AI Insourcing Support — Deciding Which of the Five Layers You Own

When we are asked about AI insourcing support, the first thing we ask back is a single question. What exactly do you want to bring in-house? In most cases the answer is “AI itself.” And that is where the discussion stalls. In-house versus outsourced is not a two-way choice. The work of putting AI into daily operations splits into five layers, and the real question is who holds each layer. What finally decides which layers you can hold yourself is neither technical skill nor budget. It is a far more mundane condition. Can you hire the people who staff that layer, and will they stay?

This article is written for the people who run manufacturing sites in Thailand and the wider ASEAN region, including local management, administrative heads and IT staff at Japanese-owned plants, who have tried generative AI once, watched it fail to take root, and are now unable to decide how much their own organization should own. We will work through what published surveys in Japan show about the state of AI talent, set that against the numbers from the Thai employment market, and use both to design where the line falls layer by layer. Every figure carries its source. Where a number is our own recommendation or an assumption used for illustration, we say so.

Before You Look for AI Insourcing Support — In-House Versus Outsourced Is the Wrong Question

Let us start with the conclusion. The question “should we build AI in-house or outsource it” has no answer, because the question is framed too coarsely.

Putting AI into daily operations is really five separate kinds of work.

First, deciding which problem in which operation AI should solve. Second, gathering and preparing the data that this requires from inside the company. Third, choosing the model or tool and building something that runs. Fourth, embedding it into the actual work procedures so that it gets used. Fifth, keeping it running and improving it.

These five layers demand different skills, and they differ enormously in how far an outsider can stand in for you. Layer 3 has suppliers all over the world, and once the specification is fixed it can be sent out. Layers 1 and 4, on the other hand, can only be done by someone who knows your operations and the human relationships on your floor. However capable your vendor is, they cannot judge for you whether the daily report entry on this process is genuinely painful for the operators, or who reads this inspection record and for what purpose.

So the correct question becomes this. Of the five layers, which do we hold ourselves, which do we send outside, and which do we hold jointly? And the answer differs from company to company. It shifts with the size of the plant, the state of the existing systems, the headcount in the IT function, and above all the depth of the talent pool you can recruit from at your Thailand site.

If you are searching for an outside partner using the phrase “AI insourcing support,” you have already given up on holding everything yourself. In that case the first decision is not which support company to pick. It is where to draw the line. When support is engaged before the line has been drawn, the supporting side can only deliver Layer 3, the building work, and the engagement lands on the single most common ending of all. Something runs, and nobody uses it.

Below we break down how to design that line, one step at a time. We start with the underlying reality of the talent supply.

Setting the Baseline — Even in Japan the Supply of AI and DX People Falls Short

Before we get to the Thailand site, it is worth checking the situation in Japan. This matters because a shaky picture here breaks a very common design assumption, namely that the parent company holds the capability and the Thai site simply consumes it.

According to the FY2025 edition of the DX Trend Survey published by the Information-technology Promotion Agency, Japan (IPA) on July 16, 2026, responses of “somewhat short” and “severely short” regarding the quantity of people driving DX together reached 85.5%. The survey covered executives, IT departments and DX promotion departments at companies in Japan, ran from April 17 to June 12, 2026, and drew responses from 1,799 companies.

The important part is that this 85.5% sits at roughly the same level as FY2024 and FY2023. In other words, three years with no improvement. Over the years in which generative AI worked its way into daily work and tools became cheaper and easier to use, the perceived shortage of people has barely shifted.

The same survey also shows the gap in AI adoption by company size.

That 16.6% is the adoption rate among companies with 101 employees or fewer. It is not a figure for small and medium enterprises as a whole. What is certain is that a substantial gap separates them from the largest companies. And many Japanese-owned plants in Thailand belong to a parent with more than 1,001 employees while the local entity itself runs from a few dozen to a few hundred people. Looked at as a standalone entity, the local company sits closer to the second environment, and treating it that way matches reality better.

