AI Adoption Ongoing Support | Making It Stick After Rollout

The generative AI accounts have been handed out, but only a handful of people actually use them. The PoC succeeded, yet not a single workflow has changed. A Teikoku Databank survey shows the share of companies using generative AI in their operations has climbed to 34.5%, but among companies that have already adopted it, more than 70% of managers report that colleagues who cannot use the tools are creating friction in day-to-day work. Installing a tool and building a state where the shop floor keeps using it are two different projects. This article looks at AI adoption ongoing support, the hands-on partnership that carries a rollout through to lasting results: what you can actually ask a partner to do, how it differs from insourcing support and outsourced development, and what extra work is required at sites in Thailand and the wider ASEAN region.

Why generative AI rollouts fail when the project ends at installation

Adoption rates have grown, but results are unevenly distributed

Teikoku Databank surveyed corporate trends in generative AI between 17 and 31 March 2026, collecting responses from 10,312 companies. Of those, 34.5% said they were using generative AI in their operations. The picture varies sharply by company size.

Among companies that are using it, 86.7% said they felt a real effect. That breaks down into 25.2% reporting “significant results” and 61.5% reporting “some results”.

At first glance these numbers read as good news. Change the lens, though, and a different picture appears. The 86.7% satisfaction figure applies only within the one in three companies that answered “we are using it”. Most of that group chose “some results”, and only a quarter of active users were confident enough to say “significant results”. In other words, a large number of companies are stuck in the middle ground: the tool is in, but the payoff is thin.

The reason they stall is that only some people use it

A generative AI usage survey conducted by PRIZMA Research between 28 and 29 January 2026 shows what that stall looks like more directly. Among 1,008 managers and supervisors at companies that had already deployed generative AI, more than 70% said that colleagues unable to use the tools were disrupting work. The breakdown is 22.2% who “strongly agree” and 49.1% who “somewhat agree”.

This is not a technology problem. Under the same tool, the same licence and the same internal rules, one person switches their daily document work to the new method while another leaves the browser tab open, never uses it, and closes it again. When that gap hardens inside an organisation, work piles onto the people who can use the tools, the steps owned by those who cannot become bottlenecks, and the process as a whole gets more complicated than before. That is the real source of the feeling that “we adopted it and work did not get any easier”.

Only a small minority reach enterprise-wide deployment

The same structure shows up even more starkly with AI agents. According to McKinsey’s research report “The state of AI in 2025”, 62% of companies are interested in AI agents and have begun experimenting, while only 23% have managed to deploy them at enterprise scale. Just 6% report an enterprise-level contribution of 5% or more to EBIT.

What stands out is that the report locates the cause of the plateau on the organisational side rather than the technical side. Specifically, it identifies three dividing lines: whether business processes have been redesigned, whether accountability is clearly assigned, and whether the culture tolerates failure. Progress does not stall because model performance falls short. It stalls because nobody has settled who holds the authority to change how the work is done, and who makes the call when something does not work.

The list of concerns bites after go-live, not before

The Teikoku Databank survey also asked what companies see as concerns or obstacles in using generative AI. Respondents could select up to three answers.

Read these five items as “worries at the evaluation stage” and the countermeasure ends up being nothing more than holding a review meeting before you buy. In practice, all five reach their real test after deployment. Accuracy only becomes a live issue once the shop floor starts using outputs in operational decisions. Which tasks to apply it to cannot be separated out until people have tried it across the board. Rule-making is catch-up work that begins when unanticipated uses appear.

PwC Japan’s six-country comparison in its Spring 2026 generative AI survey points the same way, showing that creating value from generative AI depends on “AI Readiness” — the ability to embed it into business operations, data and processes. The dividing line is not skill at selecting tools but how deeply you can work it into your own operations and data.

There is a class of problems that arises after go-live and can only be solved after go-live. That is precisely why ongoing support exists as a distinct type of engagement.

What ongoing support means and how it differs from insourcing support and outsourced development

Definition

A column published by NTT East describes ongoing support as an engagement that runs “from the stage of evaluating a new technology, tool or service, through actual use in operations, to embedding it on the ground and reaching self-sufficiency”. The key point is that the scope is not cut off anywhere between evaluation and self-sufficiency. Rather than selling evaluation, build or training as separate slices, it treats the path to working results and a running operation as a single continuous line.

