Generative AI Implementation for ASEAN Manufacturing Sites: Cost, Roadmap and Governance 2026

Across manufacturing sites in Thailand and Vietnam, the same conversation keeps repeating in one form or another: the group has started using generative AI somewhere — at the parent company, at the regional headquarters, in a corporate function — but nobody at the plant knows what the local entity is actually supposed to do about it. Tool selection is the visible question. It is almost never the real one. Before you choose a platform, you need to establish where you currently stand, in what order to move, how to count cost and return, how to handle security and local regulation, and how to design the programme so that it does not die as another proof of concept.

This article approaches generative AI implementation from the perspective of a plant or a regional entity in ASEAN rather than from the perspective of a corporate centre. It uses published survey data from Japan, Thailand, Vietnam and a global enterprise study, each cited with its source, sample size and survey period, so that you can judge for yourself how far each data point travels. Where a number is specific to one country, it is labelled as such.

Where Generative AI Implementation Actually Stands

Before debating anything, it is worth fixing an objective sense of position. When a leadership team argues from feel — “we are behind,” “we are ahead of our peers” — the investment decision ends up being driven by whoever speaks with the most conviction rather than by evidence.

For groups with a Japanese parent company, the most recent large-sample reference point is the Teikoku Databank survey on corporate trends in generative AI, conducted from 17 to 31 March 2026. It covered 23,349 companies and received 10,312 valid responses, a 44.2% response rate. This is a Japan-specific dataset: every respondent is a company operating in Japan, and it should be read as an indicator of what a Japanese parent company or group IT function is likely to have already done, not as a proxy for ASEAN.

In that survey, 34.5% of companies said they use generative AI in their business, broken down into 4.4% “using it extensively” and 30.2% “using it somewhat.” Roughly one company in three has brought generative AI into operations in some form, while fewer than one in twenty rates itself as a deep user. Adoption rises with size: 46.5% among large enterprises, 32.4% among small and medium enterprises, and 28.0% among micro enterprises. By headcount the gap widens further, from 63.6% among companies with more than 1,000 employees and 51.9% in the 301–1,000 band, down to 29.6% among companies with five employees or fewer. By industry, services lead at 47.8%, followed by finance at 38.6% and real estate at 34.9%, while transport and warehousing (27.5%) and construction (26.4%) sit at the bottom.

The Uncomfortable Concentration in Text Work

For anyone designing an implementation at a production site, the adoption rate is not the interesting figure. The breakdown of use cases is. In the same Japan-specific survey, 45.1% of companies cited “writing, summarising and proofreading text” as their use case — far ahead of information gathering at 21.8%, planning and ideation at 11.0%, data aggregation and analysis at 7.4%, and programming assistance at 5.9%.

Read that again with a factory in mind. Data aggregation and analysis — the category that would include production output, equipment logs, quality inspection records and inventory — does not even reach one in ten. The primary data a plant generates every single day is, for the overwhelming majority of companies, still untouched by generative AI. That gap is the thread running through the rest of this article. Making documents faster to write is genuinely useful, but it does not move a manufacturing site’s profit and loss statement.

The same survey found that 86.7% of users reported a positive effect (25.2% “significant,” 61.5% “some”). Plenty of companies feel the benefit. But if that benefit consists mainly of saving time on writing, the honest management question is not whether generative AI works — it is what comes after the writing.

Local Companies in Thailand Are Already Moving

If you look only at the parent company’s numbers, a comfortable sequencing suggests itself: let corporate adopt first, then cascade to the regions. Put Thai data next to it and that assumption starts to look shaky.

The Thailand-specific picture begins with Thailand Digital Outlook 2026, a national survey of 834 business operators. It reported that Thai companies reached an intermediate level of digital maturity for the first time, with an average score of 2.12 out of 4, up sharply from 1.56 in 2025. By size, large enterprises scored 2.58, mid-sized enterprises 2.41 (up from 1.97), and small enterprises 2.01 (up from 1.45) — the first time small operators crossed the 2.0 line. Among advanced technologies, AI had the highest adoption rate at 21.2%, ahead of blockchain at 20.1%, big data analytics at 19.9%, IoT at 19% and robotics at 17.8%. The survey flagged development of digital products and services, and R&D, as the country’s weak points.