What follows from this is that the assumption “the parent company in Japan knows AI, so Thailand can just be taught” is, with fair probability, not going to hold. The parent’s IT department is short of people too. And the parent’s priorities point first at its domestic plants and its domestic core systems. The pattern in which AI at the Thai site stalls waiting for head office support follows naturally from these shortage figures.

The same survey also records where AI sits within DX overall. “AI used as part of DX promotion” and “AI-centered use for the sake of DX promotion” together come to 54.2%. Meanwhile, companies answering that “DX and AI are pursued separately” account for a quarter of the companies engaged in DX. AI and DX running as two separate projects inside one company is by no means unusual.

What Is Missing Is Not People Who Can Build AI — This Is the Heart of Insourcing

This is the core of the article.

The same IPA survey reports the sufficiency of AI-related talent broken down by type of person. While “in short supply” makes up a majority for many of the types, the direction of travel clearly splits by type.

Read that table again, slowly.

What is easing is the shortage of “employees who can apply AI tools to their own business.” People who can operate the tools, write prompts and use generative AI in daily work are becoming more numerous inside companies. Understanding among executives has, relatively speaking, advanced as well.

What is not easing is the shortage of “employees with shop-floor knowledge plus basic AI knowledge who can drive AI adoption in their own company” and of “people with data management knowledge who can plan and drive it internally.”

The gap between those two explains the entire insourcing debate.

What is missing is not people who can build AI. What is missing is people who know your operations, who also know what AI can do, and who can connect the two. The first half alone you can hire, in Thailand or in Japan. The second half alone you can outsource. But the work of joining the two inside one person’s head is something an outsider cannot do in principle, because learning your operations takes time.

Put differently, the complaint “we cannot make generative AI work for us” is, in most cases, not a problem of knowing how to use the tool. The number of people who know how to use the tool has grown. What that person lacks is the authority and the information to decide how to point it at, say, the transcription work in your incoming inspection records. Meanwhile the section chief who knows that work inside out does not know what AI can do. This disconnect is what a stalled proof of concept actually consists of.

The Corporate IT Utilization Trend Survey 2026, conducted by the Japan Institute for Promotion of Digital Economy and Community (JIPDEC) in January 2026, likewise cites “a shortage of people or skills able to develop and operate AI” as the largest barrier before adoption. In that survey, companies putting AI into practical use stopped at 36%. Note that this 36% comes from a different organization, a different period and a different respondent base than the IPA survey, so it cannot be lined up beside the IPA figures in the previous section for comparison. The correct reading is that two separate surveys, each from its own angle, are saying that people are in short supply.

Take this talent structure as given and the design principle for insourcing follows automatically. The layers that are hardest to hire for from outside are exactly the layers you have no choice but to grow internally. And what can be bought outside, you buy. The five-layer line converges on that unremarkable conclusion.

Where AI Is Working and Where It Is Not — Work Backwards to the Layers You Should Own

Which layers you should own can be derived backwards from where AI is actually working today. The IPA survey tabulates both the uses AI is put to and the concrete effects obtained from adoption.

First the uses. All of the figures below come from multiple-response tabulations, so they do not add to 100%. Do not add the uses together.

The top three are all jobs that handle words. Summarize, write, find. What sinks to the bottom is the work of deciding how things move. Planning support for production and logistics stops at 6.0%.

Next, the effects actually obtained. These are multiple responses as well.

Lay those two tables over each other and the outline of AI use today comes into focus. What is working is efficiency and speed in daily work (91.6%), while contribution to revenue or profit (3.9%) is barely reported at all. Companies citing a reduction in overtime hours came to 29.2%, considerably fewer than those citing efficiency gains. Things got faster, but not yet far enough for the gain to be booked as hours saved.

On overall assessment of the effect of AI adoption, companies choosing “beyond expectations” or “as expected” stopped at 31.8%, while the middling verdict “there was a certain effect” accounted for 50.6%. A large number of companies sit in the position of not disowning it, yet not reaching what they hoped for.

Now back to the insourcing line. The uses at the top, summarizing, writing and searching, get a long way on general-purpose tools. Honestly, the case for assigning your own engineers to build something bespoke there is weak. Comparing commercial tools and picking one is faster, cheaper and more consistent in quality. We set out how to think about tool selection in our comparison of enterprise generative AI.

