Requests to get staff using AI have become routine at Japanese-affiliated plants across Thailand and Vietnam. So has the follow-up three months later: the session went well, the room was engaged, nobody uses it now. Generative AI training is not an investment that ends when the lecture ends. In a plant where a few Japanese expatriates lead a Thai- or Vietnamese-speaking workforce, a programme built for a Tokyo head office does not connect on arrival.
This article covers why training fails to stick, what the four formats cost, what Thai and Vietnamese policy means for hiring, how to design for a multilingual site, what each job layer needs, how to run 90 days, and how to convert the result into baht. It is written for plant managers, HR and L&D leads, and regional heads defending the budget line.
Why generative AI training became a management issue
Through 2023 and 2024 the question was whether to use generative AI at all. In 2026 the gap between users and non-users shows in the speed of ordinary work.
51.0% of manufacturing staff already use it in some form
Arsaga Partners surveyed 311 manufacturing employees on 22-23 July 2025 (online panel):
- Using it actively: 16.3%
- Using it in parts of the job: 29.3%
- Piloting it: 5.4%
- Interested but not using it: 17.2%
- Neither interested nor using it: 31.7%
The first three total 16.3 + 29.3 + 5.4 = 51.0%. The group that has not touched it is 17.2 + 31.7 = 48.9%; the two sides sum to 99.9% because of rounding.
Read that half carefully: only 16.3 points represent active use, and the remaining 51.0 − 16.3 = 34.7 points sit at “parts of the job” or “piloting”. In Thailand, expatriates summarise meeting notes on personal accounts while local staff have been issued nothing. That is not adoption but an unmanaged information risk.
87.6% of those who use it report a gain
Among users, on work efficiency:
- Improved a great deal: 21.3%
- Improved somewhat: 66.3%
- Did not improve much: 7.1%
- Did not improve at all: 1.2%
Gains total 21.3 + 66.3 = 87.6%; no meaningful gain totals 7.1 + 1.2 = 8.3%. At 87.6 ÷ 8.3, approximately 10.6, ten people report a benefit for every one who does not. A company seeing no return usually never started.
The number one blocker is not knowing what to use it for
Reasons for non-adoption (multiple answers, pt):
| Reason for not adopting | Response (pt) |
|---|---|
| Do not know what to concretely use it for | 46.7 |
| Hard to connect with existing systems | 25.7 |
| Lack of implementation knowledge and skills | 19.0 |
| Return on investment unclear | 13.3 |
| Security concerns | 11.4 |
The top answer leads the second by 46.7 − 25.7 = 21.0 pt, a ratio of 46.7 ÷ 25.7, approximately 1.8 times. Read the composition, not the ranking: first (46.7 pt) and third (19.0 pt) are people problems, while the second (25.7 pt) is a data problem. Training alone solves one of the two. PwC Japan’s survey work adds the same point qualitatively: competition is shifting from adopting AI to capturing value from it, and readiness is not a licence count but knowing who can access which data and how output is evaluated.
Our generative AI implementation guide maps the application areas process by process, which makes choosing exercise material easier.
We ran the training and nobody uses it: four root causes
Cause 1: the exercise material is not your own work
Standard programmes use cross-industry material such as drafting an email, teaching the buttons but leaving nothing about tomorrow morning. What lands on a factory floor looks like this:
- Turn a Thai-language defect report into a Japanese summary for the head office
- Search three years of corrective action reports and draft a prevention plan
- Turn an English equipment manual into a Thai work instruction for operators
- Read a customer’s specification-change email and list the affected processes
None can be practised without your own documents. Committing to prepare ten real internal documents most determines whether the day works.
Cause 2: accounts and permissions were never issued
Participants use the instructor’s environment, then return to a desk with no account. Where IT decisions sit with the Japanese head office, approvals take weeks to months and momentum is gone before access arrives. The other half is rules: with no guideline on permitted inputs (customer drawings, cost data, appraisal comments), the most conscientious employees stop first. Provision accounts and distribute input guidelines in the local languages before the training date; if that cannot be done, move the date.
