Generative AI Consulting in Thailand 2026 — Four Contract Types and a Five-Layer Cost Model

If you start your evaluation with the question “should we hire someone for generative AI consulting,” you will almost certainly end up with a stack of proposals that cannot be compared. The term covers four completely different contracts — buying a diagnosis, buying a set of boundaries, buying someone to stay with you until adoption sticks, and buying a working system. The deliverables, the duration, the cost and the way each one fails are all different. This article replaces the question with a better one — what exactly are we contracting to buy — and works the five-year economics through with real figures from a Japanese-owned manufacturer in Thailand.

Why inquiries about generative AI consulting are rising in 2026

Something changed in the questions we get from managers at Thai subsidiaries this year. Through 2024 and 2025, the typical inquiry was exploratory — what is generative AI, could we use it here. In 2026 the questions are different. “Head office has instructed us to apply it to real work by year end.” “We ran a PoC last year and it never made it into production.” “We were told to define what is allowed, and nobody can decide.” These are all problems that occur after the project has already started moving.

There is data behind the shift. Inquiries are not rising because interest is rising. They are rising because adoption moved one step forward, and a new class of blockage appeared everywhere at once. Let us take the pieces in order.

Adoption rates went up. What stalled is everything after that

According to the Deloitte Thailand Digital Transformation Survey 2026 (reported August 2026), the AI implementation rate among Thai companies has reached 61%, up sharply from 47% the previous year. More than half of all companies now have AI in their operations in some form. That number tells you the evaluation phase is over.

The breakdown is where the problem lives. In the same survey, only 19% of companies reached full enterprise-wide deployment. The remaining 81% are stuck in experiments and pilots. So what the 61% adoption rate actually describes is a population of companies that have tried it but have not embedded it into their work. And the meaning of that 81% is that the cause of failure is not tool selection. If the tools were bad, the adoption rate itself would not have climbed. It climbed and then stalled — which means the blockage sits downstream of introduction, in the preparation of the business side, in adoption, and in ongoing operation.

The reported barriers point the same way. The largest barrier was a shortage of technical talent and skills at 71%, and 40% of companies said they could not identify usable use cases. Not being able to identify a use case is not a technology problem. It is a problem of how clearly a company sees its own work. On the results side, only 9% of companies said AI generated new revenue, while 50% achieved cost reduction. That asymmetry matters. Generative AI barely functions yet as a tool for creating revenue, and works for roughly half of adopters as a tool for cutting time out of existing work. The moment you design your internal approval request around a promise of new revenue, you are betting on the 9% side.

There is one more number in the Deloitte survey that tends to get overlooked. Roughly 70% of companies treat AI as the responsibility of the IT department rather than as a business-led initiative. Whether generative AI succeeds is determined by whether business processes can be changed — and yet ownership sits with a function that has no authority to change them. Sign a consulting contract while that structure is left in place, and your counterparty becomes the IT department, and the deliverable becomes a report that begins and ends inside IT.

What is happening in Thailand is “head office moved ahead, the local site was left behind”

The situation specific to Thai subsidiaries is a gap in progress between head office and the local operation. Japanese head offices spent 2024 and 2025 signing enterprise licences, writing internal policies, and collecting departmental use cases. Meanwhile, at the Thai site, usage in the overwhelming majority of cases is still individuals using free versions to clean up English emails.

This is where the idea of “rolling out the head office contract” appears — the assumption that applying the generative AI consulting contract head office already signed will be the fast path in Thailand. The conclusion up front is that it does not work. There are four reasons, covered in detail later, but here is the short version. The language of the shop floor is Thai. The business process differs from head office. Documents are scattered across paper and local servers. And the regulatory environment is different. All four express themselves the same way — deliverables produced at head office cannot be used here as they are.

What Thai SMEs are doing is also worth watching. In the UOB Business Outlook Study 2026 (June 2026), 265 Thai SME owners responded, and more than 70% had already adopted AI. Among adopters, 58% reported cost reductions and 44% reported productivity gains. Local companies are not waiting. While a Japanese-owned site spends a year waiting for a head office policy, the Thai company in the same industrial estate is winning on quotation turnaround speed.

Market numbers describe spending, not results

Market forecasts appear in almost every proposal, and misreading them leads to bad decisions. Gartner forecasts worldwide AI spending of 2.59 trillion USD in 2026, growing 47% (press release dated 19 May 2026). Total worldwide IT spending is expected to reach 6.15 trillion USD, with IT services at 1.866 trillion USD growing 8.7% (reported February 2026). Spending on generative AI models is growing at a standout 80.8%.