Conversely, the uses at the bottom, planning support for production and logistics and enhancement of your own products and services, are territory that general-purpose tools do not reach. The reason is plain. The data they need exists only inside your company, and it is not in order. The JIPDEC survey likewise cites “a large volume of information that has not been turned into data for AI to learn from” as a problem that remains after adoption.

So it comes to this. Send outside the problems that outside strength can solve, and put your own people on the problems that need data and operational knowledge only you have. That is the conclusion of working backwards.

Splitting AI Insourcing into Five Layers — What to Own and What to Send Out

Now we turn all of the above into an actual line. We split the work of putting AI into operations into five layers and decide, for each, where it is held.

Of the five layers, two cannot be outsourced (Layers 1 and 4), one is easy to send out (Layer 3), and two are hybrid (Layers 2 and 5). That accounts for all five.

The talent type at the top of the shortage list in the IPA survey we looked at earlier, “employees with shop-floor knowledge plus basic AI knowledge who can drive AI adoption in their own company,” is precisely the person who staffs Layers 1 and 4. These two layers alone cannot be filled by outsourcing, however much budget you stack on them. That is the backbone of this article.

Layer 1 — Problem Definition and Prioritization Cannot Be Outsourced

What you do in Layer 1 is not about AI. It is an inventory of the work. In which process, by whom, for how many hours a month, doing what. And of those, which are painful for the person doing them, and which are a loss for the company. Those two do not always coincide.

For example, and all of the numbers here are assumptions used to make the point, suppose a production control staff member spends an hour every morning compiling daily reports. It is painful for that person, but what the company is losing is twenty hours of labor a month. Now suppose instead that a shipping decision slips by a day and inventory sits. Nobody feels any pain, yet the company takes a loss. These two are measured in different units, labor hours and inventory, so which is larger cannot be known until you put your own numbers against them and compare. Deciding the order means actually doing that comparison.

Deciding which to solve first is beyond an outsider, because most of the information the judgment needs is tacit knowledge held inside the company. Hand this over wholesale to a vendor and the vendor will propose “the themes that generally produce results.” That is not wrong, but whether it is best for you is another matter entirely.

Layer 2 — Data Preparation Is Held as a Hybrid

Layer 2 takes the most effort and gets the least respect.

For AI to judge anything, the material for the judgment has to be structured. But information in a plant is scattered across paper, Excel, the production control system, equipment logs and people’s memories. The problem the JIPDEC survey points to, “a large volume of information that has not been turned into data,” is pointing exactly here.

This layer can be split. Only your own people know which data sits where inside the company, so identifying locations and judging access rights stays in-house. The hands-on work of extracting, reconciling and shaping, on the other hand, can go outside. If your single IT staff member absorbs all of this alone, everything else that person handles comes to a stop.

Note that pulling data out of an existing core system is less an AI project than a system integration project. We set out how to think about the cost of that in our article on business system development and ERP integration. Estimating it in a budget line separate from AI saves you a surprise later.

Layer 3 — Model and Tool Selection and Build Is the Easiest to Send Out

Layer 3 is the most “AI-like” of the five layers and, at the same time, the easiest to send out.

The reason is simple. The work in this layer can be reduced to a specification. The input is this, the output is that, and these conditions must hold. Once that can be written down, there are people outside who can implement it. Turn that around and it says something else. If you outsource this layer while the specification is not yet written, the outsourcing of this layer will fail without exception, because writing the specification is the output of Layers 1 and 2.

Insist too hard on insourcing here and a familiar sequence unfolds. You hire one junior engineer, put that person in charge of building the model, something runs, the person moves to another employer two years later, and nobody understands the internals. A decision to hold Layer 3 in-house is meaningless unless it is paired with the question of whether you can sustain that arrangement for five years.

Layer 4 — Embedding into Operations Cannot Be Outsourced

Layer 4, alongside Layer 1, cannot be outsourced. It is also the layer where failures occur most often.