Cause 3: the data does not exist in a form AI can use
Hand real work to AI and you hit the question of where the information behind the judgement lives.
- Production results are in the system, but daily report remarks are handwritten
- Trouble history sits in veterans’ memory and personal spreadsheets
- Drawings sit on a shared folder with no naming convention, unsearchable
- Supplier documents arrive as PDFs and are re-keyed
In that state AI can only produce text from general knowledge; translation still pays, but answers grounded in your own history do not. We cover the mechanism in RAG implementation for factory knowledge; the point is that this groundwork runs in parallel, not afterwards. For back offices buried in paper, AI-OCR for back-office automation is the faster entry point and makes good training material, because results show up as numbers.
Cause 4: nobody agreed how success would be measured
Measure with a satisfaction survey and training becomes a one-off event; high scores tell you nothing at budget time. Measure what people stopped doing and what they shortened, converted into money through hours saved × headcount × hourly cost.
An external provider can solve cause 1 and part of cause 4; causes 2 and 3 belong to the buyer. Missing that split produces a good instructor and an unused tool.
Four types of corporate AI training and what they cost
Cost has no single answer: the order of magnitude changes with format. The bands below are Japanese domestic rates in yen, not a quote for Thailand.
| Format | What it is | Price band (Japan domestic) | Best suited to |
|---|---|---|---|
| 1. E-learning | Video on basic operation and risk, at each person’s own pace | A few thousand yen up to JPY 30,000 per person | Basic knowledge company-wide |
| 2. Instructor-led group training | Half-day to full-day session on a general curriculum | JPY 200,000 to 500,000 per session | Giving one department the basics |
| 3. Customised training | Built around your own operations; one day for 5 to 20 people runs JPY 500,000 to 700,000 | JPY 500,000 and above in total | A department’s real work |
| 4. Accompanied practical programme | Multiple sessions alongside actual process improvement | JPY 1,000,000 to over 3,000,000 | Company-wide adoption with KPIs |
Per person, formats 2 and 3 swing with headcount. The one-day customised session at JPY 600,000 is 600,000 ÷ 20 = JPY 30,000 per person for 20 people and 600,000 ÷ 5 = JPY 120,000 for five: a four-fold spread on identical content. E-learning at JPY 10,000 per head for 50 people is 10,000 × 50 = JPY 500,000, the same bracket, and an accompanied programme at JPY 2,000,000 for 20 people is 2,000,000 ÷ 20 = JPY 100,000 per person. Combine formats by headcount and objective.
Choosing the right mix
Under 100 employees, use e-learning for management plus one group session, to create a few people internally who can hold the conversation. From 100 to 500, run customised training department by department with e-learning for the baseline, because the work varies too much for one curriculum. For a DX talent pipeline, anchor on the accompanied programme as an annual plan. The rule of thumb: do not buy format 4 first, and do not stop at format 2.
A note on Japanese training subsidies
Groups with a Japanese parent ask about Japan’s Human Resources Development Subsidy. As of FY2026 it covers 75% of expenses for SMEs and 60% for large companies, with a wage subsidy of JPY 1,000 per person-hour for SMEs (raised from JPY 960 in FY2026) and JPY 500 for large companies, and FY2026, ending March 2027, is expected to be the final year. For an SME spending JPY 2,000,000 on 12 hours of training for 20 people: expense subsidy 2,000,000 × 0.75 = JPY 1,500,000, leaving 2,000,000 − 1,500,000 = JPY 500,000; wage subsidy 1,000 × 20 × 12 = JPY 240,000; net cost 500,000 − 240,000 = JPY 260,000. The scheme applies to employers in Japan and cannot be used for training a Thai entity funds for its own staff. In practice the parent claims for head-office employees, the Thai entity budgets separately, and both share one set of material. Confirm eligibility with the labour bureau.
The policy environment in Thailand and Vietnam
National workforce policy feeds the recruitment market, which sets your labour cost and competition for talent.
