
"We bought expensive tools, and in the end only the same few people ever use them." "We ran training once, but a month later everything was back to normal." "We do not even know what to teach, or to whom." These are the three concerns we hear most often from executives, HR leaders, and DX managers who have been handed responsibility for AI adoption.
Bringing in a tool and getting employees to master it are two entirely different problems. As the barrier to adopting generative AI has fallen, the decisive battleground has shifted from "adoption" to "adoption that sticks." And what determines whether it sticks is not features but people. This article lays out how to develop talent so that your investment does not miss the mark, raising AI literacy across the organization, in a form you can put to work directly.
First, let us set the premise. METI's IT Human Resource Supply and Demand Survey projected a shortage of up to roughly 790,000 IT professionals by 2030. It is a frequently cited figure, but it is a projection made as of March 2019, on assumptions from before generative AI became widespread. The situation has changed. While tools now shoulder some tasks such as coding, the shortage of "people who can actually deliver results using AI" has become more acute than before.
A private-sector survey likewise found that as of 2024, roughly 70% of companies cited "insufficient literacy and skills" as a barrier to generative AI adoption. But the concern we hear on the ground in 2026 has shifted somewhat in character. Not "we have no usable tools," but "we adopted it and have no one who can use it well" and "we trained people but it did not stick." The problem has moved from being about the tool to being about people, and that is our honest impression.
The Ministry of Internal Affairs and Communications (MIC) White Paper on Information and Communications 2025 also notes that while Japan's use of generative AI grew significantly year over year, the gap in usage rates against the United States and China remains large. A large gap to close also means this is a moment when companies that develop their people now can pull ahead.
So should you solve it by hiring? Recruiting AI talent from outside is one option, but competition in the market is fierce, and it is not realistic for small and mid-sized companies to keep securing high-level specialists. It is more reliable, and cheaper, for most companies to develop existing employees who know the business intimately into AI-capable people.
There is one thing worth emphasizing here. AI talent development is not "technology education," it is "part of organizational transformation." Teach people how to operate a tool and stop there, and it will not stick. The real objective is to root a culture of using AI in daily work throughout the organization.
There was a time when "AI talent means engineers" was the dominant idea. The center of gravity has clearly moved since then. With the spread of generative AI agents and no-code tools, the necessity of technical specialists who build models from scratch has declined, while the value of "literacy across all employees" and "champions who keep AI adoption running on the ground" has risen. Designing your development program in that order fits the reality of 2026.
Organized by demand, from highest to lowest, the roles look like this.
| Role | Primary Responsibilities | Required Skills | Typical Development Time | How to Develop |
|---|---|---|---|---|
| AI Users (all employees) | Use generative AI safely in daily work | Basic grasp of how it works, prompt fundamentals, risk judgment | 1-2 months | Internal development |
| Front-line Champions | Drive adoption in their department, collect and spread use cases | Deep operational knowledge, interpersonal influence, ability to teach | 1-2 months | Internal selection |
| AI Adoption Managers | Adoption planning, execution, impact measurement | Business analysis, project management, ROI calculation | 3-6 months | Internal development |
| AI Workflow Designers | Design business workflows and AI agents, curate knowledge, review AI output | Logical design, domain knowledge, tool integration | 3-6 months | Internal development |
| Technical Specialists | Model selection, API implementation, operations (MLOps) | Python, API development, data infrastructure | 12+ months | Consider outsourcing |
The role once called "prompt engineer" has broadened from writing one-off prompts into designing entire business workflows and AI agents, curating internal knowledge, and reviewing AI output. That is why the table recasts it as "AI Workflow Designer." Simply being able to write a clever instruction is no longer enough.
For a company of 50 or fewer, there is no need to assign all five of these roles to different people. Raise the baseline of "AI Users" across everyone, have two or three people double up as "Front-line Champions" and "AI Adoption Managers," and outsource technical specialists to an external partner. This realistic combination works perfectly well.
Running on your own role definitions alone leaves gaps in what you teach and makes it hard to explain externally. Mapping to IPA's (Information-technology Promotion Agency) Digital Skill Standards (DSS) keeps your design from drifting. The DSS come in two parts: the DX Literacy Standard (DSS-L) for all employees, and the DX Promotion Skill Standard (DSS-P) for driving talent, and learning items related to generative AI were added between 2023 and 2024.