What you do after something runs is not technical. Rewriting the procedure document. Deciding whose approval gets dropped. Explaining to the person who used to do the work by hand how their job now changes. Getting the section chief on the floor to accept it.

This work can only be done by someone whose face and name are known inside the company. An outside consultant who walks onto the floor and says “please do it this way from tomorrow” will not move anything. In a plant in Thailand this is even more pronounced, because it only advances on the trust between Japanese managers and Thai staff.

Layer 5 — Run and Improve, With Monitoring Outside and Judgment Inside

Layer 5 is the layer that starts once things are running. Has accuracy slipped, is it being used, and if it is not, why not.

Monitoring and first-line response can go outside, because reading logs, detecting anomalies and reporting them can be reduced to a routine. But the judgment of whether a given drop in accuracy is tolerable or worth investing to fix has to stay inside. Hand that judgment outside and the vendor will keep saying it would be better to fix it. That is their job.

On how to measure effect, our article on measuring the effect of AI adoption and ROI covers measurement design. When you design Layer 5, settling the measurement items there before you assemble the operating routine avoids ending up unable to explain what the effect was.

In-House AI and a Secure Generative AI Environment — How Much of Layer 3 to Hold Yourself

Let us go a little deeper into Layer 3. The phrase “building in-house AI” in fact covers a range.

The higher up the table, the lighter the lift. The lower down, the tighter the control. And the lower down, the more it hurts when someone leaves.

A request for “a secure generative AI environment” tends to be received as meaning the bottom option. But what is actually needed is, in most cases, that confidential information is not used to train an outside model, and there are cases where the middle configuration satisfies that through contract terms. Identifying which level of control is genuinely required, against your legal requirements, your customer contract terms and your parent company’s rules, comes first.

In the JIPDEC survey, 35.3% of respondents cited “securing final human judgment and having a structure that can explain the basis and reasoning behind AI output” as a problem. This is not an infrastructure matter. It is an operating rule matter. Before debating where the model sits, it is necessary to settle who approves an answer produced by AI, and against what criteria. If you build only the environment while that decision is still missing, what emerges looks controlled while in fact nobody is able to take responsibility for it.

On the configuration that adds internal documents to what is searched, our article on building RAG over factory knowledge covers the practical data preparation required. It gives you material for estimating how heavy the Layer 2 load will be.

The Assumption That Breaks at a Thailand Site — The People in Layers 1 and 4 Will Turn Over

Everything up to this point holds equally in Japan and in Thailand. But a Thailand site carries one condition of its own. The people who staff Layers 1 and 4 turn over on a regular cycle.

There are two reasons.

One is the rotation of expatriates. Japanese managers change every few years. The person who best understands what actually hurts in this plant, and who can build agreement on the floor, finishes their posting and returns to Japan. The next person starts from learning the operations from zero. The tacit knowledge that Layers 1 and 4 require is reset each time.

The other is local staff changing employers. On the Thai employment market, there is a survey that sets out the outlook for 2026.

This table needs care in the reading. The 4.7% salary increase budget is an annual budget rate. The 15% to 30% figure is the step up in level that occurs at the moment a person moves employer. Because they measure different things over different periods, you cannot divide one by the other and say pay rises some number of times faster. We will be straight about this. The source carries no information that would let these two be converted into each other.

What can be said is only that a structure exists in which pay moves differently for those who stay and for those who move outside. And the scarcer the skill, the wider that difference. Work that touches AI falls squarely into the scarce-skill category.

Note as well that the 17.5% attrition rate is given in the source as a regional figure, and it is not stated as attrition for Thailand on its own. It cannot be asserted as a Thai number. Still, there is value in understanding that this is a market with movement at that level.

Thailand’s economic growth outlook for 2026 is about 1.6%. It is hard to assume the business as a whole grows sharply, while pay for talent is the part that moves readily. That said, the party connecting these two points is us, not the source, which does not present that analysis. Speaking as our own interpretation, this situation is what makes it hard to assemble AI capability at a Thailand site through recruitment.

On the practical differences specific to adopting AI in Thailand, our article on AI adoption in Thailand covers language and data handling among other topics.