Every one of these is a spending figure. None of them is a results figure. A growing market guarantees nothing whatsoever about whether your own investment will pay back. If anything, when spending surges, the supply side of talent thins out, and the probability rises that an inexperienced consultant arrives on your site at a premium day rate.

The most sober number on the results side comes from MIT Project NANDA (July 2025), which reported that 95% of generative AI pilots produce no measurable impact on P&L. If Deloitte’s 81% describes “never reached production,” MIT’s 95% describes “reached production and still did not move the numbers.” Put the two side by side and the real shape of generative AI investment becomes visible — of the companies that start, fewer than one in five reach production, and only a fraction of those affect the bottom line. Choose the firm that builds its design around those two numbers, not the firm that puts market growth rates in its proposal.

Generative AI consulting splits into four contract types

This is the core of the article. When you send out a request for quotation saying “we would like support introducing generative AI,” the proposals that come back cannot be lined up and compared. Firm A proposes a four-week diagnostic for 400,000 baht. Firm B proposes a nine-month implementation for 3,000,000 baht. Both are quotations for “generative AI consulting.” Pick the cheaper one and you are left holding a diagnostic report. Pick the more expensive one and you get something that works, but the risk remains that the wrong target process was chosen.

Decompose the contract into four types first, decide which one you should be buying right now, and only then request quotations. That single step makes proposals comparable.

Type A Assessment — the contract that decides where to start

What you buy in Type A is a basis for decision. The provider inventories your processes, separates the work where generative AI helps from the work where it does not, ranks the candidates and hands the ranking back. Duration is 4 to 6 weeks, and cost typically runs 150,000 to 400,000 baht.

Type A is genuinely useful for companies on the 40% side that Deloitte identified — the ones who cannot identify a usable use case. When several candidates emerge internally and nobody can decide which to start with, that is a decision-making problem, not a technical one. There is real value in having an outside party rank them.

The classic failure mode of Type A is that it ends with the report. The inventory and the use case evaluation are delivered, a readout meeting is held, and everything stops there. To avoid this, write into the deliverable definition at contract time that the output must include a conclusion narrowing the work to a single process to start in the next 90 days. Not a priority table with five candidates — a conclusion with one. A diagnosis that cannot narrow is not a diagnosis.

Type B Governance — the contract that decides what is permitted

What you buy in Type B is rules. A generative AI usage policy, the line between information that may and may not be entered, a review workflow for exceptions, a logging policy, and training material for employees. Duration is 6 to 10 weeks, and cost typically runs 250,000 to 600,000 baht.

Demand for Type B is high at Thai subsidiaries because the head office policy cannot be used as written. Head office policy is drafted against Japan’s personal information protection law and head office’s IT environment. In Thailand, the PDPA applies, most working documents are in Thai, and the management state of the devices people actually use differs from head office. A policy that has merely been translated reads, from the shop floor, as a document from which you cannot tell what you are allowed to do — so nobody reads it, and people use the tools on their own judgement anyway.

The caution when contracting Type B on its own is that writing rules first makes people freeze. Circulate a list of prohibitions while leaving the permitted scope vague, and everyone errs on the safe side and stops using anything. Rules exist to get things used. In practice, as set out in how to write a generative AI usage policy, the structure that works leads with permissions — “in this process you may go this far” — rather than with an enumeration of bans.

Type C Embedded support — the contract that stays with you until it sticks

What you buy in Type C is time. For 60,000 to 150,000 baht per month across 6 to 12 months, someone works alongside you on building the evaluation dataset, monthly accuracy checks and improvements, and developing internal champions.

Deloitte’s 81% stuck at the experimental stage and MIT’s 95% with no P&L impact are both phenomena that occur after the tool is in. So the contract type that actually drives adoption is Type C. And yet Type C is the hardest to get approved internally, because the deliverable is not a tangible thing like an implementation or a policy document — it is a continuing activity called monthly improvement.

To get Type C approved, you have to define the monthly deliverables concretely. The number of questions in the evaluation dataset, the measured accuracy result, the number of prompts improved, and the scope of work the internal champion can now perform unaided. Write those four into the contract as mandatory items of the monthly report. Contract on the phrase “we will work alongside you” alone, and the monthly meeting degrades into a status update, and six months later nobody can explain what changed.