A rough mapping goes like this. Tie the all-employee "AI Users" tier to DSS-L, and tie front-line champions and AI workflow designers to the DSS-P talent archetypes (business architect, designer, data scientist, software engineer, cybersecurity). That lets you express the required skills and proficiency against a public yardstick, and it becomes the foundation when you build an internal certification.
Throwing every employee into the same program is the most common design, and the one that sticks least. Split people into three tiers by role, and change the goal and deliverable for each.
The goal is to understand the basic mechanics of AI and be able to picture where to use it in one's own work. This tier is thin and broad, but it reaches everyone.
| Week | Topic | Format | Duration |
|---|---|---|---|
| Week 1 | AI fundamentals and how generative AI works | E-learning | 2 hours |
| Week 2 | Operating your company's tools and prompt basics | Hands-on | 3 hours |
| Week 3 | Build a prompt you can use in your own job | Workshop | 2 hours |
| Week 4 | Risks and precautions (data leakage, hallucination) | E-learning and quiz | 1.5 hours |
Adjust delivery to your scale. A 30-person design studio can run two 3-hour workshops with everyone attending. The first covers fundamentals and a demo; the second is hands-on using each person's own work as the material. Being small is exactly the advantage that lets you resolve questions on the spot. A 250-person food manufacturer can first level everyone up on fundamentals with e-learning, then hold half-day hands-on sessions by department. Change the material by team, proposals for sales, inspection reports for quality control, and you create the felt sense that "this works for my job."
What works across every industry is this "pull it toward your own work" design. For a healthcare provider, nursing records or referral-letter drafts; for retail, promotional copy or summarizing shift notices; for professional services, drafting documents or research; for local government, rewriting resident-facing documents in plain language; for logistics, daily reports or inquiry replies. Once the material becomes personal, the post-training adoption rate changes completely.
Select one or two people from each department and develop them into people who can keep AI adoption running in their own department. This is the highest-return investment of the three tiers.
| Week | Topic | Content | Deliverable |
|---|---|---|---|
| Weeks 1-2 | Prompt and workflow practice | Advanced instruction design, mapping to business workflows | 5 templates |
| Weeks 3-4 | Business analysis and identifying AI opportunities | Department workflow analysis, ROI estimation | AI opportunity candidate list |
| Weeks 5-6 | No-code and AI agent use | Workflow automation, simple agent design | 1 automation prototype |
| Weeks 7-8 | Change management for adoption | How to teach, handling resistance | Department rollout plan |
There is a clear right answer for how to choose champions. Do not pick "the person who is good with IT," pick "the person who knows the work deeply and is trusted by colleagues." To spread the appetite for adoption to those around them, interpersonal influence matters more than technical skill. Get this wrong and you fall into the classic failure described later.
This targets members of IT and data analytics teams, covering AI model selection, API integration, building data infrastructure, MLOps (monitoring, retraining, deployment), and security and governance. Developing them over 3-6 months while running real projects is the realistic approach.
That said, not maintaining this tier in-house is entirely valid. Now that generative AI agents and managed services are widespread, a company of 50 or fewer can outsource the technical side to an external partner and pour everything into strengthening Tiers 1 and 2. That yields a better return on investment.
Generic AI training builds knowledge, but the front line does not move without the "I can use this tomorrow" feeling. Always include exercises based on the tools you actually use and your own business data.
Even if people feel "AI is amazing" during training, they forget once time passes before they use it on the job. Within the training, have each person build one prompt that helps their own work and try it right there. That small win changes behavior.
Build in the devices that keep a single training event from being the end of it, from the start.
| Initiative | Frequency | Purpose |
|---|---|---|
| Internal study group | Monthly | Share use cases, introduce new features |
| Q&A channel | Always on | Resolve stumbling points on the spot |
| Prompt library | Always on | Accumulate and share instructions that worked |
| Utilization contest | Quarterly | Compete on improvement ideas and celebrate them |
| Internal certification | As needed | Make proficiency visible and set the next goal |
As a yardstick for proficiency, a growing number of companies combine internal certification with an external generative-AI certification. Using an outside credential as "proof of having acquired the fundamentals" makes the learning goal concrete and easier to fold into HR evaluation.
Measuring training effectiveness with the four levels of the Kirkpatrick model makes it easier to explain to leadership.
| Level | What to Look At | How to Measure | Target |
|---|---|---|---|
| 1. Reaction | Satisfaction | Immediate post-survey | 4.0/5.0 or higher |
| 2. Learning | Acquisition | Post-training test | Pass rate 80% or higher |
| 3. Behavior | Use in work | Usage logs and interviews at 1 month | Utilization rate 60% or higher |
| 4. Results | Business improvement | KPI comparison at 3 months | 15% or greater time reduction on target tasks |
As a starting point, measuring where your organization stands on AI literacy first lets you see the gain after training. When you want a quick read, starting with a simple check like the AI Literacy Check is a good approach.
