefit up, but as shown in the sensitivity section it pushes the cost side up at the same time, so net benefit grows more slowly than gross benefit does.

The direct rate of 50 baht/hour is anchored to Thailand’s minimum wage. Per JETRO Business Briefs (4 July 2025), the rate effective 1 July 2025 set Bangkok at 400 baht per day across all industries, up from 372 baht, while four provinces and one district including Chonburi and Rayong had already been at 400 baht since January 2025. The level was held into 2026, with a continuing range of 337 to 400 baht per day. Dividing 400 baht by 8 hours gives the 50 baht used here for direct labour.

Layer 1 pays back in about 7.5 months

The scenario: generative AI is rolled out to the 20 people in indirect departments, cutting 30 minutes per person per day from routine document work.

Dividing the 150,000 baht initial cost by the 239,100 baht annual net benefit gives 0.63 years — roughly 7.5 months to payback.

The inclusion of 18,900 baht of internal administration effort in the annual cost matters. Costing generative AI on seat licences alone drifts away from reality. Account administration, answering questions, sharing prompts and updating the policy all belong to somebody, and that person’s time is not free. Half an hour a day is on the conservative side.

Even so, Layer 1 pays back. The standard advice to start with generative AI is, in this specific sense, correct. There are not many moves that clear their initial cost inside a year on labour savings alone, with no data preparation required. The problem is what happens when a company stops there.

Layer 2 does not pay back on labour savings alone

The scenario: shop-floor forms are digitised and production data capture is introduced, including ten tablets, licences, configuration and floor-level training.

Dividing the 900,000 baht initial cost by the 97,200 baht annual net benefit puts payback at about 9.3 years. A capital request carrying a 9.3-year payback is unlikely to clear.

That outcome is not the product of a lazy estimate. The labour-saving benefit is structurally small. Cutting six minutes a day off the entry time of 100 direct workers produces only 126,000 baht, because the rate is 50 baht/hour. At a site with low labour costs, converting direct-department time savings into money shrinks the number the moment you do it.

Which is exactly why Layer 2 should not be evaluated as an efficiency investment. The real value of Layer 2 is not the time it saves but the data it generates. What that data is worth becomes visible in the Layer 3 calculation.

Layer 3 does not pay back on its own either — payback comes from avoided loss

The scenario: automated visual inspection, on the assumption that the Layer 2 data already exists.

Dividing the 1,500,000 baht initial cost by the 102,400 baht annual net benefit puts payback at about 14.6 years. Combining Layers 2 and 3 into a single programme does not improve matters.

Twelve years. The most common misreading of that number is to conclude that AI is premature for smaller manufacturers. The correct reading is different. Counting only labour savings as the benefit, shop-floor AI does not pay back. What makes shop-floor AI pay back is not reduced inspection effort but avoided loss — defect escapes, downtime, late deliveries.

And that amount can only come from your own actuals. At this point a back-calculation is useful.

Required annual avoided loss = initial investment / target payback years + annual running cost – annual benefit from labour savings

Feeding in the model figures (initial 2,400,000 baht, annual running 380,000 baht, labour savings 579,600 baht) gives the following.

So if the target is a five-year payback, the decision criterion becomes whether reduced defect escapes and downtime can avoid the equivalent of 280,400 baht a year. Expressed monthly, that is around 23,400 baht — roughly 58 person-days at the 400 baht daily rate. Whether that is a lot or a little cannot be said in the abstract. Without knowing how much loss your own site incurs each year, the question is unanswerable.

One caution about that table: it is not claiming that avoided losses of this size exist. It states a necessary condition and nothing more. The actual loss figure can only be computed from your own defect rate, shipment-hold frequency and minor-stoppage hours. Our article on measuring AI ROI covers the measurement approach in more depth, but whichever framework is used, the formula stays empty without actuals to put into it.

Sensitivity analysis — what moves when the hourly rate moves

It is worth checking how much the result depends on the assumptions. Taking Layer 1 and varying the indirect hourly labour rate, the important detail is that the cost side moves at the same time as the benefit side. The 126 hours of internal administration effort inside the annual cost are valued at that same hourly rate.

The cost formula is seat licences of 120,000 baht plus 126 hours times the hourly rate. Estimates that move only the benefit side, and conclude that a higher rate means a bigger effect, are simply ignoring the administration effort.

Rearranging the table into a formula for net benefit gives the following.