Type D Implementation included — the contract that buys the working thing

What you buy in Type D is a working system. Build, integration with existing systems, testing and handover, across 3 to 9 months, at a cost with a wide spread of 800,000 to 3,000,000 baht.

That spread is itself a piece of information. The difference between 800,000 baht and 3,000,000 baht is, in most cases, determined by how far you integrate with existing systems. Configuring a generative AI tool on its own so that internal documents become searchable sits at the low end. Pulling specification data from the core system or PLM, referencing a database of past quotations, and dropping the drafted answer into the existing quotation format sits at the high end. A proposal whose breakdown does not state the number of systems to be integrated will generate change orders later.

Before choosing Type D, confirm that the target process really has been narrowed to one. Enter Type D without narrowing, and you follow the path of Scenario A described below.

Where the line falls between consultancies, AI development firms and system integrators

The four types map roughly onto the categories of supplier. Strategy and business consultancies are strong at Type A and Type B, and subcontract Type D implementation. AI development firms live in Type D, and their Type A diagnosis is often cursory. System integrators are strong on Type D work where integration with existing systems carries most of the weight, and many of them can also take on Type C operations.

Do not decide where to go first. Decide which type you are buying first. Once the type is fixed, the appropriate supplier narrows on its own. How to assess each category of firm is covered in detail in how to choose an AI development company, which is worth reading before you go out for competitive quotations.

Break the cost into five layers

Comparing costs as “firm X charges this much” is meaningless. The same total with a different internal allocation produces an entirely different outcome. Here we break a twelve-month contract into five layers, and place side by side the case where the scope is widened (Scenario A) and the case where it is narrowed to a single process (Scenario B).

The model company is a Japanese-owned machined-parts manufacturer in Chonburi, Thailand. The Thai site has 320 employees, including 6 Japanese expatriates. Head office has deployed generative AI company-wide, but locally usage is limited to individuals on free versions. The internal labour rate is 450 baht per hour. The target process is quotation response. There are 180 inquiries per month, each taking an average of 95 minutes of internal information gathering, searching past quotations and confirming specifications, of which 40 minutes is document search and transcription.

First, Scenario A, which starts five processes simultaneously.

The annual running cost from year two onward, corresponding to Layer 5, is 580,000 baht.

Next, Scenario B, narrowed to quotation response alone.

The annual running cost from year two onward, corresponding to Layer 5, is 220,000 baht.

The first thing to notice when the two tables sit side by side is that the shares barely move. The totals differ enormously — 2,800,000 baht against 1,000,000 baht — yet every layer’s share stays within two percentage points. What that means is that changing the scale does not change the shape of the cost. The size of the total is determined not by the quality of the work but by the number of target processes.

Why Layer 2 is always the largest

In both scenarios the largest layer is Layer 2, business-side preparation, at 29.3% in Scenario A and 28.0% in Scenario B. It is larger than implementation in Layer 3.

Most internal approval requests miss this. When people hear “the cost of introducing generative AI” they picture tool licences and build cost. In reality the largest expense is the work of getting your own documents into a state where the model can read them. In the quotation example, past quotations are scattered across individual staff folders, there is no file naming convention, and the same part is called one thing on the drawing and another in the ERP. Search against that state and what comes back is an irrelevant old quotation.

Layer 2 is also a layer that consultants cannot finish alone, because only people inside the company can judge which term refers to the same thing as which other term. So a proposal that allocates only around a tenth of the total to Layer 2 is assuming that the buyer will do this work for free. Before signing, confirm in person-days how much effort the buyer must put into Layer 2. Leave that vague and, halfway through the project, overtime rises on the shop floor and resentment toward the generative AI initiative itself takes root.

Do not import Japanese price levels into Thailand unchanged

For reference, here are typical price levels for generative AI consulting inside Japan. Monthly advisory arrangements run JPY 100,000 to 300,000 per month, project-based engagements JPY 500,000 to 2,000,000, and PoCs JPY 400,000 to 10,000,000.

Please do not convert these figures into baht. The moment you convert, you misjudge. There are three reasons. First, delivery in Thailand adds tasks that do not exist in Japan — handling Thai-language documents, delivering training in Thai for local staff, and unifying terminology across two languages. Second, some parts are genuinely lighter than in Japan. The site is smaller and the number of departments in scope is limited, so the cost of building company-wide consensus is lower. Third, the supply side is structured differently. In Thailand, people who can understand manufacturing operations across Japanese, Thai and English are scarce, and that scarcity is priced into the rate.