How it differs from outsourced development and insourcing support

External support around AI broadly splits into three shapes. The names sound similar, but the contracted deliverable, the exit condition and the way risk is held all differ.

None of this makes outsourced development bad or insourcing support inferior. If what needs to be built is clearly defined, outsourced development is the shortest path. If the area needs continuous refinement, insourcing is the right answer. Neither, however, directly solves the situation where a tool has been deployed but nobody on the floor uses it. A contract to build ends when the build ends, and a contract to teach ends when the teaching ends. Ongoing support is the one that commits to staying until the thing is actually used.

For the question of which layers to keep in-house and which to hand outside, we set out a layer-by-layer framework in AI insourcing support and deciding where each layer belongs. If you do hand development outside, the contract and estimation traps are covered in How to choose an AI development company in 2026. This article sits neither before nor after those, but covers the work itself of embedding what you deployed and keeping it delivering.

What ongoing support actually involves

Because the term invites vague language, here are the typical tasks in concrete terms.

• Observe how work is actually done and identify the steps where generative AI or AI agents can be inserted
• Prioritise those steps and decide the order of attack and the criteria for judging each one
• Build prompts and settings around the shape of the work, then distribute them as templates per department
• Cross-reference usage logs against operational metrics to separate the tasks that are being used from those that are not
• Ask the shop floor why a task is not being used, and decide whether to fix the tool or fix the process
• Add verification steps and guardrails wherever incorrect outputs or leakage risk have surfaced
• Develop internal champions in each department and gradually reduce the external footprint

None of these are one-and-done. Staff rotations change who the users are, model updates change behaviour, and when the work itself changes, so do the points where AI belongs. Being designed around continuous adjustment is the essence of this engagement model.

Three outcomes ongoing support delivers

Outcome 1 — Adoption on the floor

The most obvious outcome is that more people use the tools. Setting a goal of “every employee uses it daily” is a mistake, though. Adoption is not about pushing a usage percentage up. It means a specific task has been replaced by a new method and does not revert.

Three conditions have to hold for a task to be replaced. First, it must be clearly faster or easier for the person who owns that task. Second, there has to be a fallback for when it does not work. Third, the method must be written down as a team standard rather than an individual’s personal trick. In an ongoing support engagement, you keep narrowing the target task until all three are in place, then move on to the next task once they are. Behind the finding that more than 70% of managers see friction from colleagues who cannot use the tools sits exactly this issue: no standardised method has been distributed.

Outcome 2 — Making ROI visible

Every ongoing support engagement eventually hits the question “so how much did it actually save?”. Leave that vague and you will have nothing to say when the budget comes up for renewal, and the programme gets cut.

Measurement cannot be designed in a hurry after go-live. You need a baseline reading of processing time and volume for the target task before you touch it. In ongoing support, the moment a target task is chosen you also decide what will be measured, when, and how. The design of measurement and how to apply it in practice are covered step by step in Measuring AI impact and designing ROI.

It is worth noting that McKinsey’s figure of only 6% contributing 5% or more to EBIT reflects two things at once: that results themselves are thin, and that few companies manage to connect their results all the way through to enterprise financial metrics. Hours saved in individual departments do not show up in the financials unless they translate into changes in headcount allocation or outsourcing spend. Deciding up front who owns that connection matters.

Outcome 3 — Avoiding known failure modes

The third outcome is not repeating failures that are already well documented. Stalling at PoC. Spreading the target too wide so results get diluted. Decisions stuck because no owner is named. The shop floor afraid to use the tool because nobody has defined who is accountable for a wrong output. These patterns recur, and you can act on them in advance. We break down the structure of stalled generative AI rollouts in detail in Generative AI implementation failure in 2026 and the dividing line for Thailand sites.

The three outcomes are not independent. Without adoption there is nothing to measure, and without measurement you will not even notice failure. The sequence forms a loop: narrow the target task, define how to measure, embed the new method, confirm the numbers.

Barriers specific to Thailand and ASEAN sites

Adoption is high, but the head-office playbook does not transfer

UOB’s Business Outlook Study 2026, released on 30 June 2026 and based on a survey of 265 executives and decision-makers, reports that more than 70% of Thai SMEs have already adopted AI. That is above the regional average. Among adopters, 58% report cost savings and 44% report productivity gains, and manufacturing is identified as one of the sectors under the strongest cost pressure due to high operating expenses and supply chain complexity.