An earlier reference point is the ETDA and NSTDA study “Readiness for the Application of AI Technology for Digital Services in 2024,” which approached 3,758 organisations between July and September 2024 and received 580 responses. It found 17.8% of organisations had adopted AI, up from 15% the previous year, with 73% considering it. The barrier most frequently cited was data quality. The three areas where generative AI implementation was furthest along were product and service development and R&D, marketing and sales and customer engagement, and production processes.

More recently, the UOB Business Outlook Study 2026 surveyed 265 business owners and decision-makers in Thailand (the study’s first Thailand edition, conducted in H1 2026). It reported that more than 70% of Thai SMEs have implemented AI, above the regional average, and that more than 80% have adopted some form of digital initiative. Among companies that had adopted AI, 58% reported cost reduction and 44% reported productivity gains. The same study found more than 80% considering overseas expansion within two to three years and 53% looking to expand manufacturing within ASEAN.

How to Read These Numbers Without Misusing Them

One caution deserves emphasis. You cannot line the Japanese figures up against the Thai figures and declare a winner. The Teikoku Databank survey draws on more than 10,000 responses from companies in Japan; Thailand Digital Outlook 2026 covers 834 operators; the ETDA/NSTDA study reflects 580 responding organisations; the UOB study surveyed 265 owners and decision-makers. The populations, respondent profiles, question definitions and survey periods all differ. A headline like “70% in Thailand versus 34.5% in Japan” is a misreading that ignores methodology.

What you can legitimately take away is directional. Thai companies are converting AI from “under consideration” to “in production,” and the motivation is unambiguously practical: cost and productivity. While a regional entity waits for a group-level rollout, the local suppliers and competitors in the same industrial estate may well be several steps ahead.

The table is not a criticism of either side. It simply shows that when the corporate centre and a production site use the phrase “generative AI implementation,” they are describing two different problems with different constraints. A programme that succeeded at headquarters does not stall at a plant because the plant is less capable; it stalls because the preconditions are different.

How to Implement Generative AI: A Five-Phase Design

Once you know where you stand, sequencing is the next decision — and getting it wrong is what turns budgets of any size into abandoned pilots. For a programme anchored at an ASEAN site rather than at the corporate centre, five phases work well in practice.

Phase 0 — Inventory the work and the data

The first task is not tool selection. It is an inventory of processes and data. Which activities at the site consume how much effort, and where does the information those activities depend on actually live, in what format? Forms that circulate on paper. Ledgers maintained in spreadsheets. Tables inside the production management system. Logs emitted by equipment. Judgement criteria that exist only in one supervisor’s head. Simply listing these separates problems that generative AI could plausibly address from problems that cannot be addressed at all because the underlying data does not exist in usable form.

This is also the phase to fix your evaluation metrics. “It feels more convenient” will not secure a second-year budget. Tie the target to metrics you already track: processing time for the target task, error rate, lead time, overtime hours.

Phase 1 — Distribute a common tool and establish rules

Next, give employees at the site a safe environment to work in. The point here is not feature richness. It is that everyone uses the same environment under the same rules. Unmanaged use of personal accounts is a risk exposure before it is a productivity question.

The Japan-specific March 2026 Teikoku Databank survey found that 25.5% of companies cited “rules have not been established” as a challenge, and that the most frequently reported negative effect was “a widening gap between people who can use it well and people who cannot,” at 18.8% (67.7% reported no negative effects, and 0.7% reported information leakage). That gap is usually not a tool problem — it is the absence of shared patterns of use. At an ASEAN site, publishing usage guidelines and a library of standard prompts in both the local language and the corporate language changes the speed of adoption on its own.

Phase 2 — Connect your own documents and records

Distributing a general-purpose tool caps the benefit at faster writing. The substance begins here: bringing internal policies, work standards, equipment manuals, historical troubleshooting records and quality criteria into scope for retrieval, and requiring the system to answer with citations back to the source document.

This is the point at which a new employee gains access to the same basis for judgement as a twenty-year veteran. Manufacturing sites in ASEAN often run on short assignment cycles for expatriate managers, which makes knowledge handover a structural rather than an occasional problem. Improving the searchability of documentation alone has an outsized effect on that.

Phase 3 — Embed AI into the process and move toward agents

Once document retrieval is established, the next step is embedding AI in the process itself. Instead of a person going to ask the model a question, the model sits inside the workflow and carries multi-step tasks forward automatically.