So use Japanese price levels only as a shared vocabulary for discussing cost with head office. Build the Thai budget itself from the five-layer model above. When explaining to head office, framing it as “how much goes into which of the five layers” gets approved more easily than “cheap or expensive relative to Japan.”

How to write the internal approval request

Write the cost as a single total line and the approver judges on the size of that total alone. Write it in five layers and the discussion shifts to which layer to cut. That difference is large.

The recommended format is one line per layer from Layer 1 to Layer 5, stating the amount and what the company gets from it, and then stating explicitly that Layer 2 includes internal effort. Then state up front that first-year benefits will not accrue in full because of the ramp-up period. In the model below, the first-year benefit is taken at 50% and calculated as 324,436 baht. An approval request that claims full benefit from year one becomes impossible to defend twelve months later.

For how to construct the ROI case itself, measuring the impact of AI adoption and designing ROI sets out the procedure for designing the measurement indicators. Deciding how you will measure comes before writing the approval request.

Contract size has nothing to do with results

This is the part of the article we most want to convey. Lay Scenario A and Scenario B side by side over five years, and the sign flips.

Note that the five-year total benefit is the same 2,919,924 baht in both scenarios. That is not a typographical error. In this model, whether you start five processes at once or narrow to one, the benefit obtained over five years is identical. The reason follows below.

The result is a five-year ROI of -43.0% for Scenario A and +55.3% for Scenario B. Plus and minus swap places. Scenario A never pays back, even over five years. Scenario B pays back in about 2.6 years.

To be honest about a conclusion that works against us — the scenario a supplier like TOMAS TECH would prefer is the larger one, Scenario A. We still do not recommend it. A project that has not paid back after five years does not lead to a second contract.

How to build the three benefit pillars

Now the benefit breakdown. It is built from three pillars, each multiplied by a 70% realisation rate. Both the pre-multiplication and post-multiplication figures are shown.

Here is the design thinking behind the three pillars.

Labour reduction is the easiest to measure and the easiest to overstate. Of the 95 minutes spent per case, generative AI only touches the 40 minutes of document search and transcription. Judging whether the specification is sound, estimating cost, and setting terms in light of the customer’s situation all remain. So we do not write that 95 minutes halves. We write that 40 minutes becomes 16 minutes — a saving of 24 minutes per case. At 180 cases per month that is 4,320 minutes, or 72 hours. Multiplied by the 450 baht hourly rate that is 32,400 baht per month and 388,800 baht per year. Apply the 70% realisation rate and you get 272,160 baht. This discipline of restricting the saving to one part of the task is what determines whether the approval request is credible.

Avoided lost orders produces a large number, so the underlying assumptions have to be held tightly. With 2,160 inquiries per year, an average order value of 158,000 baht and a gross margin of 18%, gross profit per order is 28,440 baht. Assume 42 orders per year are lost because of slow responses, and treat only 14 of them — one third — as recoverable. We do not claim all of them can be won back. Even with faster responses, the orders lost on price and the orders lost on technical requirements remain. The result is 398,160 baht, or 278,712 baht after the realisation rate.

Faster ramp-up is the most frequently overlooked benefit. A new sales engineer used to take 6 months to run quotations independently, and being able to search past quotations in natural language shortens that to 4 months. With 2 hires per year, 2 months saved and a monthly productivity gap of 35,000 baht, that is 140,000 baht per year, or 98,000 baht after the realisation rate. Turnover at Thai sites is faster than at Japanese head offices, so this benefit compounds year after year.

The three pillars total 648,872 baht per year. The first year is taken at 50%, calculated as 324,436 baht. A 70% realisation rate may look conservative, but set against MIT’s 95% figure it is on the optimistic side.

Why widening the scope does not increase the benefit

Now the explanation of why Scenario A, which starts five processes at once, still produces only the single-process benefit of 2,919,924 baht. This is the heart of the article.

Widen the scope to five processes and all five have to pass through the same stages — preparing documents and data, unifying terminology, building the evaluation dataset, training the shop floor, and agreeing operating rules. Only the first half of that can be bought with money. The second half, the part where the shop floor actually keeps using the thing, involves a different owning department for each process, and each department moves according to its own priorities.

Five departments do not change the way they work in parallel inside a twelve-month contract. What actually happens is that the single most pressured department engages seriously, and the remaining four attend the training and report at the monthly meeting that they are still trialling it. That is the reality behind Deloitte’s 81%, and behind MIT’s 95%. The pilots are running. They are simply not moving the P&L.

