Why Generative AI Does Not Stick Inside Companies

Hello, this is Ryuta Hamamoto from TIMEWELL.

A few years ago the question was how to get started with generative AI. It is not any more. People are using it. And nothing has changed.

This piece is for companies in that state. It is about retention, not adoption.

The short version:

• Usage is no longer the problem. Intent to use sits at 68.9% of Japanese firms
• What is missing is organisational: 27.0% report no organisational effort
• It fails to stick because the practice stays private to individuals
• Start by writing down the safe envelope, not by listing prohibitions
• If you appoint a lead, hand over time and authority together

People are using it. Japan is not behind on that

Let me remove one assumption first. "Japanese companies are not using generative AI" no longer matches reality.

In the 2026 edition of Japan's information and communications white paper, the combined share of firms answering that their policy is either to use generative AI actively or to use it in limited areas was 68.9%, up sharply from 49.7% the previous year. Large firms were at 74.0%, small and medium firms at 58.1%¹.

The white paper notes that Japan's increase was larger than in the other three countries surveyed — the US, Germany, and China. This is a catching-up picture.

But the same white paper carries another figure. On transforming how work is done, 27.0% of Japanese firms report no organisational effort at all — a low level of engagement compared with the US, Germany, and China².

Individuals are using it. The organisation is doing nothing. That is where things stand.

The breakdown of effects makes it sharper

Layer on IPA's July 2026 survey of 1,799 Japanese companies and the outline gets clearer³.

Asked what AI actually delivered, 91.6% cited work becoming more efficient or faster. Against that, 3.9% cited higher revenue or profit, 4.5% improved customer satisfaction, and 2.7% an expanded customer base.

Usage follows the same shape. Summarising, translating and proofreading text or audio at 82.5%; drafting documents and reports at 80.5%. Meanwhile, upgrading the firm's own products and services sits at 10.9%, and planning support for production, logistics or service delivery at 6.0%.

Individual tasks got faster. Nothing that leaves the building changed.

Reading that as a limitation of generative AI is premature. It is simply what happens when individuals apply a tool to their own tasks with no organisational effort behind it.

Four structures that stop it sticking

Working with companies, the places it stalls are fairly consistent.

1. Practice is never shared. The good users are genuinely good. How they phrase a request, where they draw the line between delegating and doing it themselves, which failure modes they avoid. All of it lives inside one person. The colleague next to them opens the same tool and stops at "so what do I type?"

2. The permitted scope is vague. Can I put a customer's name in? Can I paste a contract? Can I send the output straight to a client? Where this is unclear, the careful people stop using it — and only the careless continue. That is the worst combination available.

3. There is no recognition. Someone finishes faster using AI and receives proportionally more work. Their pay is no different from a colleague who did not bother. Keep that up and the clever ones start hiding the fact that they use it.

4. The lead is part-time. Many companies have appointed someone. Often on top of an existing role, with no time, no authority, and no budget. That is not an appointment. IPA found that 85.5% of firms describe themselves as somewhat or severely short of people to drive DX, with the shortage of employees who combine operational knowledge with basic AI literacy remaining especially high³.

None of these gets fixed by changing tools. They are questions of organisational design.

There is an order to this

First, write down what is permitted.

This comes first because nothing else works without it.

One trick on how to write it. Describe the safe envelope rather than listing prohibitions. "Do not enter personal data" leaves the reader stuck on what they may enter. Instead: "internal published material, text you wrote yourself, and public information can go in as they are"; "redact customer names and contract values"; "verify the facts yourself before any output goes outside." Written that way, people start using it.

Government material helps here. The AI Business Operator Guidelines published by Japan's Ministry of Internal Affairs and Communications and METI, and the Digital Agency's guidebook on managing risks in text-generation AI use, both work as a base for an internal policy. There is no need to start from nothing.

Second, collect what works and circulate it.

Unglamorous and effective. Once a month, hold a slot for "this was useful this month." Make it a formal presentation and nobody will contribute — keep it at the weight of a single line in a chat channel.

Turn the good users' methods into an organisational asset. Unshared, they leave when the person does.

Third, give the lead time and authority.

Appointing is not enough. How many hours a week may they spend on this? What can they decide alone? Can they progress a contract, or does everything need approval? Without answers, the lead cannot do much beyond encouraging people.

Fourth, put it into how people are assessed.

I have placed this last because it is the hardest. But without it, the first three do not hold. The person who uses it and delivers must not lose out. Add it to objectives, recognise people who share what works — the mechanism matters less than getting out of the state where using it depends on individual goodwill.

Do not leave the freed-up time unassigned

One more thing separates retention from results.

When generative AI cuts task time, where does that time go? Leave it undecided and efficiency stays efficiency. The "91.6% efficiency, 3.9% revenue" figure is largely this.

This is not something to leave to the floor. Set it as a departmental objective: where the freed-up time is redirected. More customer visits, better proposals, time on something new. The choice matters less than making one, because otherwise the time quietly fills itself.

There is more on the pilot-to-production decision in getting out of PoC limbo, and on the build-versus-buy line in how far to insource AI.

Teach the boundary, not the prompt

Finally, on training.

Generative AI training usually teaches prompt writing. That has its place, but what actually drives retention is the boundary.

How much to delegate and where to take over. How to verify output. When something goes wrong, whether to fix the prompt or accept that the task is a poor fit. When more people can make that call, the organisation becomes capable of using the tool.

Teach prompt technique alone and you start again every time the model changes. In the training we design through WARP, more time goes on the basis for judgement than on procedure. Tools change; the way you draw the boundary does not.

In summary

• Usage is no longer the problem. Intent to use is at 68.9%
• The gap is organisational. 27.0% report no organisational effort
• Effects concentrate on efficiency. 91.6% efficiency, 3.9% revenue and profit
• It stalls on four things: unshared practice, vague scope, no recognition, a part-time lead
• The order is write the envelope, circulate what works, resource the lead, reflect it in assessment
• Decide where the freed-up time goes, or efficiency is all you get
• Training should teach the boundary, not prompt technique

To talk through an internal rollout or a training design, get in touch.

References

Footnotes

1. 2026 Information and Communications White Paper, "Policies on generative AI use" (Ministry of Internal Affairs and Communications, Japanese). The combined share answering that their policy is to use generative AI actively or in limited areas was 68.9% in Japan (49.7% the previous year), 74.0% at large firms and 58.1% at SMEs. The note that Japan's increase was larger than in the other three countries comes from the same white paper ↩

2. 2026 Information and Communications White Paper, "Organisational efforts" (Ministry of Internal Affairs and Communications, Japanese). The figure of 27.0% of Japanese firms reporting no organisational effort around transforming work, described as a low level compared with the US, Germany, and China, comes from the same white paper ↩

3. Key points on DX and AI adoption trends among Japanese companies (IPA, 16 July 2026, Japanese). The reported effects of AI adoption, the usage breakdown, the 85.5% shortage of people to drive DX, and the state of AI-related talent all come from this material. 1,799 responses, fielded 17 April to 12 June 2026 ↩ ↩²