Japan's AI Gap Is a Workflow Gap

Every few months an article goes around claiming Japan is behind on AI. The numbers quoted are usually real. The conclusion drawn from them is usually wrong, and I think the mistake is expensive, because it sends companies shopping for tools when their actual problem is upstream of any tool.

Here are the numbers, from Japan's own Ministry of Internal Affairs and Communications, published in the 2025 Information and Communications White Paper.

In the 2024 survey, 26.7% of people in Japan had used a generative AI service. In the United States it was 68.8%. Germany 59.2%. France 81.2%. Read alone, that is the gap everyone quotes.

Now read the year before it. Japan was at 9.1% in 2023. It roughly tripled in twelve months. The US went from 46.3% to 68.8%, Germany from 34.6% to 59.2%. Japan grew faster than any of them in relative terms. It started lower.

A country that triples its adoption in a year does not have a curiosity problem.

Where the gap really opens

The more useful survey is IPA's DX Trends 2025, which asks companies rather than individuals and compares Japan against the United States and Germany.

Start with the good news, which nobody reports. Among Japanese companies with more than 1,000 employees, the share that have actually deployed generative AI is the highest of the three countries. Not trialing. Deployed. Japanese large enterprises moved from "testing" to "in production" between the 2023 and 2024 surveys while the trial and evaluation categories shrank.

Then IPA asks how it is being used, and the picture inverts. Across all three countries, the common answers are the same: individuals trying it out, individuals using it for their own work, drafting documents, generating ideas. On one specific option, Japan falls behind: whether the tool is built into the department's work process.

That is the entire gap in one line. Japanese companies bought the tool. People are using it personally. Almost nobody changed how the work runs.

I see this constantly. A team has licenses. Individuals use them, quietly, mostly for first drafts and translation. Nobody has removed a step from any process, nobody has changed a handoff, and the official review gates are identical to what they were two years ago. When leadership asks what the return has been, the honest answer is that some people are finishing some tasks faster, which is real but does not show up anywhere.

Why the tool never becomes the process

Three things are going on, and only one of them is about technology.

The first is that embedding a tool in a process means deciding whose job changes. That is not an IT decision. It requires someone senior enough to say the review step is now unnecessary, or that this approval moves, or that this role does something different starting Monday. IPA's own summary of Japan's pattern is that it is inward facing and partially optimized, with weak links between management, IT and the business units. If the person who could remove the step is three departments away from the person using the tool, the step does not get removed.

The second is that most companies are aiming the tool at the wrong half of the work. IPA found Japanese companies report DX outcomes in cost reduction and efficiency, where American and German companies report them in revenue and profit. Point AI at efficiency and you get a slightly faster version of what you already do. That is worth having and it is also self limiting, because the ceiling on making an existing process faster is the process.

The third one is the strangest and it is the one I would put in front of a board. IPA asked which AI roles companies need. In Japan, 56.4% said AI researchers are "not needed at our company", and 40.7% said the same about AI developers. Those are defensible answers for most firms, and IPA reads them as Japan choosing to apply AI rather than build it. Fine.

But if you are not building the model and you are not changing the process, you have not made a decision about AI. You have bought a subscription.

The small company problem is different and worse

The averages hide a split that matters more than the international comparison.

Among Japanese companies with 100 or fewer employees, the share answering either "interested but no plans" or "no plans at all" for generative AI is close to 80%. In the United States that group is just over 20%. In Germany, just under 40%. IPA's separate question on why smaller Japanese firms are not doing DX at all gets a clear top answer: they do not understand what benefit it would bring to their company. Second: they lack the knowledge and information to judge.

That is not resistance. That is an unanswered question, and it has gone unanswered because most of the material aimed at these companies explains what the technology is rather than what it would change on a Tuesday in their office.

The same report shows the wider DX picture: 96.1% of Japanese companies over 1,000 employees are doing DX in some form, against 46.8% of those under 100. A gap of more than two to one inside the same country.

What actually works

I have had the best results by refusing to start from the tool.

Pick one process that runs every week and that somebody complains about. Write down what happens now, step by step, including who waits for whom. Most teams have never done this and the document is uncomfortable to look at, which is the point. You usually find two or three steps that exist because of a system limitation nobody has checked in five years.

Then ask which step disappears, not which step gets faster. Faster is a rounding error. Disappearing is a change in the shape of the work, and it is the only thing that shows up in a business result.

Then change one step and leave everything else alone. One. The failure mode I see most often is the pilot that redesigns an entire department, takes nine months, and dies in the approval process before it touches a customer.

The last part is unglamorous and it is the part that decides whether any of it survives: someone has to own it. Not a project team. A person, with the authority to change the step and the obligation to report what happened. IPA found that more than 80% of Japanese companies say they are short of DX people, and 19.4% are not carrying out any talent acquisition at all. Those two facts sitting side by side tell you the constraint is not the labour market.

The version I would put on a slide

Japan does not have an AI adoption problem. Adoption tripled in a year, and its largest companies deploy at world leading rates.

Japan has a workflow redesign problem, which was true before generative AI existed and will be true after the current models are obsolete. AI made it visible by giving every company a capable tool and then showing, very publicly, which of them could change anything about how they work in order to use it.

That is not a technology question. It is the oldest design question there is: what does this person do now, and what should they do instead.

References

• Ministry of Internal Affairs and Communications, 令和7年版 情報通信白書, individual AI use, July 2025. Japan 26.7% in 2024 against 9.1% in 2023; US 68.8% from 46.3%; Germany 59.2% from 34.6%; France 81.2% from 56.3%.
• Ministry of Internal Affairs and Communications, 令和7年版 情報通信白書, enterprise AI use. 49.7% of Japanese companies had a policy to use generative AI in FY2024, up from 42.7%.
• Information technology Promotion Agency, DX動向2025 and its full report. Deployment by company size (fig 2-10), specific usage patterns including process embedding (fig 2-11), AI role need at 56.4% and 40.7% (section 2.3), DX participation 96.1% versus 46.8% by size (fig 1-2), reasons smaller firms are not starting (fig 1-4), talent acquisition at 19.4% (fig 3-5).