As a result, the only thing that grows when you widen to five processes is the cost. Layer 2 preparation is incurred five times over, Layer 3 implementation is incurred five times over, and Layer 5 running cost becomes 580,000 baht per year for five processes. But only one process sticks, so only one process worth of benefit accrues. Over five years the cost swells to 5,120,000 baht while the benefit stops at 2,919,924 baht. Net, that is an ROI of -43.0%.

Understand this structure and you read proposals differently. A proposal for “cross-functional application to five processes at once” is not better because it is bigger. It means the force required to make things stick is five times larger, and unless the proposal says who supplies that fivefold force, it is simply a proposal that costs five times as much. The structural reasons why PoCs never reach production are covered from the same angle in why AI agent projects stall at the PoC stage.

Four reasons head-office-led contracts do not work at Thai sites

Take the generative AI consulting contract head office signed and roll it out to the Thai subsidiary as is. Here are the four reasons why that apparently rational decision does not function.

Reason 1 The language of the shop floor is Thai

At a Thai site, most of the documents involved in responding to a quotation — drawing annotations, internal specification memos, records of coordination with purchasing, past complaint handling records — are written in Thai. With 6 Japanese expatriates among 320 employees, the documents written in Japanese are only the parts an expatriate touched.

Head office’s generative AI design assumes internal documents are uniformly in Japanese. The prompt templates, the evaluation criteria and the training material are all in Japanese. Bring that to Thailand and you end up running a Japanese design philosophy through translated templates, and retrieval accuracy against Thai documents never improves.

The remedy is to build the evaluation dataset in Thai. The 30 evaluation questions built in Phase 1 below should be written in the language the shop floor actually uses, phrased the way the shop floor actually asks. Build them in Japanese and translate afterwards, and you get a system that passes evaluation in polished translated Thai and fails against the colloquial Thai people really type.

Reason 2 The business process differs from head office

Even for the same “quotation response,” the steps differ between head office and the Thai site. At head office, sales, engineering and cost control are separated, and each has information in its own system. At the Thai site, a small group of sales engineers handles the entire quotation process end to end, cost is managed in spreadsheets, and part of the judgement depends on an expatriate’s experience.

Where head office’s design says “pull automatically from the cost management system,” the Thai reality is a step where a staff member remembers a similar past case and opens a spreadsheet. Bring the head office design across without closing that gap, and implementation stops because the system you were supposed to integrate with does not exist.

For that reason, the Type A step of inventorying the current business process cannot be skipped at a Thai site. The most dangerous judgement is that it is unnecessary because head office already ran a diagnosis.

Reason 3 Documents are scattered across paper and local servers

At Thai manufacturing sites it is still common for approved drawings and inspection records to be kept on paper. Even the digitised documents are scattered across departmental file servers, individual PCs, shared drives and chat attachments.

To let generative AI search internal documents, you have to fix where the target documents live, sort out access rights, and remove duplicates and superseded versions. That work is Layer 2, business-side preparation, and even in Scenario B it accounts for 280,000 baht, or 28.0% of the total. Even if head office had already consolidated everything into SharePoint and barely needed this work, in Thailand it arises in full.

The reason horizontal rollout of the head office contract does not add up financially is mainly that this layer is missing from the estimate. Use head office’s rate card as it stands and the layer that consumes the most money and effort in Thailand goes unbudgeted.

Reason 4 The regulatory environment is different — Thailand has no AI law in force

Misinformation is circulating here, so let us be blunt. Thailand has no AI law enacted as of August 2026. The ETDA (Electronic Transactions Development Agency) is still refining its Draft Principles of AI Law, and enactment is expected to take two to three years. This follows the analysis in Baker & McKenzie’s “Thailand: 2026 AI Regulatory Landscape for Businesses” (March 2026).

Several secondary articles online state that a Thai AI law took effect on 1 March 2026. That is a confusion with Vietnam’s AI law, which did take effect in March 2026. Vietnam is in force. Thailand is at draft stage. Build internal policy while conflating the two and you spend effort complying with an obligation that does not exist while missing the regulation that actually applies.

What is actually in force in Thailand is the PDPA (Personal Data Protection Act) plus sector-specific regulation in areas such as finance and healthcare. So build your generative AI internal policy on PDPA compliance, not AI law compliance. Concretely, decide three things first — the scope of personal data that may be entered into generative AI, how consent is handled for data sent to external services, and log retention periods and the response to data subject access requests. Settle those three and any future AI law obligations can be layered on top.