In other words, the debate in Thailand over whether to adopt AI is already closing, and how you run it afterwards is becoming the competitive variable. But that “how you run it” does not hold up if you simply import the approach that worked at the Japanese head office.

The issues that come on top

The third and fourth are the easiest to overlook. Compliance with the personal data protection act is a legal question and, at the same time, a question of whether the floor may use the tool at all. If nobody can answer “may I put this customer list into a summary?”, people default to caution and stop using it. Showing which data may be used, task by task, does more for adoption than a list of prohibitions.

It is also common for group-wide rules set by the Japanese head office to sit badly with local practice. Head office writes rules for head-office circumstances, without assuming the workload or staffing profile of the local site. This calls for someone standing between the two, translating the rules into local procedures while preserving their intent. The value of having an ongoing support partner physically in-country lies in being able to carry out that translation continuously.

Language is not translation, it is moving the shape of the work

Multilingual support tends to be treated as a translation problem, but in practice it is thornier. Japanese work procedures carry a large volume of unwritten assumptions. Who to check with. How far you may decide on your own. What to do when an exception arises. Work that ran on those tacit assumptions in Japanese will not function simply by being rendered into Thai.

Inserting generative AI into a process forces you to put those tacit assumptions into words. Writing them out as prompts and operating rules is laborious, but a by-product is that the work itself gets tidied up. That is why ongoing support at a multilingual site often produces more than an AI rollout alone would.

How to choose an ongoing support partner — four things to check

Check 1 — Do they have adoption support and a training programme?

The NTT East column also notes that support during the operating phase is indispensable, not just help at the initial rollout, and that this requires continuous training and the development of guidelines. That is exactly where your evaluation of a prospective partner should focus. A proposal may say “implementation support” while the substance amounts to initial configuration plus a kick-off training session.

What to check is what the contract commits them to doing six months and a year after go-live. What gets reviewed at the regular meetings, who attends, and what changes if the metrics do not improve. A partner who can answer that concretely has actually seen an engagement through to adoption.

Check 2 — Can they discuss security and data handling at the operational level?

An explanation that stops at “our service is secure” gives you nothing to decide with. What you need is an explanation at a granularity you can use for operational judgements: how far your own business data may be entered, where entered data is retained, and if it is retained, who can delete it.

The Teikoku Databank figures — 33.5% naming leakage risk and 50.4% naming accuracy as an obstacle — also reflect the fact that explanations at this granularity are not reaching the shop floor. When the criteria do not reach the people doing the work, even permitted uses grind to a halt.

Check 3 — Is measurement built into the design from the start?

Following the “AI Readiness” logic from PwC’s survey, whether you get results depends on your ability to embed AI into operations, data and processes. A simple way to test that ability is to see whether the partner can express, at proposal stage, what success will look like in numbers.

Be wary of proposals that put usage rate alone forward as the success metric. There are real cases where usage climbs while neither working hours nor costs change. Check that the measurement targets connect through to business metrics — processing time and volume for the target task, and changes in outsourcing spend or overtime.

Check 4 — Do they understand the work and the local organisation?

At a manufacturing site, production control, quality, purchasing, accounting and HR have very different operational characteristics. The steps where generative AI helps, and the way risk shows up, vary by step. A partner without industry and process understanding cannot prioritise where to start, and the result is a thin proposal amounting to “let’s use it across the company”.

On top of that, at a Thailand site, whether the partner has people on the ground determines the pace of real work. Support delivered across a time difference and a language barrier can sustain a monthly review, but it cannot respond within days when the floor hits a blocker.

AI adoption ongoing support from TOMAS TECH

TOMAS TECH is based in Bangkok, delivering production management and energy management systems to Japanese manufacturers and providing hands-on execution support for factory IT and DX. In the AI space, our centre of gravity sits less on installing tools and more on building the state where what was installed keeps running on the shop floor.

In practice we support the full sequence: identifying and prioritising target tasks, building prompts and operating procedures that fit the work, producing procedure documents and training content in both Japanese and Thai, monitoring usage alongside operational metrics, and developing internal champions. We are also set up to handle in-country the friction points specific to overseas sites, such as reconciling head-office rules with local practice and sorting out how personal data categories are handled.