That leaders understand AI and actually use it makes a big difference to the development effect across the whole organization. At one 100-person consulting firm, the president publicly stated internally that "every morning I spend 30 minutes using AI to summarize industry news," which created an atmosphere of "if the president is doing it," and participation rates rose. The opposite failure is covered in the next section. The approach of running governance (preparing for data leakage and hallucination) in parallel with development is the direction indicated in METI's AI Guidelines for Business as well.
Failure stories tend to scatter throughout an article, so we consolidate them here. This is probably the section readers most want.
Making an IT-loving junior employee the champion and leaving them isolated. At a 120-employee real estate company, IT-loving junior staff were appointed as each department's AI champion. But veterans would not accept it, with reactions like "I'm not taking direction from that kid," and the champions themselves became isolated. After switching to mid-career employees trusted on the ground, the utilization rate improved from 25% to 65% in three months. The fix is simple: choose champions by trust, not technical skill.
Leadership saying "leave it to the young people." At a 200-employee manufacturer, leadership said "AI is for the young employees to handle," and managers took it to mean "this has nothing to do with me." As a result, managers became a bottleneck, approvals did not come through, and company-wide adoption stalled. Only after leadership declared "I will use AI at least once a week myself" did the atmosphere finally change. A leader's declaration of personal ownership can work better than a training budget.
Training that is one-off and left to fizzle. Even if satisfaction is high after a single session, without follow-up it is back to normal in a month. The fix is to launch the "mechanisms for continued use," study groups, a Q&A channel, a prompt library, at the same time as the training. Design training as a line, not a point.
Bringing in the tool first and deferring the adoption design. When the contract runs ahead and it becomes "the rest is up to the field," the tool gets left with both the reason to use it and the way to use it vague. Which task, made easier how? Draw the blueprint for adoption before you distribute it.
Whether to develop AI talent in-house or outsource it comes down to the following criteria.
| Criterion | Favors Building | Favors Buying |
|---|---|---|
| Importance of business knowledge | High (company-specific know-how is essential) | Low (addressable with general-purpose technology) |
| Frequency of use | High (used daily) | Low (project-based) |
| Speed of technology change | Slow (a stable domain) | Fast (cutting-edge model development) |
| Ease of securing talent | High (a foundation exists internally) | Low (hiring specialists is difficult) |
One footnote on the reality of 2026: the necessity of keeping technical specialists in-house has declined. With the spread of generative AI agents and no-code, work that once assumed in-house development can now be handled by external services. Conversely, "AI Users" and "Front-line Champions" rooted in business knowledge cannot be grown by outsourcing. Here, building in-house is the only choice.
Recommendations by scale look like this. Under 50 people: build Tiers 1 and 2 in-house, fully outsource Tier 3. From 50 to 300: build Tiers 1 and 2 in-house, and move Tier 3 from external training toward in-house. Over 300: internalize all tiers while partnering externally only on the most advanced technical domains. If you are unsure about the training design itself, having an external partner handle only the initial design in a hands-on model and then bringing operations in-house is also realistic.
You can use national subsidies for AI talent development. The centerpiece is the Ministry of Health, Labour and Welfare's Human Resource Development Subsidy, whose Business Restructuring Reskilling Support Course and Investing in People Promotion Course, aimed at DX and reskilling, are a good fit for AI training. These courses were still running as of May 2026 (per the May 14, 2026 disbursement guidelines).
What to watch for is that the expense subsidy rate, upper limits, and application deadlines move with annual program revisions. Historically, smaller companies have had more generous expense subsidies and a portion of wage subsidies available, but the specific figures change. Always confirm the current values in MHLW's latest brochure before applying. In particular, the Investing in People Promotion Course is described as a time-limited measure and is expected to be near its final year. Planning while you can still use it is the wise move.
Subsidy applications require an advance plan filing, and unless you begin preparation two to three months before the training, you will not make it in time. We recommend moving early even just on the logistics.
When explaining the return on investment to leadership, the following formula is useful.