For enterprise reality on this, Anthropic’s study “How enterprises are building AI agents in 2026,” conducted with Material and published on 9 December 2025 among more than 500 technical leaders, is a useful global reference. It found 57% running AI agents in multi-step workflows and 16% running them in cross-functional processes, with 81% planning to take on more complex use cases in 2026. Software development dominates the use cases: roughly 90% use AI for development assistance and 86% run agents against production code. Data analysis and report generation stood at 60%, internal process automation at 48%, and 56% expected to implement research and reporting agents within the year. Our own take on what this means for a plant floor is set out in AI agents and manufacturing DX.

Phase 4 — Replicate across sites and build internal capability

The final phase is horizontal replication — to adjacent processes and to other country sites — and bringing operations in-house. When one process shows a verified result, the same pattern moves next door. What matters is that it was designed to be movable in the first place. If the working solution depends on prompts and procedures hand-built by one individual, replication is where it stops.

Durations are indicative only and shift with process complexity, the state of existing systems and local staffing. The one thing to build into the plan from day one is that the data preparation required from Phase 2 onward takes real time.

Generative AI Implementation Cost and ROI

Cost is where most evaluations stumble. Teams write a funding request without a usable sense of market pricing, and the compromise — “let’s start small” — often means starting below the scale at which any effect could be observed.

Published ranges from vendors in the Japanese domestic contract development market put a proof of concept at JPY 1.5–5 million and a full production implementation at JPY 15 million and above. A second source presents a different build-up: JPY 1–5 million for a PoC, and production implementation at JPY 800,000–2.5 million per month multiplied by the number of active months.

Treat these numbers carefully. They are benchmarks for contract development procured in Japan. They are not procurement prices in Thailand or Vietnam. Published statistics on per-man-month rates for system development in Thailand are thin, and this article will not assert a figure it cannot source. In practice, importing Japanese pricing wholesale produces quotes that do not match the local market, while assuming “it must be cheaper here” tends to omit the cost of multilingual support and coordination with the parent company, and the project runs out of budget mid-flight. The workable approach is to obtain quotations from several providers and compare them at the same level of granularity across five line items: initial build, licences, usage-based charges, maintenance, and training. For how to evaluate a development partner in this market, we have set out the criteria in how to choose a system development company in Thailand.

Define the numerator and the denominator before you spend

The most common ROI failure is deferring the definition of the numerator — the benefit. Asking people after go-live “how much did this help?” without a baseline collects impressions, not evidence.

Some reference points. In the Anthropic study cited above, 80% of responding organisations said they were already seeing measurable economic returns from their investment in AI agents. Named examples in that study include eSentire reducing threat analysis time from five hours to seven minutes, and Doctolib shipping features 40% faster using Claude Code. On the Thailand side, the UOB Business Outlook Study 2026 reported that 58% of AI adopters realised cost reductions and 44% realised productivity gains.

These are other companies’ results and guarantee nothing about yours. What they share is the unit of measurement: time. For a factory, that means measuring specific process times before you start — changeover duration, time from anomaly detection to first response, time to produce the monthly report, time to reconcile import/export documentation. Whether you took that one extra step at the beginning determines, more than anything else, how easy the second-year budget is to defend.

On the denominator, resist looking only at licence fees. In practice the majority of the cost lands on internal effort to prepare data, integration work with existing systems, and the training and hands-on support needed to make the change stick on the floor. Buy licences without budgeting for those and you will pay a monthly fee for accounts nobody opens.

Why Generative AI PoCs Fail

“The PoC went well but it never went to production” is a story heard as often at ASEAN sites as anywhere else. The causes are structural, not motivational.

The most instructive evidence is the ranking of obstacles in the Anthropic enterprise study: integration with existing systems first at 46%, data access and data quality second at 42%, and change management third at 39%. The wall enterprises hit is not model intelligence. It is that the model does not connect to the systems where work happens, that the data behind those systems is not in a usable state, and that people and organisations do not change.

The Japan-specific survey points the same direction. Teikoku Databank’s March 2026 challenges ranking was: accuracy of information 50.4%, shortage of specialist talent 41.3%, unclear scope of application 40.0%, information leakage risk 33.5%, and inadequate rules 25.5%. That accuracy tops the list at more than half is arguably the mirror image of asking a general-purpose model questions without ever giving it your own primary data. On the Thai side, the ETDA/NSTDA study named data quality as the barrier to adoption. Different countries, different methodologies, same finding.