Annual net benefit = 2,520 hours x hourly rate – 120,000 – 126 hours x hourly rate = 2,394 x hourly rate – 120,000

That expression crosses zero at an hourly rate of about 50 baht/hour. In other words, Layer 1 turns a profit on labour savings alone as long as the target population’s hourly rate is somewhere above 50 baht/hour. At the indirect rate of 150 baht/hour there is comfortable headroom; at the direct rate of 50 baht/hour the calculation sits exactly on the break-even line. Extending the same generative AI accounts to direct workers unchanged would therefore not pay back on labour savings. The reason Layer 1 belongs in the indirect departments shows up here as a number rather than an opinion.

The same operation applied to Layer 2 behaves differently. The 180,000 baht annual running cost is licences and maintenance, so it does not move with the hourly rate; only the benefit side does. At an indirect rate of 200 baht/hour the annual net benefit becomes 147,600 baht and payback about 6.1 years; at 100 baht/hour it becomes 46,800 baht and about 19.2 years. However the rate is set, labour savings alone never produce a payback period that survives an investment review. Changing the assumptions does not change the conclusion — which is what this exercise establishes.

“No concrete use case in sight”, 35.6%, really means nobody is measuring

In the Shoko Chukin Bank survey, the leading obstacle in the pre-adoption phase was “no concrete use case in sight” at 35.6%. Across all phases, it is the single largest barrier. In the evaluation phase it is followed by “no one to drive adoption” at 32.0% and “employee knowledge gaps and psychological resistance” at 29.2%; in the post-adoption phase the leading item becomes “internal rules and policies not keeping pace” at 25.1%.

That 35.6% is usually interpreted as an information deficit. The SMRJ survey supports the reading at first glance: 83.3% report inadequate access to success stories and usage examples, 79.8% report inadequate information for selecting an appropriate vendor or product, and the public support most requested includes provision of case-study information at 70.5%, second only to subsidies for adoption costs at 77.9%. So policy naturally heads towards publishing more case studies.

But reading a hundred case studies does not tell you where AI would work in your own plant. The reason is simple. Whether it works depends on how much money you are currently losing, and where.

The right-hand column is what the 35.6% actually consists of. The reason use cases are invisible is not a failure of the owner’s imagination. It is that nobody has measured where and how much the company is losing. As long as the loss figure is blank, no case study can produce anything beyond “that looks like it might help us”, which is not a form that survives an investment review.

The 21.3% who report post-adoption that effects cannot be measured and results stay invisible come from the same root. A company that did not measure before deployment cannot measure after it, because there is no baseline.

From here the positioning of Layer 2 changes. Layer 2 is neither “the stage before AI” nor “an efficiency project to get rid of paper”. It is the investment that makes a Layer 3 investment decision possible.

That reframing changes how the request is actually written.

• The conventional version — “Digitise shop-floor forms and cut 151,200 baht a year from daily-report transcription effort.” (Rejected on a 9.3-year payback.)
• The reframed version — “Build the ability to report defects, stoppages and waiting time in monetary terms every month, and use those figures to decide whether to fund shop-floor AI next year at an initial scale of 1,500,000 baht.”

The second version explains 900,000 baht not as an efficiency measure but as the cost of improving the accuracy of a 1,500,000 baht investment decision. Because it is now compared against the cost of getting that decision wrong, the basis of evaluation shifts. And as a by-product, the 151,200 baht of transcription savings and the 126,000 baht of shorter entry time still arrive. Reversing the order of the argument is enough to make the same investment clear.

Four conditions change at an ASEAN site

Everything so far rests on Japanese primary surveys. Applying it to a Japanese-owned manufacturing subsidiary in Thailand or elsewhere in ASEAN, four conditions change.

Seven in ten Thai SMEs use AI, but fewer than two in ten run it company-wide

According to the UOB Business Outlook Study 2026 (first-half edition, covering 265 business owners and decision makers in Thailand), more than seven in ten Thai SMEs have implemented AI — above the ASEAN regional average. Adoption of digital solutions exceeds 80%. Among companies that have adopted AI, 58% report reductions in operating costs and 44% report immediate productivity gains.

Taken alone, that reads as ahead of Japan. The same study has a second half. Company-wide AI adoption plateaus at 18% to 24%, with 73% still in the preparation stage of digital transformation.

That 18% to 24% band overlaps almost exactly with the Japanese figures: 16.3% company-level generative AI deployment and a 20.4% overall AI adoption rate. In other words, many companies have AI in use by individuals or single departments, but only around two in ten have embedded it into company-wide business processes — and that structure is the same in Japan and in Thailand. Neither “Thailand is behind” nor “Thailand is ahead” captures the reality. The bottleneck is in the same place.