Thailand: THAILAND² and the five-ministry skills overhaul
On 21 July 2026, under the THAILAND² strategy, the Thai government announced a skills overhaul across five ministries: Higher Education, Science, Research and Innovation; Social Development and Human Security; Education; Labour; and Agriculture and Cooperatives. The priority skills are digital literacy, data analysis and collaboration with advanced technology, against a backdrop of moving from low-cost contract production to higher value-added industry. No budget figure has been published. The significance is directional: as policy lifts the digital skill level of Thai employees, the line between what you develop internally and what you hire will move.
Thailand’s National AI Strategy and its three-tier talent target
| Tier | Target | Role |
|---|---|---|
| AI professionals | 10,000 people | Building and implementing AI systems |
| AI innovators | 20,000 people | Applying AI to business and operations |
| AI literacy | General public | Basic understanding and everyday use |
The Thailand National AI Strategy and Action Plan 2022-2027 sets the tiers above; the upper two total 10,000 + 20,000 = 30,000 people, and smart manufacturing is a priority sector. The World Bank has separately warned of a structural labour crisis in Thailand and urged reskilling for the AI era.
Vietnam: Southeast Asia’s first standalone AI law
- AI Law: in force from 15 March 2026, the first standalone AI law in Southeast Asia
- Five-year personal income tax exemption for AI specialists: a direct incentive to attract high-level talent
- Programme to train 20,000 practical AI specialists: responding to Resolution 57, launched by VinUni in January 2026, with roughly 500 people in the first cohort that April and 2,000 planned across three cohorts
- Decree No.179/2026/ND-CP: making digital literacy compulsory across all majors
Keep the scale in proportion: against the 20,000 target, 2,000 across three cohorts is 2,000 ÷ 20,000 = 10%. Still, a country running a numerical target, a legal framework and a tax incentive at once cannot be ignored. With a Vietnamese site, check with local counsel how your AI usage is classified.
Folding policy into your training plan
Do not assume staff start from zero: younger employees have often used AI tools privately, so open with an anonymous usage survey. Rebalance hiring against development, because competition for specialists will intensify and the higher return is managers who can decide what to hand to AI. And do not wait for policy, which takes years to show results.
Designing generative AI training for a multilingual plant
Build the glossary before you split by language
The first step is not a curriculum per language. It is a glossary of your own internal vocabulary. Process abbreviations, equipment nicknames, defect codes and form names circulate as a mixture of Japanese, Thai and English and drift between departments. Leave that unresolved and one process is named three ways, a Japanese query never reaches the Thai records holding the answer, and output uses terms nobody recognises.
Build a Japanese, Thai and English glossary of roughly 100 to 200 core operational terms, adding Vietnamese where relevant, as a precondition not a by-product. It also underpins later system integration and search, and can itself be an AI exercise.
Which language for which layer
Japanese for the executive layer and Japanese managers, because these sessions deal with investment judgement and where to draw the line on risk. Thai for line leaders, supervisors, QC and Thai back-office staff, because in Japanese or English their energy goes into decoding the language, not the application. The mixed forum with both together is hardest to design. With Vietnamese sites, keep the core material common and swap only the exercise documents.
The hidden costs of training through an interpreter
Prompts get translated. An interpreter rendering a prompt into Thai leaves participants unsure what to type; prompts belong on screen as literal strings.
Technical terms drift. Hallucination, token, context: no settled translations exist, so put AI terminology in the glossary.
Everything takes 1.5 to 2 times as long. Consecutive interpretation stretches the same content to roughly 1.5 to 2 times the time, so a six-hour Japanese-language day loses its exercise time until only the lecture survives. Design for 60 to 70% of the content volume.
The questions stop. Asking through an interpreter carries a social cost, so use small group work and collect questions by group.
The best answer is an instructor who teaches in Thai. Failing that, developing Thai managers first to teach the floor sticks better than interpreting.
Use generative AI itself as the translation tool
Generative AI produces its most legible wins here, so it belongs at the centre of the material:
- Convert a Thai defect report into a Japanese summary for the head office
- Rewrite a head office instruction email as plain-language Thai work steps
- Pull the relevant sections out of an English equipment manual, in Thai
- Restructure a Japanese internal regulation as a Thai FAQ
These attack work that has long consumed expatriate time. One rule is non-negotiable: nothing translated leaves the company unchecked, and customer-facing or contractual text passes through a human.