Conversion Rather Than Recruitment

Securing talent is not only a matter of recruitment. There is another route, converting the people you already have.

According to an article published by The Nation Thailand on July 17, 2026, the major Vietnamese IT company FPT Corporation converted 30,000 conventional software engineers into “AI-augmented” talent, as stated by the company’s CEO Levi Nguyen. The company is described as having 170,000 students through various university and training frameworks.

That said, this is a matter of an entirely different scale. What an IT company with tens of thousands of engineers can do differs from what a Japanese-owned plant with a few dozen to a few hundred people can do. A Japanese-owned plant is not going to carry out a conversion on the same scale. We cite it here for one reason only. It illustrates a structure, namely that securing AI talent has a route other than hiring from outside, which is converting the people you already have.

The same article also notes that the Economic Research Institute for ASEAN and East Asia (ERIA) points to severe talent shortages and mismatches slowing AI adoption across ASEAN. Talent scarcity is not confined to Thailand. It is a structural feature of the region. Waiting until the recruitment market improves is therefore not a very realistic option.

A Design That Does Not Depend on Individuals Is the Precondition for Insourcing — What Remains When People Leave

The conclusion of the previous section can be summarized as follows. At a Thailand site, the turnover of the people staffing Layers 1 and 4 has to be designed for as a premise, not treated as an exception.

This is not a reason to give up on insourcing. It is a reason to change the design. Concretely, it means building “not held in anyone’s head” into the arrangement from the start.

If your reaction to that table is that it is all obvious, that reaction is correct. There is nothing special in it. But on real sites, this obvious material is in most cases not there.

The reason is understandable. AI adoption is usually driven by one particular enthusiastic individual. That person moves fast. To move fast, they postpone the record-keeping. And by the time results arrive, there is no time left to write it up.

Guarding against dependence on individuals therefore has to be secured by procedure rather than by good intentions. Here are the three we recommend. The headcounts and frequencies below are not benchmarks drawn from a source; they are our own recommended values.

First, state in the purchase order to the vendor that documents are included as deliverables. That is, do not make a working system the only thing subject to delivery. Second, put two internal drivers on it from the beginning. With one, everything stops the moment that person leaves. Third, fix a day once a quarter to check whether the handover material still matches reality. If you do not fix it, it does not happen.

Holding something in-house means keeping the knowledge inside the company. If it walks out with the person, then it was never held in-house. This is the single biggest trap in insourcing at a Thailand site.

What to Pin Down When You Ask for Hands-On AI Support — Scope and the Point of Release

The phrase “hands-on AI support” has become common over the past few years. It is not a bad phrase, but it is vague. The word says nothing about how long the support runs alongside you or where it lets go.

When you engage hands-on support, here is what to pin down in the contract or the agreement.

The most important of these is the third row, the conditions for letting go.

The purpose of hands-on support is for you to reach the point of running under your own power. So if the ending is not defined at the time of contract, what you have is not hands-on support but plain continuing outsourcing. Continuing outsourcing is not a bad thing in itself. Layer 3, and the monitoring portion of Layer 5, are in fact more stable when placed outside on a continuing basis. The problem is when something contracted for the purpose of insourcing quietly turns into continuing outsourcing and nobody notices.

There is one trick to writing the conditions for letting go. Write them as actions, not as deliverables. Not “documents are delivered” but “our own assigned person can take one new operation from Layer 1 through Layer 5 without help from the supporting side.” Written as an action, whether it has been achieved is obvious to everyone.

There is one more thing the requesting side has to supply. The time of your own assigned people. Hands-on support does not work if there is nobody on your side to receive it. A request along the lines of “we are busy, so please handle all of it” is, by definition, not hands-on support. Fail to pin this down at the outset and the supporting side will, out of goodwill, shoulder the load in your place, and the engagement ends with nothing left behind.

Three Classic Ways AI Insourcing Fails

Building on everything above, here are three classic failure patterns for insourcing. Each of them arises from skipping one of the five layers.

Failure 1 — Hiring Someone First and Working Out What They Do Later

The sequence goes “we hired an AI person, now what shall we have them do.”