If a proposal contains a line item for “Thai AI law compliance,” ask which statutory provision it refers to. If the firm cannot point to one, it has not researched Thai regulation. The approach to policy development including regulatory compliance is set out in how to write a generative AI usage policy.

Ten questions to ask before you sign

At the stage of issuing the request for quotation and reading proposals together, put these 10 questions directly to each firm. The point is not to eliminate firms that cannot answer, but to compare them on the substance of their answers.

Question 1 Which of the four types is this contract. Ask whether it is Type A, B, C or D, and if it is a blend, ask for the proportion of each type expressed in money. A proposal that returns a vague answer here will produce a disagreement over deliverables later.

Question 2 How many target processes are there. Is it narrowed to one, and if there are several, is the sequence for embedding them shown, together with the adoption criteria for each.

Question 3 How many person-days of buyer effort does Layer 2 business-side preparation require. A proposal that cannot answer in person-days is counting on that effort being free.

Question 4 How many evaluation questions will be created, and who creates them. A project where evaluation design is not in the contract ends without any way to measure the outcome.

Question 5 How will Thai-language business documents be handled. Translated, or handled natively in Thai. If translated, who checks the quality.

Question 6 How many integrations with existing systems are there. Ask for the names and the count of the systems to be integrated. This is the single largest source of change orders.

Question 7 What will we still have in month twelve. A working system, a policy, trained people. Ask the firm to articulate which of those three remains.

Question 8 What is the scope of the handover to in-house operation. Which parts of operations will we be able to run ourselves. Exit design is decided at contract time.

Question 9 Which statute does Thai regulatory compliance rest on. The correct answer is the PDPA. If the answer is “the Thai AI law,” ask for the basis.

Question 10 Under what circumstances would this contract fail to pay back. A firm that can state conditions unfavourable to itself understands the risk. Do not trust a firm that says there are no circumstances in which it fails to pay back.

Six common failure patterns

Failure 1 Going out for competitive quotations without deciding the type. A diagnostic-only quote and an implementation-inclusive quote sit side by side and the decision gets made on the price difference alone. This is the most frequent failure of all.

Failure 2 Rolling out the head office contract unchanged. For the four reasons in the previous section, the contract begins without Layer 2 costs budgeted, and a supplementary budget request becomes necessary partway through.

Failure 3 Not narrowing the target process. Widen to five processes and one still sticks. Only cost grows, and the five-year ROI comes out at -43.0%.

Failure 4 Starting without an evaluation dataset. With no shared yardstick for arguing whether accuracy is good or bad, the monthly meeting becomes a discussion of impressions, and three months later everything stops with a vague “it does not really work.”

Failure 5 Making the IT department the owner. Roughly 70% of companies are in this structure, as Deloitte noted. Put ownership in a function with no authority to change how work is done, and the deliverable begins and ends inside IT.

Failure 6 Writing the approval request at full benefit from year one. Write 648,872 baht per year without factoring in the ramp-up period, and the moment year-one actuals land at around 324,436 baht, the verdict is that it is not working and the year-two budget is cut. This is the most painful failure of the six.

A 90-day approach to getting started

Here is Scenario B’s route mapped onto the first 90 days. The goal in this period is not company-wide deployment. It is to reach a state where one process is running, is measurable, and is being used continuously by the people who do the work.

Phase 0 (weeks 1 to 2) Decide on one target process

List the candidates and narrow to one. There are three narrowing criteria — high volume, a measurable breakdown of working time, and an outcome expressible in money. Quotation response was chosen because it has the volume at 180 cases per month, because the 95 minutes per case decomposes into 40 minutes of document search and transcription, and because a lost order converts to a gross profit figure of 28,440 baht.

In these two weeks, measure the current state of the target process down to the breakdown of time. A process you have not measured cannot be shown to have improved even after it improves.

Phase 1 (weeks 3 to 6) Build the 30-question evaluation dataset first

Build the evaluation dataset before you build anything else. 30 questions. Collect 30 questions that sales engineers actually ask, in their actual phrasing, and settle internally on a model answer for each.

Do not reverse the order. Implement first and the evaluation gets written against the answers that implementation produces, which hollows it out. Build the evaluation first and the quality of the implementation can be judged objectively. Those 30 questions also remain usable as an asset if you later switch tools.