Building the curriculum by job layer
Identical training for everyone is easy to budget and has the lowest retention of any approach, because the target outcome per layer differs. The design below assumes 100 to 500 employees, one day counted as six hours.
| Job layer | Target outcome | Core material | Language | Indicative time |
|---|---|---|---|---|
| Executive (MD, plant manager) | Can decide investment and where to draw the risk line | Application map, leakage risk, KPI design, budget allocation | Japanese | Half day (3 hours) |
| Managers (department and section heads) | Can break the team’s work down and decide what AI takes on | Work inventory, use case selection, output review, appraisal linkage | Japanese plus Thai | 2 days (12 hours) |
| Line leaders (supervisors, QC) | Can replace routine daily tasks | Work instructions, defect report summaries, multilingual communications | Thai / Vietnamese | 1 day (6 hours) |
| Back office (purchasing, finance, HR, production control) | Can automate document and form handling | Quotations and invoices, meeting summaries, regulation search, Excel formulas | Thai with Japanese support | 1.5 days (9 hours) |
Managers get 12 hours, executives 3. That is deliberate.
Why management AI training comes first
Only managers can decompose the work. Deciding that AI drafts and a human checks needs someone who knows the whole process and the quality standard.
A tool the manager does not use will not take root below. We have never seen a tool embed itself where the supervisor does not understand it but tells the team to go ahead.
Only managers can connect it to evaluation. Without a design for how you appraise someone who shortened a task, telling people to be more efficient teaches the floor that finishing early only brings more work. For the two manager days, day one covers fundamentals plus a work inventory and day two prototyping the selected use cases.
How much prompt engineering training do you need?
For general employees, advanced technique is not necessary: model quality has improved to the point where usable output no longer needs elaborate instructions. The floor needs three habits:
- State the premise and goal first (“You are a quality control engineer. Summarise the defect report below in 300 Japanese characters for the head office.”)
- Specify the output format (bullets, table, length)
- When the result is wrong, say what is wrong and have it revised
The exception is internal IT and production control staff embedding AI into business systems, who need deeper prompt design and API knowledge; budget a dedicated course for them. That work sometimes involves technologies outside the generative family, such as AI visual inspection on the line, so an internal IT plan covering image recognition and data analysis fits real plant problems.
A 90-day programme that makes the training stick
| Period | Objective | Main activities | Exit criteria to move on |
|---|---|---|---|
| Days 0-30 | Common language and safety rules | Kickoff, guidelines per language, account provisioning, first glossary, foundation courses by layer | 90% hold an account and log in weekly / input rules distributed in every language |
| Days 31-60 | Apply it to your own operations | Work inventory, three use cases per department, shared prompt library, data groundwork started | 2 or more patterns used weekly per department / hours saved being measured |
| Days 61-90 | Embed it and scale sideways | Results review, use cases documented and integrated, champions appointed, KPI review | Monthly hours saved at 70% or more of target / 3 or more departments confirmed for rollout |
Days 0-30. The objective is a common starting point, not results: what may and may not be entered, the rule that output is always checked by a person, and the agreed vocabulary. Account provisioning is critical, so start head office approval a month before kickoff. The exit criterion is a number because this is where projects run on impression; move on at 80% and the last 20% never come back.
Days 31-60. Each department picks three tasks that are high frequency, routine in judgement, and non-critical if they go wrong. This is when the data problem surfaces, and only data groundwork in parallel moves the programme forward. Check here whether production results are entered correctly and forms can be extracted as structured data.
Days 61-90. Move from “people who want to use it, use it” to “it is part of the procedure”. Proven use cases go into standard operating procedures, prompts into a departmental template library, and where possible into business systems so nobody writes a prompt. Appoint an internal champion per department, because the instructor leaves at day 90, with formal approval to spend 5 to 10% of working time on it.