It looks like a forward-leaning investment, but it skips Layer 1. What does someone hired while the problem is still undecided do first? They go looking for a problem themselves. But they have just joined and do not know the operations. The result is that they land on an easy-to-find problem, that is, a theme generally held to produce results. And there is no guarantee that this coincides with where you actually hurt.

This pattern has a secondary problem as well. The person hired becomes isolated without producing results. Those around them treat them as “the AI person” but do not teach them the operations. As we saw in the previous section, the Thai market has a structure in which high-skill people move readily. An isolated hire is not going to stay.

The correct order is the reverse. Narrow down the problem in Layer 1, and once the required role is clear, hire for that role or assign someone already on staff to it.

Failure 2 — Outsourcing a PoC and Stopping There

In the spirit of trying something small first, a proof of concept is commissioned from a vendor. Three months later a working demo exists. It is shown in a meeting. The verdict is that it is well made. And there it stops.

This one skips Layer 4. The success condition for a PoC is that it runs. The success condition for production is that it gets used. Those are two different jobs. And the contracted scope of a PoC usually does not include Layer 4.

Outsourcing a PoC is not in itself a bad thing. The problem is that at the point the PoC ends, who does what next has not been settled. At the time you commission the PoC, settle who carries it onto the floor if it succeeds and where that person’s time is going to come from. That alone goes a long way toward preventing this pattern.

Failure 3 — Handing Out Tools and Leaving the Usage to the Floor

Generative AI licenses are distributed to all employees. One training session is held. After that it is left to whatever the floor comes up with.

This pattern is not rare either. Recall that in the IPA survey, the shortage share for “employees who can apply AI tools to their own business” had fallen relative to the other types. The number of people who can operate the tools is growing. So distribute licenses and there will be some level of use. It lands on the top-ranked uses, summarizing, translating and writing documents.

But it goes no further than that, because nobody is assigned to the work of embedding it into the operating procedure, which is Layer 4. Individual productivity rises a little, while the organization’s business processes do not change. Nor can the effect be measured. It ends at a vague sense that things got more convenient.

In the IPA data from the earlier section, 91.6% answered that work became more efficient and faster while only 29.2% cited a reduction in overtime hours, and that can be read as suggesting how widespread this pattern is. Things got faster, but the gain has not been recovered as time. Recovering it requires rebuilding the procedure itself.

Designing the Stages — Where to Start With Generative AI

Now we can assemble an answer to the question of where to start with generative AI out of everything above. What we recommend is the following four stages.

Of the four stages, Stages 0 and 2 are led by you. Stage 1 lends itself to using outside strength, and in Stage 3 the hands doing the measurement can go outside while the evaluation and the selection of the next theme stay with you. As the mapping to the layers shows, these four stages are designed to go once around all five layers in order.

The important thing is not to skip Stage 0. Skip it and you fall into both Failure 1 and Failure 2 from the previous section. And Stage 0 cannot be outsourced. What an outside supporter can do is show you how to run the inventory, ask questions, and supply other companies’ cases as raw material. Producing the answer is yours.

Turning the Effect into Money

When you measure the effect in Stage 3, you may be asked to convert it into money. Here are the points to watch.

First, fix a single baseline. Define one “state before adoption” and measure every effect as a difference from it. Start calculating without declaring this and the numbers will fail to reconcile later, without exception.

Second, for the same piece of work, do not stack a reduction in internal labor and a reduction in vendor fees on top of each other. If your own people used to do a piece of work, what you can cut is internal labor. If it was outsourced, what you can cut is the vendor fee. Not both. An estimate that adds the two together is booking an effect that does not exist. This is a mistake we see often in practice.

Third, this article does not give absolute money figures. The sources referenced carry no figures for cost savings from AI adoption. What can be written is the procedure only. Put in your own labor rate, the monthly volume of the target work and the time per case, all measured in your own operation, and calculate it yourself. Applying an amount published in another company’s case to your own has almost no meaning, because the premises differ.