Training ROI (%) = (Business improvement from training - total training cost) / total training cost x 100
Let us run a rough estimate, purely as an example, for an 80-employee precision-parts manufacturer. Training cost is 1.2 million yen in instructor fees plus 400,000 yen in labor cost for attendee time, 1.6 million yen in total. Then subsidies come in. Assuming you can receive roughly 70% of training expenses, the effective cost falls to somewhere around 700,000 yen (subsidy rates change by year, so always calculate with the latest values). On the benefit side, 28 hours of monthly task savings, converted at 3,500 yen per hour across 12 months, comes to about 1,176,000 yen per year. ROI works out to roughly 68%. And because the effect continues with no additional cost from the second year on, the numbers grow further on a cumulative basis.
Beyond satisfaction surveys, present this "time saved converted to money" and the "effective cost after subsidies" side by side. That is the surest route to getting on the agenda for next year's budget.
Follow-up after training, more than the training itself, determines the adoption rate. Setting goals in defined intervals prevents momentum from stalling.
The 30-day goal is a state where all trainees use AI at least once a week. Send weekly tips, set up a rapid-response Q&A channel, and encourage habit formation with a "30-day challenge to try one thing every day."
The 90-day goal is a state where each department's champion has produced at least one use case in their department. Run monthly use-case sharing sessions, peer reviews among champions, and results reports to leadership as a set.
The 180-day goal is a state where AI use is built into operational manuals and has become "the norm." Reflect AI procedures in manuals, add the curriculum to new-employee training, and include AI use in performance evaluation. Once you reach here, development starts to run on its own as a system.
Q. Where should we start with AI talent development? Rather than immediately building company-wide training, the fastest route is to first measure where your organization stands on AI literacy. Once you can see who can do what, the content and sequence of what to teach become clear. From there, a two-track first step is hard to get wrong: foundational training for all employees, plus selecting one or two champions per department to develop.
Q. Can a small company with no IT department still develop AI talent? Yes. In fact, the smaller the company, the more realistic it is to develop existing employees who know the business into AI-capable people rather than hiring specialists from outside. Leave technical domains like model development to an external partner, and concentrate internally on raising everyone's ability to use generative AI in their work. Drawing that line clearly improves cost-effectiveness.
Q. Is training on how to use ChatGPT enough? It is useful as an entry point, but it will not stick on its own. Even after people learn how to operate a tool, they stop using it on the job if they cannot see where and how it applies to their own work. You need to design exercises based on your real business, mechanisms for continued use, and impact measurement as a set.
Q. What should we do if employees resist AI? Ordering people to use it is counterproductive. In most cases, resistance comes down to either fear of losing one's job or the assumption that it looks too difficult. What works is preparing a small win first, where a person's own tedious task genuinely gets easier, so they come to see AI as an ally rather than a tool for evaluation or pressure.
Q. Should we have employees earn a generative-AI certification? It is not essential, but it works well as a learning goal and a yardstick for proficiency. Using an external certification as "proof of having acquired the fundamentals" prevents the problem of training that stays vague. The practical use is to combine it with internal certification and make it a factor in performance evaluation.
Q. Which should we do first, external training or in-house? We recommend borrowing external hands-on support for the initial design and launch, then bringing operations in-house. Designing everything from scratch on your own takes time and is prone to gaps. Have a specialist set up just the running start, then keep it going internally, and you strike a balance between cost and self-sufficiency.
"We have neither the time to design training nor an in-house specialist." When that is the case, using external hands-on support for the initial design and launch is the realistic move. TIMEWELL's WARP provides AI-adoption consulting and training, with former senior DX and data strategy professionals accompanying you on a monthly basis. There are three plans depending on your stage.
If you would like to talk through the approach that fits your organization, see the WARP service page, or share the specifics of your design in a free WARP consultation. If you would like to start by understanding where you stand, try the AI Literacy Check first.
Related articles
This article was produced with the help of AI. A human verified the primary sources and edited the text before publication.
A four-phase roadmap for turning generative AI adoption into real business results, explained with the latest data from PwC and Japan's Ministry of Internal Affairs and Communications. Covers KPI design to prevent PoC death, governance, work redesign for the AI-agent era of 2026, and how to structure the effort by company size.
A practical guide to raising AI literacy across your entire organization, covering current-state assessment, role- and level-based development, systems for sustained adoption, and measurement, backed by the latest 2026 data.
A practical framework for measuring AI investment ROI across three axes -- direct benefits, indirect benefits, and adoption (utilization rate) -- covering the formula, a worksheet, 3-year TCO, measurement methods, and sensitivity analysis. It also maps out the 2026 shift toward outcome-based evaluation driven by agentic AI.