Measuring the impact: KPIs and ROI
| Layer | What it measures | Example indicators | Frequency |
|---|---|---|---|
| Layer 1: activity | Is it being used | Weekly active usage rate, sessions per person | Weekly |
| Layer 2: proficiency | Can people use it | Use cases per department, templates registered | Monthly |
| Layer 3: operations | Has the work changed | Time per task, lead time, rework incidents | Monthly |
| Layer 4: financial | Did it reach the P&L | Hours saved in money, outsourcing cost, overtime | Quarterly |
Most companies look only at layer 1 and declare victory on usage rates, or reach for layer 4, fail to measure it and give up. Layer 3 is the highest-value measurement, and layer 4 is then arithmetic: ask people to self-report minutes per week on a task before the programme, and again after 90 days.
Converting hours into money
Monthly saving = hours saved per person per day × number of people × working days per month × hourly cost
Keep the currency consistent on both sides. The examples below are calculated entirely in Thai baht.
Pattern A: 20 back-office staff, 30 minutes saved per person per day
- Hours saved: 0.5 hours × 20 people × 20 days = 200 hours per month
- Hourly cost: THB 30,000 monthly salary ÷ 160 working hours per month = THB 187.5 per hour
- Monthly saving: 200 hours × THB 187.5 = THB 37,500 per month
- Annualised: 37,500 × 12 = THB 450,000 per year
Pattern B: 60 line leaders, 15 minutes saved per person per day
- Hours saved: 0.25 hours × 60 people × 20 days = 300 hours per month
- Hourly cost: THB 20,000 monthly salary ÷ 160 working hours per month = THB 125 per hour
- Monthly saving: 300 hours × THB 125 = THB 37,500 per month
- Annualised: 37,500 × 12 = THB 450,000 per year
Both land on THB 37,500 per month: a large saving for a few and a small saving for many can be worth exactly the same. Together, 37,500 + 37,500 = THB 75,000 per month, or 75,000 × 12 = THB 900,000 per year.
Two adjustments apply. Employment on-costs: social security, bonuses and benefits push the real cost above salary. At a factor of 1.3, Pattern A becomes 187.5 × 1.3 = THB 243.75 per hour, so 200 hours × 243.75 = THB 48,750 per month; Pattern B becomes 125 × 1.3 = THB 162.5 per hour, so 300 hours × 162.5 = THB 48,750 per month. Combined, 48,750 + 48,750 = THB 97,500 per month, or 97,500 × 12 = THB 1,170,000 per year. Retention: at a conservative 60%, against the base case without on-costs, 75,000 × 0.6 = THB 45,000 per month, or 45,000 × 12 = THB 540,000 per year.
Frame the decision as the range between the conservative THB 540,000 and optimistic THB 1,170,000 per year, using your own salary levels, working days and on-cost ratio.
Reinvest the freed-up time or the return is zero
Every calculation above assumes the saved time is redirected into valuable work. If it disappears into waiting, the effect is zero: no headcount is reduced, no new work absorbed, and the hours appear nowhere in the accounts. Decide during planning what the released time is for:
- Preventive maintenance and improvement that never had capacity
- Reducing overtime, which lowers actual labour cost
- Absorbing higher volume without adding headcount
- More time on quality checks and customers
The choice changes what you measure: overtime hours if you target overtime, output per person if you target volume. Whether an AI adoption support partner will discuss this is a meaningful selection criterion.
Five common failures and how to avoid them
1. Launching company-wide at once. What executives need and what line leaders need do not overlap, so identical content is half-useful to all. Split by job layer, and where budget is tight start with one department of about 20 people, then expand.
2. Taking material from generic examples. Assemble ten or more real internal documents at least three weeks ahead, and do not delegate that to the vendor.
3. Running training before accounts exist. Complete provisioning before the training date, or postpone it, and distribute permitted-input rules as a one-page sheet in each working language.
4. Relying on interpretation. Glossary first, and where possible an instructor teaching in the local language. If interpretation is unavoidable, cut content to 60 to 70% and present prompts on screen.
5. Making satisfaction scores the KPI. A satisfaction survey evaluates an instructor, not an investment. Measure how long the target tasks take before the training, and put the three- and six-month reviews in the calendar now.