This is the hardest phase, because settling on model answers inevitably surfaces questions where opinions inside the company diverge. But that divergence is itself a finding — a problem that predates generative AI, now captured.

Phase 2 (weeks 7 to 10) Implementation and operating rules

Load the target documents and get search and draft generation working. In parallel, settle the operating rules. Who may use it, what must never be entered, who approves a generated draft. Fix the PDPA-based boundaries here.

Placing implementation and operating rules in the same phase is deliberate. Write the rules first and they become something nobody follows. Write them afterwards and you are writing them after an incident. Setting the rules while looking at something that works is the fastest route.

Phase 3 (weeks 11 to 13) Limited release and the improvement cycle

Release it to 5 sales engineers only. Do not release company-wide. The reason for limiting it to five is to get daily usage and to capture the cases that did not work immediately. Distribute it company-wide and you never recover the reasons it went unused.

Over these three weeks, re-measure against the 30-question evaluation dataset and record the prompts improved and the documents added. What you report on day 90 is not “we have introduced it” but “how many of the 30 evaluation questions are now answered correctly, and how many minutes were cut from each case.” Only at that point do you have the basis for deciding whether to widen to a second process.

Exit design — how to hand over to in-house operation

The thing least often designed into a generative AI consulting contract is the exit. Who operates it after the contract ends. Start without settling that and either operations stop the day the contract ends, or the contract can never be wound down and the monthly fee runs on.

There are four things to hand over — updating the evaluation dataset, adding target documents and removing superseded versions, revising prompts and operating rules, and first-line triage when accuracy degrades. Once those four can be run internally, external engagement can be scaled down.

Place the handover with the department that owns the target process, not with IT. The reason is simple — only people who know the work can judge the model answers in the evaluation dataset. IT’s remit stops at access rights and log management, and accuracy management belongs to the business side. The Deloitte figure of roughly 70% noted earlier, the share of companies treating AI as an IT responsibility, is a measure of how many companies get this placement wrong.

Fix the timing of the exit at contract time. For a Type C embedded support contract of 6 to 12 months, write into the contract that the final three months are a period in which the internal champion is the primary owner and the external party only reviews. An embedded support contract without that clause tends to mutate into an open-ended monthly retainer. The sequencing of insourcing, and the judgement of how much external support to retain, is set out in how to approach AI insourcing support.

To be honest, writing an exit design into the contract reduces the supplier’s revenue. It should still be written. A contract with no exit gets terminated abruptly in year three on the grounds that nobody can say what they are paying for. A design that scales down in stages tends, in the end, to produce a longer relationship.

Frequently asked questions about generative AI consulting

How much does generative AI consulting cost

The order of magnitude changes with the type you buy. Type A (assessment) runs 150,000 to 400,000 baht, Type B (governance) 250,000 to 600,000 baht, Type C (embedded support) 60,000 to 150,000 baht per month, and Type D (implementation included) 800,000 to 3,000,000 baht. Price levels inside Japan are JPY 100,000 to 300,000 per month for advisory arrangements, JPY 500,000 to 2,000,000 for project engagements, and JPY 400,000 to 10,000,000 for PoCs — but do not convert those into baht and call it a Thai budget. Thailand adds the step of preparing Thai-language documents, which does not exist in Japan, while the smaller site size reduces the effort of building consensus. Building up from the five-layer model is the correct method.

How far does an AI PoC need to go to be enough

It is enough at the point where you have measured accuracy against the 30-question evaluation dataset and confirmed by actual measurement that the working time of the target process has fallen. Put the other way, a PoC that has not confirmed that will never be enough no matter how many months it runs. MIT Project NANDA (July 2025) reported that 95% of generative AI pilots produce no measurable P&L impact precisely because most PoCs stop at “it worked” and never advance to being measured in time or money. Write the PoC’s pass criteria into the contract as numbers.

Should we look at custom-built AI or off-the-shelf tools first

Off-the-shelf tools first. The reason is that an off-the-shelf tool is sufficient to verify whether the target process really has been narrowed down. If the score on the 30-question evaluation dataset does not improve with an off-the-shelf tool, the cause is usually not the tool but Layer 2 document preparation. Proceed to custom development in that state and you end up searching unprepared documents with an expensive mechanism. Custom development becomes necessary when the number of integrations with existing systems grows, and when process-specific judgement logic has to be embedded.