For context, Thailand has around 3.3 million SMEs, representing 99.5% of all enterprises and employing 13.6 million people. At the same time the country is estimated to be short of roughly 80,000 digital workers, and more than 80% of SMEs are considering overseas expansion within the next two to three years, with Singapore, Vietnam and Malaysia at the top of the destination list. Talent stays thin while business complexity keeps rising.

The language of the shop floor, and the absence of any real IT function

The first thing a Layer 2 design runs into at an ASEAN site is language. Operators work in the local language, the management layer is a mix of local staff and Japanese expatriates, and head-office reporting is in Japanese. That structure cuts two ways.

It works in Layer 1’s favour. Translation, summarising and minute-taking — generative AI’s strongest suit — are the daily routine. The 378,000 baht annual benefit calculated above is, at a genuinely multilingual site, arguably a conservative estimate.

It works against Layer 2. Unless the entry screens and pick-lists in the electronic forms are complete in the operators’ own language, the floor will not fill them in. A screen with Japanese field labels and a local-language footnote degrades within weeks of go-live into everyone selecting whichever option sits at the top of the list. Defect classification codes have to be designed at a granularity and in a vocabulary that a local inspector can choose from without hesitating. Cut corners here and the data that accumulates is unusable in Layer 3.

The delivery structure differs too. The Shoko Chukin Bank survey records “no one to drive adoption” at 32.0%, “employee knowledge gaps and psychological resistance” at 29.2%, and “no support partner or adviser available” at 11.5% — and at an overseas site all three are amplified. There is often no dedicated information systems role at all, with general affairs or accounting staff covering it as a side duty. On top of that, expatriate postings typically run three to five years, so there is a structural risk that the project stops the moment the sponsor rotates out. Where to keep work in-house and where to hand it out is covered in our article on AI insourcing and external support; at an ASEAN site, whether local-language work instructions and a documented handover to local members are included in the deliverables makes more difference to the outcome than it would on a domestic project.

Wage levels change the shape of the payback calculation

The third condition is labour cost. Bangkok’s minimum wage is 400 baht per day, which puts the direct hourly rate at the 50 baht/hour level used here. That is a very different level from the equivalent work in Japan.

The gap changes the shape of the investment case. A proposal justified by direct-department labour savings rarely stands up in Thailand. The Layer 3 calculation, where reduced inspection effort came to only 302,400 baht and payback stretched to 14.6 years, is exactly this effect. The same equipment installed in a high-wage country could pay back in a few years on labour savings alone.

Avoided loss, by contrast, does not track local wages. The cost of letting a defect escape is set by material cost, freight, customer handling and, in the worst case, a stopped production line. At an export-oriented site the customers are in Japan, Europe or North America, and the handling costs arise at those countries’ cost levels.

The conclusion is that at an ASEAN site the centre of gravity of the payback shifts from labour savings to avoided loss, more strongly than it does in Japan. Basing the case on avoided loss requires a monetary figure, and producing that figure requires records. The lower the labour cost at a site, the relatively more important Layer 2 becomes.

Thai support schemes, and what a Japanese-owned subsidiary needs to verify

The fourth condition is public programmes. Thailand has its own frameworks for pushing SME digitalisation forward.

SMEs Growth 2026, led by ETDA (the Electronic Transactions Development Agency), includes depa, OSMEP, SME D Bank and TCG as participants. For 2026 it expands to 16 provinces across 4 regions, offering roadshows, workshops, consultation clinics, technology showcases and business matching. For a site stuck at “we do not know which vendor to ask for what”, the technology showcase and business matching elements are a practical entry point.

On funding, OSMEP has an SME loan scheme of 1.2 billion baht at 1% annual interest with a maximum term of 5 years. Transformation SME, aimed at modernisation, digital technology and innovation, goes up to 10 million baht; Enhancement SME offers 10 million baht for small enterprises and 15 million baht for medium ones. Applications opened on 5 May 2026. As the name suggests, the intended use of Transformation SME overlaps with Layers 2 and 3 of this article.

There is, however, an important caveat. Eligibility conditions such as nationality and shareholding ratio are not stated in the published announcements, so whether a wholly Japanese-owned local entity qualifies cannot be asserted either way. Before a scheme is built into an investment plan, it is worth confirming directly with OSMEP. Thai SME support programmes are a mix of those that carry a Thai-ownership requirement and those that do not, and writing one into a plan without checking the conditions tends to mean rebuilding the funding plan later.