Five recurring failure patterns

The first is choosing the wrong first target. Picking root-cause analysis of quality defects because the prize looks big means spending six months on data preparation before producing anything. The first target should be a process whose data is already digital, whose effect is easy to measure, and whose failure does not halt operations.

The second is the absence of acceptance criteria. Start a PoC without defining what success looks like and it ends in “seems promising,” which is not a basis for a production investment. Set the line in advance: “reduce a task that currently takes 30 minutes to under 10,” or “achieve 85% accuracy or better.”

The third is staffing. It is entirely normal for a site to have one IT person who is already carrying day-to-day operations and support, and who runs the PoC as an additional duty. Under that structure, the moment the PoC ends, nobody owns it. Bring whoever will operate the solution into the project during the PoC, not after.

The fourth is the relationship with the corporate centre. A site selects a solution independently, then finds it fails the group’s security standard and is sent back — or discovers it duplicates a platform the group already licenses. Aligning scope with group IT before the funding request removes this failure mode entirely.

The fifth is consent on the shop floor. Change management ranked third at 39% for a reason: it is universal. For an operator, a new system raises two questions — am I being monitored, and is my job being reduced? Explaining the purpose in the local language, and starting with use cases that visibly make the operator’s day easier, looks like a detour and is in fact the shortest path.

Security and Data Leakage Controls in Practice

Every leadership discussion about overseas deployment reaches data leakage. In the Japan-specific March 2026 Teikoku Databank survey, information leakage risk ranked fourth among challenges at 33.5%. Yet only 0.7% of companies reported information leakage as an actual negative effect they had experienced, and 67.7% reported no negative effects at all. High concern, limited reported harm — that is the honest current picture, and it is a picture from Japan rather than a global measurement.

Limited reported harm does not mean controls are optional. At regional sites in particular, a familiar pattern emerges: the group has written a policy, the policy has never been translated, and in practice everyone quietly uses personal accounts. That is the most dangerous configuration available — formally prohibited, universally unenforceable.

It helps to think about controls in five layers. The first is the contract and tenancy layer: choose an enterprise agreement and confirm the configuration under which your inputs are not used for model training. The second is access rights. Whoever can see a document today will be able to retrieve it through AI tomorrow, because your existing permission design carries straight through. Point a retrieval system at internal documents while permissions are loose and information that was previously invisible becomes available to anyone who asks a question. Auditing permissions before deployment is non-negotiable.

The third layer is data classification: drawings, costing, HR records, customer information — decide what may and may not be sent to an external service, calibrated to local reality rather than to a template. The fourth is logging and audit: record who submitted what, and make it reviewable on a schedule. The fifth is education. Circulating a list of prohibitions does not change behaviour. Staff need to hear, in their own language and with concrete examples, *why* a particular input creates a problem.

One further note for manufacturers: as information networks and control networks converge, extracting equipment data for AI use has to be preceded by securing the OT side. That is a separate discipline with its own risk model, and it should be reviewed in parallel at the point you start considering equipment data as an AI input.

Regulation in Thailand, Vietnam and Across ASEAN

If your programme touches an ASEAN site, home-country law alone is insufficient. Here is what to track as of 2026.

In Thailand, AI legislation is being drafted. As reported, the framework under consideration would regulate by use case across three tiers — prohibited, high-risk and general use — led by the Ministry of Digital Economy and Society, with ETDA drafting the bill and the AI Governance Center (AIGC) supporting. Passage is targeted within the fiscal year, and for public agencies the discussion includes practical sanctions applied through personnel evaluation. Unauthorised use of copyrighted works for training is one of the contested points. The critical qualifier: this is a bill and it is not yet enacted. Briefing your board on the premise that “Thailand’s AI law has passed” will require a correction later. The practical response is to check whether any of your planned use cases could fall into a high-risk category — recruitment screening, credit decisions, control of critical infrastructure — and, if so, to design for explainability and log retention now rather than after enactment.

The more immediate exposure in Thailand is the PDPA, which is already in force. Penalties are split between administrative and criminal tracks: administrative fines are reported at up to THB 5 million, and in serious cases criminal penalties including imprisonment may apply separately. The governing principle is that personal data may only be used within the purpose notified at the point of collection. The generative AI implication is direct: if “training an AI model” was not part of that original purpose, additional consent is required. Before ingesting employee or customer data into an internal AI platform, go back and read the consent language you actually collected.

