日本語 Get in touch
All articles

AI adoption

Why your AI pilot never became a product

27.0% of Japanese companies report no organisational effort around generative AI. The same question puts the United States at 1.4%. What separates them is not the tooling.

Published 3 min read

Running a generative AI pilot is no longer unusual for a Japanese company. Having that pilot turn into daily work still is.

Japan's 2026 White Paper on Information and Communications contains a number that explains the gap directly. Asked whether the company has any organisational effort around generative AI, 27.0% of Japanese firms answered that they have none. The figure is 1.4% in the United States, 4.9% in Germany and 2.6% in China.

The same survey found that 86.4% of Japanese companies use generative AI somewhere in the business, against 90.9% in the US and 91.6% in Germany. On adoption, Japan is not behind.

So almost every company is touching AI. Roughly one in four is only touching it.

Using it and using it as an organisation are different things

Both numbers can be true at once because the usage sits with individuals. Somebody uses a model to get through their own work faster. That is happening. The company then rebuilds a process around the result. That is not.

The same white paper measures how many firms have redesigned a business process around AI. Depending on the task type, the answer runs between 6.0% and 11.9%. Fewer than one company in ten.

In practice the reason a pilot stops here is nearly always the same. The pilot was scored on whether it worked. It worked. But between working and being used every day sits a long list of things nobody has designed yet.

Who provides the input. Who checks the output. Who fixes it when it is wrong. Where it sits inside the approval flow that already exists.

Skip that and the team quietly reverts to the old method, which is the rational thing for them to do.

The government reached the same diagnosis

The Basic Plan for Artificial Intelligence, a Cabinet decision dated 23 December 2025, states in its first chapter that AI is not actively used in Japanese life or work, and that the main obstacle is a lack of implementation rather than a lack of research.

That is unusually blunt for a Cabinet document. Its fourth pillar then asks for a clear division of roles between people and AI.

That is not a technology question. It is a question of who does what, written into a Cabinet decision. It is a design brief.

What it takes to leave the 27%

Having an organisational effort means, at minimum, that four things have been decided.

  1. The task is specific. Not "use it in sales" but something with a start and an end, like producing the first draft of a quotation.
  2. Someone owns quality. Who reviews the output, and against what standard. With no standard, everyone checks everything, which is slower than the original process.
  3. There is a route for failure. When the output is wrong, where does it go and who handles it. Without that route, one bad result ends adoption.
  4. Something is counted. Handling time, rework rate, volume closed. Numbers, not impressions.

None of these four have anything to do with model quality. All four are process design.

A pilot does not stall because the model is weak. It stalls because nobody built the structure the model was supposed to sit inside.

The order matters

Most companies start by selecting a tool. That is backwards.

Decide first which task, which step, and what the finished state looks like. Get that right and the design holds whichever model you use. Get it wrong and the best model available still leaves you inside the 27.0%.

What the white paper actually shows is not that Japan was slow to adopt AI. Adoption is done. The step after it is not.

Sources

  1. MIC Japan, White Paper on Information and Communications 2026 (published 2026-07-24, surveyed Jan to Feb 2026, Japan, US, Germany, China)
  2. Cabinet Office, Basic Plan for Artificial Intelligence (Cabinet decision, 2025-12-23)

Recognise the problem

We design, build and hand over the layer that turns a stalled AI pilot into something people actually use. Tell us where you are stuck.

Get in touch