On designing the measurement items, our article on measuring the effect of AI adoption and ROI sets out how to count, and for the case where forms and documents are the target, our comparison of AI-OCR sets out how to measure in practice.

Frequently Asked Questions

Where should we start with generative AI?

Start with an inventory of the work, not with tool selection. This corresponds to Stage 0 above. Concretely, list by department the work that eats time while adding little value, and write out the monthly volume and the time per case. As a rough guide of ours, spending two to four weeks on this is perfectly acceptable. If you can narrow the candidates to three or fewer here, every judgment after that gets much faster.

Why can we not make generative AI work for us?

In most cases the reason is not how to use the tool. In the IPA FY2025 survey, the shortage share for “employees who can apply AI tools to their own business” has fallen compared with the other talent types. In other words, the number of people who can use it is growing. What has not eased is the shortage of “employees with shop-floor knowledge plus basic AI knowledge who can drive AI adoption in their own company.” The bottleneck, then, is not operating skill but the absence of anyone inside the company placed in the role of connecting the operations to AI.

Who should we talk to about adopting AI?

Before choosing who to talk to, decide which of the five layers you want to talk about. For Layer 3, the build, a system development company or systems integrator suits. For Layer 1, problem definition, you need a counterpart who knows a domain close to your own operations. For a plant, a partner who understands work on the production floor moves the conversation along faster. Conversely, consulting an outside party about Layer 4, making it stick, will not get it done for you. That one proceeds on the premise that you carry it, taking advice on how to go about it.

Can smaller companies make use of generative AI?

In the IPA FY2025 survey, close to 80% of companies with 1,001 employees or more had adopted AI, while the adoption rate among companies with 101 employees or fewer was 16.6%. The gap is certainly there. But that figure is about whether AI has been adopted, not about whether results followed. Being small actually works in your favor in Layers 1 and 4, because who does what is visible and there are fewer people whose agreement a procedure change requires. Where it works against you is Layer 3 and the monitoring portion of Layer 5, and covering those with outside strength is the realistic approach.

Should we hire AI talent or grow it?

The harder a layer is to hire for from outside, the more you have no choice but to grow it inside. Engineers to staff Layer 3 may well be available on the market. But the people who staff Layers 1 and 4, meaning people who know your operations intimately and also understand the basics of AI, are not on the market. Knowing your operations intimately requires time spent working at your company. So the realistic answer is to select and develop people for Layers 1 and 4 from among existing employees, and to fill Layer 3 by hiring or by outsourcing. That converting existing people is a viable route can also be read from the ASEAN example above.

How secure does a generative AI environment need to be?

The level of control required follows from your legal requirements, your customer contract terms and your parent company’s rules. As technical options there is a progression from using a commercial SaaS as it comes, to connecting internal data to it, to placing the model in your own environment, and the further down that progression you go the more permanent operations staff you need. In many cases what you actually want to protect is that the information you enter is not used to train an outside model, and there are cases where contract terms satisfy that. Put into words what you want to protect before starting the debate about the environment.

How does hands-on AI support differ from ordinary outsourcing?

The difference is whether the ending is defined. Hands-on support ends at the point you can run under your own power. So the contract needs to state what you must be able to do for it to be over. Hands-on support without that written into it is continuing outsourcing in substance. Continuing outsourcing is not a bad thing, but if the budget was taken for the purpose of insourcing, then the purpose and the means have come apart.

AI talent does not stay at our Thailand site. What can we do?

Start by rebuilding the design on the premise that they will not stay. A survey of the Thai employment market for 2026 puts the average salary increase budget at about 4.7%, and the pay uplift on changing jobs in high-skill fields at 15% to 30%. These two measure different things and cannot be converted into each other, but a structure in which pay rises more for those who move can be read from them. The regional attrition rate is additionally given as about 17.5%. On the premise that people move, keeping the list of problems, the data definitions, the judgment criteria and the procedure documents as documents becomes the precondition for insourcing.

What should we tackle first to make the internal case easier?