Because we have come into factory processes through production management system rollouts, we can choose where to insert AI with an understanding of which steps are managed by numbers and which run on human judgement. We also work with companies that have already deployed another vendor’s tools and cannot get past that point.

Frequently asked questions

What is AI adoption ongoing support?

It is an engagement that runs from evaluating generative AI or AI agents, through actual use in operations, to embedding the tools on the shop floor and reaching self-sufficiency. Unlike outsourced development, which ends when the system is delivered, or insourcing support, which ends when skills have been transferred, ongoing support stays involved until the target task has genuinely been replaced by a new method and the results can be confirmed in numbers. The main activities are periodic review of the work, refinement of prompts and operating procedures, monitoring of usage, and development of internal champions.

Where should I go for advice on adopting AI?

It depends on what you are stuck on. If what you want to build is defined and you lack development capacity, a development company fits. If you want to build an in-house development capability, insourcing support fits. If you have already deployed something but the floor does not use it, an ongoing support partner fits. At a Thailand site, choosing a partner with a local presence who understands both head-office rules and local practice will cut the time lost waiting on decisions. If you cannot yet tell which you need, the fastest route is to talk to someone who will start by organising your current operations and pain points with you.

Where should we start when building AI capability in-house?

The standard approach starts with choosing the target task, not choosing the tool. Good first candidates are tasks with high volume, established procedures, and recoverable consequences if something goes wrong. Next, measure the current processing time and volume for that task. Without that baseline you will not be able to explain the impact later. From there, run a small trial, document what worked as a team standard, and extend to the next task. The framework for deciding which layers to keep in-house is covered in detail in our article on AI insourcing support.

What does ongoing support typically cost?

Costs vary widely with scope and the frequency of involvement, so quoting a single market rate would be misleading. When comparing costs, do not look only at the figure — line up monthly hours of involvement, the number of target tasks, whether in-country support is included, and whether the creation of training content is in scope. A contract that looks cheap but covers only regular meetings with no hands-on work is a different product from one that includes writing the operating procedures. A practical way to judge is to estimate first the hours and costs you expect to save on the target task, and then see whether the fee fits inside that range.

Summary

Generative AI adoption rates have risen. The Teikoku Databank survey shows 34.5% of companies using it in their operations, with 86.7% of those companies reporting a real effect. Yet only 6% achieve an enterprise-level contribution of 5% or more to EBIT, and more than 70% of managers at companies that have adopted it report friction from colleagues who cannot use the tools. What closes that gap is not a new model or a new tool. It is the unglamorous work of narrowing the target task, deciding how to measure, writing the method into the procedures people follow, and keeping it in use.

Ongoing support is the engagement model for running that work alongside an external partner. Where outsourced development aims at building something and insourcing support aims at becoming able to build, ongoing support aims at the thing being used and delivering results. When selecting a partner, check four points: whether they have adoption support and a training programme, whether they can discuss data handling at the operational level, whether measurement is designed in, and whether they understand both the work and the local organisation. At sites in Thailand and ASEAN, multilingual delivery, local staff training, compliance with the personal data protection act, and dual governance with head office all come into play on top.

Even if you have already deployed the tools, the approach can be restructured from here. In fact, an organisation that has been through one round is often better placed to decide the next move, because it has first-hand experience of where things get stuck.

Perhaps you have deployed AI but cannot settle on the next step, usage is not spreading internally, or you cannot produce the numbers to justify it. Any of those stages is fine. We are happy to start by organising your current operations and pain points together. It is also perfectly fine if you are still evaluating and have not decided whether to engage anyone. Enquiries are welcome via our Contact page.

References

• Teikoku Databank — Corporate trends survey on generative AI, March 2026
• Commerce Pick — Generative AI usage survey by PRIZMA Research
• McKinsey — The state of AI in 2025: Agents, innovation, and transformation (November 2025 PDF)
• Nikkei xTECH — Reading McKinsey’s report on the perception gap between executives and employees on generative AI
• UOB — Business Outlook Study 2026
• NTT East — What is ongoing support
• PwC Japan — Generative AI survey, Spring 2026