The Vietnam-specific position is further along. Vietnam’s Law on Artificial Intelligence takes effect on 1 March 2026. It sets out principles for development, testing and application, and provides for a regulatory sandbox, incentives for R&D, data and computing infrastructure, and protection of enterprise rights. The background is Resolution 57/NQ-TW, issued by the Politburo at the end of 2024, which sets the target of placing Vietnam among the top three ASEAN countries in AI research and development by 2030. Vietnam has also approved a national digital transformation strategy for 2026–2030, under which more than 40% of SMEs are expected to adopt MaaS, AI agents and shared AI solutions, with the domestic ecosystem led by VNPT, Viettel, FPT and CMC.

For a group with sites in both countries, the efficient design is to standardise the parts that do not change — data classification, logging, permission management — and to handle only the country-specific deltas separately.

Running Generative AI on a Multilingual Shop Floor

There is an operational dimension to generative AI implementation at ASEAN plants that rarely appears in generic guidance: language. On a typical floor, expatriate managers work in the corporate language, local staff work in Thai or Vietnamese, and English is the medium with equipment vendors and international customers. Work standards exist as an original plus a translation, equipment manuals arrive in English, daily reports are written in the local language, and reporting to the parent company happens in a fourth register. That is normal, not exceptional.

One Thailand-specific technical point is worth knowing: Thai script does not use spaces between words. Models that have not been trained sufficiently on Thai can therefore segment words incorrectly. Output may read fluently while the boundaries of a proper noun or a technical term have been misplaced — which quietly degrades retrieval and classification accuracy.

Work on Thai-specialised models is active. SCB 10X’s Typhoon family is the leading example: Typhoon-7B is published as open source, Typhoon2.1 4B is described as a Thai/English bilingual model built on Gemma3, and Typhoon2-Audio adds speech capability. The community-driven OpenThaiGPT takes a different route, building on Llama2 and adding 24,554 Thai tokens to the tokenizer.

That said, it would be wrong to conclude that general-purpose models are unusable for Thai. Frontier models have improved their Thai handling considerably year over year and are entirely adequate for some workloads. The correct conclusion is narrower and more useful: verify per use case. For internal document search, form reading, work-instruction translation and daily-report summarisation, build a sample from your own real data, measure accuracy, and then decide. Skipping that verification and rolling out company-wide is how a tool gets labelled useless on the floor and never gets a second chance.

There is one more design decision worth making explicitly: separate the language of the answer from the language of the source. Answer a Thai operator in Thai and a manager in the corporate language, while both draw on the same underlying source document. Designed that way, the chronic manufacturing problem of translated work standards drifting out of sync with the original becomes structurally less likely.

Which Use Cases Actually Require Your Own Data

A major reason generative AI plans stay vague is that wildly different levels of difficulty get bundled under one phrase. Classify use cases by whether they require integration with your own data and the sequencing decides itself.

The top three rows can, at the extreme, be started with a contract and a training session. The bottom three succeed or fail not on model performance but on whether the data exists. Most companies correctly recognise that the value sits in the lower rows, then start at the top and stop there — or leap straight to the bottom, discover the data is missing, and abandon the effort. Those are the two dominant patterns.

The realistic path is to run both tracks at once: use the upper rows to build organisational familiarity while separately investing in the data preparation the lower rows require. Data preparation is unglamorous work that is not, strictly speaking, an AI project. But without it, generative AI remains a document tool permanently. On the document-reading side specifically, we have covered the practical handling of the form types common in this region in AI-OCR for back-office automation.

Value Appears When AI Connects to Factory Systems and Floor Data

Pulling the threads together, what separates a successful generative AI implementation at a manufacturing site from an abandoned one is fairly clear: whether it connects to the systems where work already happens, and whether the data on the other side of that connection is in a usable state.

The Anthropic study ranked integration with existing systems first at 46% and data access and quality second at 42%. The ETDA/NSTDA study named data quality as the primary barrier for Thai organisations. The Teikoku Databank survey showed Japanese companies concentrating 45.1% of usage in text work while data aggregation and analysis sat at 7.4%. Three independent studies, three methodologies, one conclusion: generative AI moves operational numbers only when it is connected to operational data.

At a plant, there are broadly three things to connect. The first is production management data — plans, actuals, inventory, costing, work-in-process status. It sits inside the production system, but is often hard to extract because the system is old, because each site runs something different, or because it is shadowed by a parallel spreadsheet. It is common for an AI evaluation to end up triggering a production management system review; we cover that selection process in choosing a production management system for factories in Thailand.