How should we read case studies of enterprise generative AI adoption

Check three things. First, how many target processes there were. Where a case study says five processes at once, check how many of those actually stuck. Second, whether a realisation rate has been applied to the benefits. This article applies 70%, and 50% in the first year. Figures with no realisation rate applied are upper bounds. Third, whether the cost includes business-side preparation, Layer 2. A total that excludes it represents only about seventy percent of what will actually be spent.

What is the difference between a consultancy and an AI development firm

The types they are good at. Consultancies live in Type A and Type B, and are strong at selecting the target process and building rules. AI development firms live in Type D and are strong at building working things. Many system integrators can take on both Type D work where integration with existing systems dominates and Type C operational support. Do not choose by category of firm. Decide which type you should be buying, then engage a firm whose core business is that type.

Do we need to comply with a Thai AI law

As of August 2026 there is no Thai AI law to comply with. The ETDA is still refining its Draft Principles of AI Law, and enactment is expected to take two to three years (Baker & McKenzie, March 2026). Some articles claim a Thai AI law took effect on 1 March 2026, but that is a confusion with Vietnam’s AI law, which took effect in March 2026. What is actually in force in Thailand is the PDPA and sector-specific regulation, so build internal policy on the PDPA.

Summary

When considering generative AI consulting, the first thing to decide is not who to hire but which type of contract to buy. Type A (diagnosis), Type B (boundaries), Type C (adoption) and Type D (implementation) differ in deliverables, duration, cost and in the way each one fails. Go out for competitive quotations without deciding the type and you get proposals that cannot be compared.

Break the cost into five layers. The largest layer is not implementation but Layer 2, business-side preparation, at 29.3% in Scenario A and 28.0% in Scenario B. Settle before signing how many person-days of buyer effort that layer requires.

And the most important conclusion is that contract size has nothing to do with results. The 2,800,000 baht contract that starts five processes at once and the 1,000,000 baht contract narrowed to one process both produce the same five-year benefit of 2,919,924 baht, because in the end only one process sticks. As a result the five-year ROI flips sign — +55.3% for the narrowed approach and -43.0% for the company-wide approach, which never pays back even over five years.

At a Thai site, the contract head office signed does not work as it stands. The language of the shop floor is Thai, the business process differs from head office, documents are scattered across paper and local servers, and the regulatory environment is different. On regulation, the correct starting point is to recognise that Thailand has no AI law in force and to build policy on the PDPA.

There is one thing to do in the first 90 days. Decide on one target process, build the 30-question evaluation dataset first, and run it with 5 sales engineers while measuring. That is all.

TOMAS TECH delivers PEGASUS production and energy management systems along with factory automation, control and IoT projects for Japanese-owned manufacturers in Thailand, and as an extension of that work we support the selection of target processes for generative AI, the design of evaluation, and adoption on the ground here. We are happy to start from working through which layers your own operations will consume cost in, even if you have not yet settled on a contract type. A conversation about which parts of the five-layer model apply to your site, or a second opinion on how to read a proposal you have received, is equally welcome. Enquiries are received at https://tomastc.com/en/contact/

References

• Deloitte Thailand Digital Transformation Survey 2026 coverage (Kaohoon International, August 2026): https://www.kaohooninternational.com/technology/588138
• UOB Business Outlook Study 2026 on AI adoption among Thai SMEs (ThaiPR.NET, June 2026): https://www.thaipr.net/en/business_en/3735803
• Gartner forecast of worldwide AI spending in 2026 (Gartner, May 2026): https://www.gartner.com/en/newsroom/press-releases/2026-05-19-gartner-forecasts-worldwide-ai-spending-to-grow-47-percent-in-2026
• Gartner forecast of worldwide IT spending in 2026 (CIO, February 2026): https://www.cio.com/article/4126847/gartner-it-spending-to-exceed-6-trillion-by-2026.html
• MIT Project NANDA finds 95% of generative AI pilots fail (Forbes, August 2025): https://www.forbes.com/sites/jasonsnyder/2025/08/26/mit-finds-95-of-genai-pilots-fail-because-companies-avoid-friction/
• Thailand 2026 AI Regulatory Landscape for Businesses (Baker & McKenzie, March 2026): https://www.bakermckenzie.com/-/media/files/insight/publications/2026/03/thailand-2026-ai-regulatory-landscape-for-businesses.pdf
• Price benchmarks for AI adoption consulting in Japan (Revival Asia, 2026): https://www.revivalasia.co.jp/column/ai-consulting-cost