Starting with work whose effect is easy to measure and that involves few stakeholders is the standard move. In the IPA survey, “work became more efficient and faster” dominates the concrete effects of AI adoption at 91.6%, while “revenue or profit improved” stops at 3.9%. What AI today is good at is, first of all, making work more efficient. Set a contribution to revenue as the first target and you tend to end up with a result you cannot explain. Build a track record on efficiency, establish the pattern for measurement, and then move on to the harder themes.

Summary — AI Insourcing Support Means Deciding the Boundaries Together, Layer by Layer

This article makes one claim. In-house versus outsourced is not a two-way choice. It is a question of deciding, layer by layer, where each is held.

The work of putting AI into operations splits into five layers, problem definition, data preparation, build, embedding into operations, and run and improve. Of these, Layer 1, problem definition, and Layer 4, embedding into operations, cannot be outsourced, because they can only be done by someone who knows your operations and the human relationships on your floor. Layer 3, the build, can be sent out as long as the specification can be written. For Layer 2, data preparation, and Layer 5, run and improve, the realistic answer is a hybrid that keeps the records and the judgment inside while placing the hands and the monitoring outside.

This line is consistent with the results of the survey IPA published in July 2026. In that survey, responses reporting a shortage in the quantity of people driving DX reached 85.5% and stayed roughly flat over three years. And broken down by talent type, the shortage share for “employees who can apply AI tools to their own business” has fallen relatively, while the shortage of “employees with shop-floor knowledge plus basic AI knowledge who can drive AI adoption in their own company” has not. What is missing is not people who can build AI. It is people who can connect your operations to AI. That layer cannot be bought from outside.

And a Thailand site adds one more condition. The people staffing Layers 1 and 4 turn over on a regular cycle, through expatriate rotation and through local staff changing employers. On the Thai employment market for 2026, the average salary increase budget is put at about 4.7%, the pay uplift on changing jobs in high-skill fields at 15% to 30%, and the regional attrition rate at about 17.5%. People move. So the knowledge you decided to hold in-house must not be left in people’s heads alone. Keep the basis for the problem list, the data definitions, the judgment criteria and the procedure documents as documents. That is the precondition for insourcing at a Thailand site.

The order of what to do is equally clear. Inventory the work and narrow the problems to three or fewer, pick one and build small, rewrite the procedures and put it on the floor, measure the effect and extend to the next. Of these four stages, Stages 0 and 2 are led by you. The moment you skip them, you fall into one of the three failures, hiring first and thinking later, stopping at a PoC, or handing out tools alone.

What you should expect from a service called AI insourcing support is not that it builds the thing for you. It is that it decides with you which layers you hold, gets your organization able to carry the layers you decided to hold, and then lets go. Support with no written condition for letting go is not insourcing support.

TOMAS TECH is based in Bangkok and builds production management and energy management systems for manufacturers across Thailand and ASEAN, and on AI we take inquiries from the stage before anything gets built, meaning the work of sorting out where the line falls, which operation to start from and which layers to hold yourself. It is fine if no concrete theme has been settled yet, and equally fine if you have run an inventory internally but cannot settle the priority order. For a conversation about designs built on the assumption of long operation on the ground, please get in touch through our contact form.

Reference Material

• Key Points on DX and AI Adoption Trends at Companies in Japan (IPA, July 16, 2026) — DX Trend Survey FY2025. 1,799 responses, survey period April 17 to June 12, 2026. Source of the talent, use case and effect figures in this article
• Corporate IT Utilization Trend Survey 2026 (JIPDEC) — Conducted January 2026. State of practical AI use and the obstacles before and after adoption
• Thailand Job Market 2026 Outlook — The Great Digital Divide (RECRUITdee) — Thailand salary increase budget for 2026, pay uplift on changing jobs, regional attrition rate, economic growth outlook
• Bridging the Tech Rift — How Southeast Asia’s Giants are Forcing an AI Talent Revolution (The Nation Thailand, July 17, 2026) — The FPT Corporation talent conversion example and the ERIA observation
• DX Trends 2025 (IPA) — The prior year survey. Reference point when checking year on year comparisons
• Discussion Paper on Developing Digital Talent for the AI Era (IPA) — Reference point on approaches to talent development