The second is data trapped in paper and PDFs: supplier invoices, customs documents, inspection certificates, handwritten daily reports. These become AI inputs only once AI-OCR has structured them. In this region, forms routinely mix three languages on a single page, which makes multilingual capability a baseline requirement rather than an enhancement.

The third is time-series data from equipment: operating status, power consumption, temperature, pressure, cycle time. Only when IoT-collected equipment data is cross-referenced with production and quality data do you have the material to answer a question like “why did yield drop on this particular day?”

What sits beyond those three connections is process automation — work advancing through the flow without a person stopping to ask a model a question.

TOMAS TECH CO., LTD. is a factory IT and systems integrator based in Bangkok, providing the PEGASUS production and energy management system, factory visualisation, IoT equipment monitoring, AI-OCR and custom development to manufacturers across Thailand and ASEAN. In generative AI work, our emphasis is on the design question of how a new capability connects to the factory systems and floor data you already run, rather than on deploying a tool in isolation. Being able to work locally in Japanese, Thai and English is a practical difference on a floor where multilingual operation is a given.

Frequently Asked Questions

How much does generative AI implementation cost?

Published vendor ranges from the Japanese domestic contract development market put a proof of concept at JPY 1.5–5 million and full production implementation at JPY 15 million and above; a second source models a PoC at JPY 1–5 million and production implementation at JPY 800,000–2.5 million per month multiplied by active months. These are Japanese domestic benchmarks, not procurement prices in Thailand or Vietnam. Public statistics on Thai per-man-month rates are sparse, so this article does not assert one. Practically, obtain quotes from several providers broken into the same five line items — initial build, licences, usage-based fees, maintenance, training — and compare them at equal granularity. Budget from the outset for the fact that most of the spend goes to data preparation and integration rather than to the tool itself.

How should we implement generative AI, step by step?

Start with an inventory. Map the processes and the data, set your evaluation metrics, and only then move through common tooling and rules, integration with internal documents, embedding into business processes, and replication across sites. The critical planning move is to separate use cases that require no integration with your own data (summarisation, translation, drafting) from those that do (document processing, equipment data analysis, supply/demand simulation). Building familiarity with the former while preparing data for the latter in parallel is the design least likely to stall.

Why do generative AI PoCs fail so often?

Rarely because of the model. In the Anthropic study of more than 500 technical leaders published in December 2025, the leading obstacles were integration with existing systems at 46%, data access and quality at 42%, and change management at 39%. The Japan-specific Teikoku Databank survey of March 2026 ranked accuracy of information at 50.4%, shortage of specialist talent at 41.3% and unclear scope at 40.0%. Layered on top of these are process failures: starting without acceptance criteria, running the pilot without the people who will operate it, and selecting a platform without aligning to the group’s security standard first.

How do we control the risk of data leakage?

Work in five layers: contract and tenancy, access rights, data classification, logging and audit, and education. Confirm an enterprise agreement under which your inputs are not used for training, and — most importantly — audit your existing permission design before pointing any retrieval system at internal documents, because loose permissions make previously invisible information retrievable by anyone. For context, the Japan-specific March 2026 Teikoku Databank survey found 33.5% citing leakage risk as a concern while only 0.7% reported an actual leakage incident and 67.7% reported no negative effects. Concern is high; reported harm to date is limited.

Can an ASEAN site use the same tools as the parent company?

Technically, usually yes, but three things need checking. First, licensing: is the local entity actually inside the group agreement? Where it is not, staff commonly fall back to unmanaged personal accounts. Second, data residency and cross-border transfer. Thailand’s PDPA restricts use of personal data to the purpose notified at collection, so if AI model training was not part of that purpose, additional consent is required. Third, localisation of the rules themselves — distributing a policy that exists only in the corporate language does not change local behaviour.

Will accuracy hold up in Thai on a factory floor?

It depends on the use case. Thai has no spaces between words, so models not sufficiently trained on Thai can segment incorrectly. At the same time, Thai-focused work is advancing — SCB 10X’s Typhoon (open-source Typhoon-7B, the Gemma3-based bilingual Typhoon2.1 4B, and the speech-capable Typhoon2-Audio) and the community-led OpenThaiGPT, which adds 24,554 Thai tokens to a Llama2-based tokenizer. General-purpose models are also improving, so “Thai means it will not work” is not a defensible claim. Test accuracy on your own real data for each use case — document search, form reading, work-instruction translation — before deciding how widely to deploy.