On regulation, one piece of misinformation is worth correcting. Thailand’s AI law has not been enacted as of the time of writing. It remains at draft stage and is not in force. The AI law that took effect on 1 March 2026 is Vietnam’s, not Thailand’s. Articles conflating the two circulate in both Japanese and English, so compliance policies are best built from primary sources. The broader picture of AI adoption in Thailand is collected in our Thailand AI implementation article.

A 90-day route to being able to make the next investment decision

Finally, the order of execution. The goal is not to adopt AI. It is to be in a position, 90 days from now, to argue about Layer 3 with numbers.

Two things have been deliberately left out of this roadmap.

The first is a Layer 3 proof of concept. Watching a vendor demo with no data of your own adds nothing to the decision. What it adds is an impression that the technology looks promising, and that is not a solution to the 35.6%.

The second is any requirement for company-wide rollout. The loss measurement starting on day 46 can cover the single chosen process and nothing else. An attempt to measure every process burns 30 days on designing forms and produces nothing by day 90. Once one process yields a monetary figure, that figure becomes the basis for extending to the others.

What is in hand on day 90 is a measured net benefit for Layer 1 and a loss figure for one process. With those two, the management meeting’s question stops being “should we adopt AI” and becomes “given this loss figure, how much are we willing to invest against it”. Changing the shape of the discussion is the output of the 90 days.

Frequently asked questions about SME AI adoption

Where should a smaller manufacturer start with AI?

Layer 1 — generative AI in the indirect departments — is the right starting point. In the model calculation, an initial 150,000 baht pays back in about 7.5 months and no data preparation is required. In parallel, though, it is worth starting to measure for Layer 2. Continuing with Layer 1 alone never reaches the direct-department labour shortage that 39.7% of owners name as a serious management issue. Concretely, that means picking one process that hurts and beginning to record defects, stoppages and waiting time — on paper if necessary. It costs almost nothing and it becomes the input to the investment decision three months later.

How much does AI adoption cost?

The layers differ by an order of magnitude. In the model calculation for a Thailand site with 120 employees, Layer 1 was 150,000 baht initial and 138,900 baht annual, Layer 2 was 900,000 baht initial and 180,000 baht annual, and Layer 3 was 1,500,000 baht initial and 200,000 baht annual. Market prices vary widely with site size and target process, so the formulas travel better than the amounts. Annual benefit is time saved x headcount x operating days x hourly rate. Required avoided loss is initial investment / target payback years + annual running cost – annual benefit from labour savings. The hourly rate is monthly labour cost / operating days per month / actual working hours per day. Putting your own figures into those three formulas will show which layer you can afford to fund.

Can AI be adopted without a dedicated IT person?

For Layer 1, yes. It takes roughly 30 minutes a day of administration effort, but not a dedicated role. From Layer 2 onwards the situation changes. “No one to drive adoption” is the leading evaluation-phase obstacle in the Shoko Chukin Bank survey at 32.0%, and at an ASEAN site the information systems role is frequently a side duty of general affairs or accounting, which makes it harder still. The workable pattern is to keep requirements definition and acceptance decisions in-house while handing part of the build and operation to an external partner. In that arrangement, unless local-language work instructions and a handover to local members are included in the deliverables, operations stop when the expatriate sponsor rotates out.

We deployed generative AI but cannot feel any effect. Why?

Most companies are in the same position. In the Shoko Chukin Bank survey, even among active users, 73.7% give some form of positive evaluation while only 31.7% report a positive effect on the business as a whole, and 19.3% answer that there has been no effect so far. There are two causes. The first is that 74.0% of usage is drafting emails, reports and minutes, and the time that saves belongs to the 15.3% side of the management-priority list. The second is measurement: 21.3% cite the inability to measure effects and the invisibility of results as a post-deployment issue. The remedy for the first is to move up the three-layer model; the remedy for the second is to capture a baseline of task durations before deployment. Note also that companies with a formal deployment report a positive business effect more than 10 points above those leaving it to individuals, and that rule-setting matters — the leading post-deployment issue, at 25.1%, is internal rules and policies not keeping pace.

Does the same approach work at an overseas subsidiary as in Japan?