Summary: training and data groundwork only work together
51.0% of people working in manufacturing have already touched generative AI, and 87.6% of users report improved efficiency. Yet the leading reason for non-adoption is not knowing what to concretely use it for, at 46.7 pt, followed by difficulty connecting with existing systems at 25.7 pt. The problem sits on both the people and data sides.
Training fails to stick for four reasons: the material is not your own work, accounts are missing, the data is not usable, and nobody agreed how success is measured. An external provider solves the first and part of the fourth. Costs run, in Japanese domestic terms, from a few thousand yen up to JPY 30,000 per person for e-learning to JPY 1,000,000 and beyond JPY 3,000,000 for an accompanied programme, and Japan’s subsidy, at 75% of expenses for SMEs plus JPY 1,000 per person-hour, cannot fund training paid for by a Thai entity.
Build the glossary before the language-specific curricula, weight investment toward managers, split 90 days into three phases with a numerical exit criterion each, and decide in advance where the released time goes. Generative AI training does not stick as a standalone purchase: correct production results, structured forms and searchable floor knowledge are what make it pay for itself.
TOMAS TECH is based in Bangkok, supplying the PEGASUS production control system and energy management systems to manufacturers in Thailand, so we spend our days looking at how operational data is recorded on the floor and where it stops. We are increasingly asked whether to start with training or with data groundwork. Either stage is a fine place to begin the conversation, and so is the stage before that, where you simply want an outside view on the condition of your data. Nothing has to be decided beforehand. If this matches your situation, the contact form reaches us directly.
Frequently asked questions
How much does generative AI training cost?
It depends on format. In Japanese domestic terms: e-learning from a few thousand yen up to JPY 30,000 per person; instructor-led group training JPY 200,000 to 500,000 per session; customised training JPY 500,000 and above in total; and a multi-session accompanied programme JPY 1,000,000 to over 3,000,000. Per person, a JPY 600,000 one-day session is JPY 30,000 per head for 20 people and JPY 120,000 for five, so headcount dominates.
Should training for Thai staff be delivered in Thai?
For line leaders, supervisors and Thai back-office staff, yes. In Japanese or English their attention goes into decoding the language rather than into how it applies to their work. Where a Thai-speaking instructor is unavailable, developing Thai managers to cascade internally retains better than interpreting. A Japanese, Thai and English glossary prepared in advance is a precondition either way.
Can our Thai entity use Japanese government training subsidies?
No. Japan’s Human Resources Development Subsidy applies to employers in Japan, so training a Thai entity funds for its own staff falls outside it. Training run by the parent for head-office employees, including staff preparing for overseas assignment, can be considered, so budget the two sides separately while sharing material. FY2026, ending March 2027, is expected to be the final year.
Should we train managers or the shop floor first?
Managers. Only someone who knows the whole process and the quality standard can decompose work and decide what goes to AI; a tool the supervisor does not use will not take root below; and only management can design how to appraise someone who shortened a task. Shop-floor training works better once managers have selected their department’s use cases.
Is a prompt engineering course enough on its own?
For general employees the need for advanced technique is falling as models improve. The floor needs three habits: state the premise and goal first, specify the output format, and say what was wrong when the result misses. The exception is internal IT and production control staff embedding AI into business systems, who need deeper prompt design and API knowledge.
References
- Arsaga Partners, Survey on Generative AI Usage in Manufacturing (9 September 2025)
- PwC Japan Group, Generative AI Survey 2026 Spring (six-country comparison)
- Human Resources Development Subsidy, Reskilling Support Course (FY2026)
- Cost benchmarks for AI and generative AI training
- Nation Thailand, Thailand launches five-ministry skills overhaul (21 July 2026)
- Thailand National AI Strategy and Action Plan 2022-2027
- Nation Thailand, World Bank warns of Thailand labour crisis
- Bangkok Post, Thailand intensifies AI efforts in bid to close the skills gap
- IAPP, Vietnam’s first standalone AI law: an overview of key provisions, future implications
- Vietnam.vn, Programme to train 20,000 practical AI specialists (Resolution 57)
