Only 7.0% of Japanese companies use AI to write code

Of all the usage categories in Japan's 2026 White Paper on Information and Communications, code generation shows the widest gap.

Japan 7.0%. United States 35.4%. China 33.5%. Germany 25.0%. Close to a five times difference.

Japan's overall generative AI usage has reached 86.4%. It has still barely entered engineering teams.

What the number actually means

Low code generation usage means a difference in the speed at which software gets built.

One in three American engineering teams writing implementation with AI, against one in fourteen in Japan, produces a different amount of elapsed time to ship the same feature. Sustained over a few years, that becomes a difference in how many products exist and how often they improve.

This particular gap cannot be closed by using AI in sales or back office functions. Build speed only rises where the building happens.

The reasons are not technical

In practice the objections fall into three groups.

Rules about code leaving the company. Sending source to an external service is not permitted. That is a legitimate concern, but in most cases the policy was written before generative AI existed and has not been revisited. Zero retention agreements and self hosted deployment are both available now. The policy simply has not caught up.

No review structure. Nobody has decided how AI written code gets reviewed. Without that, adoption blurs accountability for quality. Choosing not to adopt rather than blur it is a rational organisational decision.

No measurement. If implementation speed improves but the organisation has no metric that reflects it, then pushing for adoption depends entirely on individual initiative.

None of these concern model capability. All three concern policy and structure.

Read alongside the IPA numbers

IPA's DX Trends 2025 asked whether companies have adequate DX talent in terms of quality. Japan answered 3.8%. The United States answered 52.9%. Germany 25.1%.

The same report notes that while more than 80% of companies report a shortage, 19.4% conduct no hiring at all.

Not enough people. No hiring. And the tool that raises output per person sits unused. All three are true at once, and that combination is the current Japanese average.

If you cannot hire, output per person is the only lever left. 7.0% means that lever has not been pulled.

Start narrow

Aiming for company wide rollout on day one puts you in a legal review that never ends. The companies that got moving scoped it down.

1. Limit the repositories. Start with internal tooling or test code where confidentiality exposure is low.
2. Limit the use case. Begin with test generation and refactoring suggestions rather than new feature implementation. The review load is predictable and the cost of a mistake is small.
3. Set the review standard first. Simply writing down that AI authored code is reviewed to the same standard as human authored code resolves the accountability question.
4. Measure it. Time from pull request opened to approved, and number of review comments. Have a before and an after.

Work through those four and you arrive at the policy discussion holding evidence. Most companies open that discussion with nothing, which is why it stops at the first meeting.

Japanese engineers can use these tools perfectly well. What is missing is the set of conditions that lets them, and those conditions are something a company builds for itself.