The order of the three layers and the logic of the investment decision are the same. Four things change. First, the hourly labour rate, which makes proposals justified by direct-department labour savings hard to sustain. Second, language: the electronic form screens and pick-lists have to be complete in the operators’ own language. Third, the delivery structure, which has to assume no dedicated IT role and a risk that the project breaks when an expatriate posting ends. Fourth, public programmes: frameworks such as ETDA’s SMEs Growth 2026 and the OSMEP loan scheme exist, but whether a wholly Japanese-owned local entity qualifies has to be confirmed scheme by scheme. Note too that while the UOB study reports more than seven in ten Thai SMEs implementing AI, company-wide adoption plateaus at 18% to 24%. The gap between individual use and company-wide operation has the same structure as in Japan.

Conclusion

The main points of this article, in order.

• SME AI adoption is not “behind”, it has “not reached”. AI sits at 68.3% in general affairs and administration against 34.9% in manufacturing and production, while the share naming the issue serious is 39.7% for direct departments and 15.3% for indirect ones. Where AI has landed and where it hurts are inverted
• 82.6% of the AI deployed is generative AI. Image recognition at 11.2%, demand forecasting at 8.1% and anomaly detection at 4.3% all sit around one in ten. 74.0% of generative AI usage is drafting emails, reports and minutes, while shop-floor and face-to-face support is 9.5%
• The cause is not a knowledge gap. Generative AI is the only AI that runs without any of your own data. Image recognition needs images, anomaly detection needs equipment data, demand forecasting needs historical actuals
• The moves split into three layers: Layer 1 (generative AI in indirect departments), Layer 2 (turning shop-floor records into data), Layer 3 (shop-floor AI). The typical failure is jumping from Layer 1 to Layer 3 with Layer 2 missing
• In the model calculation, Layer 1 recovers its 150,000 baht initial cost in about 7.5 months. Layer 2 takes about 9.3 years, Layer 3 about 14.6 years, and the two combined about 12 years. Neither Layer 2 nor Layer 3 pays back on labour savings alone
• What makes shop-floor AI pay back is avoided loss, and that amount can only come from your own actuals. The real content of “no concrete use case in sight” at 35.6% is not having measured your own losses
• Layer 2 therefore has to be repositioned from “an efficiency investment” to “the investment that makes a Layer 3 investment decision possible”
• At an ASEAN site, four conditions change: hourly labour rate, language, delivery structure and public programmes. The lower the labour cost, the smaller the labour-saving benefit and the more the payback rests on avoided loss. Note also that Thailand’s AI law is not yet enacted as of the time of writing

What decides whether AI adoption succeeds is neither model performance nor vendor selection. It is whether you can state your own losses as a monetary figure. Once you can, whether to adopt AI answers itself.

TOMAS TECH builds the Layer 2 foundations — production data capture, electronic forms and signal acquisition from equipment — for Japanese-owned plants in Thailand and across ASEAN, and supports the move up to visual inspection and anomaly detection on top of them. Rather than opening with a Layer 3 quotation, we are happy to talk through which layer your site is currently on and which one to start from, even just to work out that split. If you tell us how records are being kept today and which process is hurting, we can help you assemble a 90-day plan for producing the decision inputs — please get in touch via our contact page.

References

• Shoko Chukin Bank, “Survey on the Use of Generative AI by SMEs (January 2026 survey)”, published 31 March 2026: https://www.shokochukin.co.jp/report/data/assets/pdf/futai202603.pdf
• Organization for Small & Medium Enterprises and Regional Innovation, Japan (SMRJ), “Fact-finding Survey on the Use of AI and Related Technologies by SMEs”, March 2026: https://www.smrj.go.jp/research_case/questionnaire/fbrion0000002pjw-att/202603_AI_point.pdf
• UOB Business Outlook Study 2026 (AI adoption among Thai SMEs, reported July 2026): https://www.thaipr.net/en/business_en/3735803
• “The Resilient Innovators: How Thai SMEs are Navigating Economic Shocks via AI and Regional Expansion”, The Nation Thailand, 3 July 2026 (detail on the UOB study): https://www.nationthailand.com/business/economy/40068200
• “Thailand Expands SMEs Growth 2026 Programme to Accelerate Digital Adoption”, OpenGov Asia, May 2026: https://opengovasia.com/thailand-expands-smes-growth-2026-programme-to-accelerate-digital-adoption/
• “OSMEP launches THB1.2 billion Thai SME loan scheme at 1% interest”, The Nation, May 2026: https://www.nationthailand.com/business/economy/40065764
• JETRO Business Briefs, “Bangkok minimum wage raised to 400 baht per day (Thailand)”, 4 July 2025: https://www.jetro.go.jp/biznews/2025/07/b21007a1ac8f7fca.html