How should responsibility be split between the group and the local site?

A workable split gives the group ownership of the defensive standards — security baselines, contracting, data classification policy — and gives the local site the lead on where and how AI is applied. The site understands local processes, languages, regulation and commercial practice; if the corporate centre dictates that detail, decisions slow to a crawl. Conversely, delegating security standards to each site produces divergent implementations that cannot be consolidated later. Aligning scope with group IT before the funding request is the single most reliable way to avoid rework.

Conclusion

Generative AI implementation at an ASEAN manufacturing site is not a tool selection problem. The Japan-specific Teikoku Databank survey of March 2026 shows adoption reaching 34.5% among companies in Japan while 45.1% of use is concentrated in writing tasks and only 7.4% touches data aggregation and analysis. On the Thailand side, Thailand Digital Outlook 2026 recorded average digital maturity rising to 2.12 out of 4 with AI leading advanced-technology adoption at 21.2%, and the UOB Business Outlook Study 2026 reported more than 70% of Thai SMEs implementing AI. Methodologies and populations differ too much for direct comparison, but the direction is not in doubt: local companies are embedding AI for straightforwardly commercial reasons.

And in Japan, in Thailand, and in global enterprise research, the same three obstacles recur — integration with existing systems, data access and quality, and organisational change. Generative AI moves operational numbers when it is connected to production management data, digitised paper records and equipment time-series data. Until it is, no amount of model capability produces anything beyond faster document drafting. Establish where you stand with actual figures, use the low-preparation use cases to build organisational familiarity, and run data preparation for the high-value use cases in parallel. That two-track design is what keeps a programme from ending at the PoC.

If any of this maps onto a question you are currently sitting with — how much of your factory data is actually usable, where the line between group standards and local discretion should fall, how to phase a roadmap across two or three countries — we are happy to think it through with you, even at the exploratory stage where no plan or budget exists yet. A conversation that starts with sorting out the current situation is often the most useful one. You can reach us through the contact form.

References

• Teikoku Databank, “Survey on Corporate Trends in Generative AI” (surveyed 17–31 March 2026; 23,349 companies approached, 10,312 valid responses) — https://www.tdb.co.jp/report/economic/20260514-genai/
• Bangkok Post, “AI adoption helps Thai firms become digitally mature” (Thailand Digital Outlook 2026; 834 business operators) — https://www.bangkokpost.com/business/general/3283789/ai-adoption-helps-thai-firms-become-digitally-mature
• Bangkok Post, “Data quality concerns a barrier to adoption of AI” (ETDA/NSTDA, surveyed July–September 2024; 3,758 organisations approached, 580 responses) — https://www.bangkokpost.com/business/general/3051260/data-quality-concerns-a-barrier-to-adoption-of-ai
• ThaiPR, “UOB Business Outlook Study 2026” (265 business owners and decision-makers in Thailand) — https://www.thaipr.net/en/business_en/3735803
• Anthropic, “How enterprises are building AI agents in 2026” (with Material; 500+ technical leaders; published 9 December 2025) — https://claude.com/blog/how-enterprises-are-building-ai-agents-in-2026
• Bangkok Shuho, “Thailand’s draft AI legislation” — https://bangkokshuho.com/thainews-1236/
• Unimon, “Thailand PDPA AI Compliance Checklist” — https://unimon.co.th/en/blog/thailand-pdpa-ai-compliance-checklist
• VietnamPlus, “Vietnam shapes AI ecosystem following 18-month implementation of Resolution 57” — https://en.vietnamplus.vn/vietnam-shapes-ai-ecosystem-following-18-month-implementation-of-resolution-57-post347781.vnp
• VietnamPlus, “Vietnam approves national digital transformation strategy for 2026-2030” — https://en.vietnamplus.vn/vietnam-approves-national-digital-transformation-strategy-for-2026-2030-post348631.vnp
• Renue, “Generative AI Development Cost and PoC Pricing Guide 2026” (Japanese domestic contract development market) — https://renue.co.jp/posts/generative-ai-development-cost-poc-pricing-guide-2026
• SCB 10X, “Introducing Typhoon 2 Thai LLM” — https://www.scb10x.com/blog/introducing-typhoon-2-thai-llm