Japan Closed the AI Adoption Gap in a Year. Then 27% Reported No Change to How Work Is Done.
A year ago I wrote that Japan's AI gap was a workflow gap, not a technology gap, and that the adoption numbers would close on their own while the number that mattered stayed put.
The 2026 white paper is out. The adoption numbers closed. The number that matters stayed put.
I would rather have been wrong about this.
What closed
The Ministry of Internal Affairs and Communications published the 2026 white paper on information and communications on 24 July 2026. The survey behind it ran in January and February 2026 across Japan, the United States, Germany and China.
Individual generative AI use in Japan went from 26.7% to 58.8% in one year. That is more than double. The United States and Germany both sat at 75.6%, China at 93.6%.
The corporate numbers moved harder. Japanese companies using generative AI in at least one business task went from 55.2% to 86.4%. The United States is 90.9%, Germany 91.6%, China 98.1%. A year ago that gap was 35 points. It is now four.
Companies with a policy to use generative AI actively or in defined areas went from 49.7% to 68.9%.
If your thesis was that Japan was slow to adopt AI, that thesis is finished. Japanese companies bought it, and they bought it fast.
What did not close
The white paper asked companies what organisational effort they were making to transform work with generative AI. The answers, by country:
| Organisational effort | Japan | US | Germany | China |
|---|---|---|---|---|
| Company wide business transformation | 21.5% | 35.6% | 20.7% | 40.9% |
| Optimising work within departments | 32.4% | 40.6% | 48.1% | 32.0% |
| Individual productivity only | 11.7% | 22.1% | 26.0% | 23.8% |
| No organisational effort at all | 27.0% | 1.4% | 4.9% | 2.6% |
27.0% against 1.4%.
More than a quarter of Japanese companies are using generative AI in at least one business task and simultaneously report no organisational effort to change how work is done. In the United States that combination is essentially nonexistent.
That is not a slower version of the American curve. It is a different thing happening.
The number that made it concrete for me
The white paper breaks usage down by task type and asks, for each, whether the business process has been rebuilt around AI, whether AI is used for a specific task inside a department, or whether employees are just using it to speed up their own work.
The share of Japanese companies reporting that the business process itself has been changed to be AI centred, by task:
- Internal help desk, 11.9%
- Minutes and email drafting, 11.3%
- Sales, 9.5%
- Research and product development, 7.6%
- Marketing and PR, 7.4%
- Procurement, 7.0%
- Manufacturing and production, 6.0%
86.4% of Japanese companies use generative AI somewhere. Between 6% and 12% have changed a process around it.
The most common use, and the one where the largest share report feeling a benefit at 72.3%, is drafting meeting minutes and emails. Which is a real saving and I use it myself. It is also, precisely, the category of work where you can adopt a tool without anybody else having to change anything.
Why this happens, from having sat in the rooms
Adopting a tool is a procurement decision. One budget holder, one vendor, one contract, one rollout email. It fits inside existing authority, so it can be done quickly and by one person.
Changing a process is a design decision. It requires deciding what a step is for, who owns the output, what happens when the model is wrong, and what stops happening because the new thing happened. It crosses departments. It touches somebody's headcount justification. There is no single budget holder who can approve it.
So organisations do the first thing, at speed, and report it truthfully as AI adoption. The white paper's own numbers are the shape of that: the tool is in, the work is unchanged.
The two capability items where the white paper explicitly says Japan lags the other three countries are worth quoting because they are the same point stated as infrastructure. They are whether the company has an environment for learning the skills needed to review business processes, and whether it has trained AI on internal data or built a database the AI can reference.
Both are prerequisites for a model being useful for anything beyond drafting text. Without your own data, the model knows nothing about your company, so the only tasks left are the generic ones. Without anyone who can review a process, there is nobody to change.
What people think the thing is
There is a softer finding in the same document that I think explains more than the capability items do.
Asked what generative AI is to them, respondents in Japan, the United States and Germany most often said a machine or a tool. The second answer diverges. In Japan it is a dictionary or an encyclopedia. In the United States and China it is a friend. In Germany, a counsellor.
A dictionary is a thing you consult when you already know your question. It does not initiate, it does not ask a follow up, and it never helps you with what you did not know to look up. If that is the model in someone's head, they will use it for lookups and drafts and stop there, which is exactly the Japanese usage distribution.
And look at what Japanese individuals do not use. Text generation, 55.0%. Image and video, 16.3%. Code, 7.0%. Voice, 6.5%.
Germany's image and video figure is 47.6%, the United States 41.0%, China 55.8%. On code, the United States is 35.4%, China 33.5%, Germany 25.0%. Japan is at 7.0%.
A country using generative AI almost entirely for text is a country using one seventh of it. The overall adoption figure hides that completely, which is why the 58.8% headline is the least interesting number in the report.
The government has said this out loud
What has changed since I last wrote about this is that the diagnosis is now official.
The Cabinet approved Japan's first AI Basic Plan on 23 December 2025, subtitled "Japan's revival through trustworthy AI". Its stated goal is for Japan to become the country where AI is easiest to develop and use in the world.
Two sentences in chapter one are more direct than anything I would have written. The first says that in Japan, AI has not become something actively used in daily life or at work. The second says that for a technology where basic research and social implementation sit close together, the lack of implementation has become a major obstacle to Japan's AI development.
Not a research obstacle. Not a compute obstacle. An implementation obstacle, stated in a Cabinet decision.
The plan's fourth pillar asks for continuous transformation towards an AI society and names, as a thing to be worked out, the division of roles between people and AI.
That sentence is a design brief. Deciding what a person does and what a machine does, and where the handover happens, is interaction design with the label removed. It is not a policy question and it is definitely not a procurement question.
What I would do on Monday
If you run a Japanese company that is in the 86.4% and honestly somewhere near the 27.0%, the intervention is not another tool and it is not more training on prompts.
Pick one process. One. Something with a real cycle time you already measure, ideally something unglamorous like a quotation, an approval, an incident triage or an inbound enquiry.
Then map it as it actually runs, not as the manual says. Mark every step where a human is producing text or a judgement that a model could produce a first draft of. For each of those, decide three things and write them down: what the model produces, who checks it, and what happens when it is wrong. Then delete the steps that only existed to prepare the input for the next human.
That last part is where the productivity is, and it is the part that gets skipped, because deleting a step means telling the person who owned it that their step is gone.
The reason to do it on one process first is that it produces the thing the white paper says Japanese companies are missing: somebody in the building who has actually reviewed a process and can do it again. You cannot hire that. You can only have done it once.
Japan closed a 35 point adoption gap in twelve months, which tells you the will and the budget are both present. The remaining gap is between 6% and 12% versus 40%, and it is not going to close by buying anything.
Sources
- Ministry of Internal Affairs and Communications, 令和8年版情報通信白書 (summary), published 24 July 2026. Four country survey fielded January to February 2026. Corporate sample sizes vary by question, roughly n=477 to n=515 for Japan.
- Cabinet Office, 人工知能基本計画, Cabinet decision of 23 December 2025. The implementation passage is in chapter 1; the division of roles appears in the fourth basic policy.
- My earlier piece on the same survey series: Japan's AI Gap Is a Workflow Gap.