AI adoption for businesses in Japan
Gennai's 110 minutes: what the staff survey actually measured
The Digital Agency says Gennai saves government staff 110 minutes a week. The figure is real, but it covers fewer people and less certainty than the headlines suggest.
On October 30 the Digital Agency published the first staff survey on Gennai (源内), the generative AI platform it runs for Japan’s civil service. The headline number travelled fast. The Japan Times reported that employees at ministries and agencies had cut their working hours by 110 minutes a week, on average.
The number is real. It also describes a smaller group, measured more loosely, than that sentence suggests. I read the agency’s full results to see what sits underneath it.
Who answered the survey?
The survey ran from August 21 to September 18, 2026, as a Microsoft Forms questionnaire with 22 questions. It drew 6,439 responses: 2,467 from ministry headquarters and 3,972 from regional offices. Gennai is being rolled out to about 180,000 government staff, so roughly one in 28 of them answered.
Not everyone who answered uses Gennai. The agency’s own breakdown of the 6,439 is the most useful table in the report.
| How often respondents use Gennai | People | Share |
|---|---|---|
| Almost every day | 669 | 10.4% |
| Three or four times a week | 711 | 11.0% |
| Once or twice a week | 1,698 | 26.4% |
| Rarely | 1,687 | 26.2% |
| Registered but never used it | 325 | 5.0% |
| Never registered | 1,349 | 21.0% |
Add up the bottom three rows and 52% of respondents barely use Gennai or have never touched it. The news coverage put the figure at 31.2% by leaving out the 1,349 people who never registered at all.
What does the 110 minutes cover?
The 110 minutes is an average across the 3,078 people who use Gennai at least once a week. That is fewer than half of the respondents, and under 2% of the staff Gennai is meant to reach.
It is also an estimate built from opinion, not a measurement of hours. Staff were asked how much their weekly working time had changed on tasks where they used AI, and they picked a band. The agency then turned the bands into a number:
- “1 to 3 hours saved” counted as 2 hours, and the other ranges at their midpoints.
- “10 hours or more saved” counted as 10 hours.
- “About the same” and “don’t know” counted as zero.
- The few who said their time went up counted as a loss.
Here is what the 3,078 weekly users said.
| Change in weekly working time | Share of weekly users |
|---|---|
| Saved 1 to 3 hours | 35.1% |
| Saved 3 to 5 hours | 10.9% |
| Saved 5 to 10 hours | 5.2% |
| Saved 10 hours or more | 3.2% |
| About the same | 34.6% |
| Don’t know | 10.2% |
| Time went up | 0.7% |
Averaged that way, the bands come to 1.83 hours, or about 110 minutes. The arithmetic checks out. The method is reasonable for a first look, and the agency is careful to call the result an estimate.
What it supports is narrower than the headline: among the staff who already use Gennai every week, about 55% feel it saves them at least an hour, and about 45% feel it saves them nothing or can’t say.
What did staff say went wrong?
The agency’s report is more candid than the coverage of it. Among the 3,078 weekly users:
- 949 had run into answers that were inaccurate or unreliable, about 31%.
- 627 found the output could not be used as it stood and needed rework.
- 478 were not sure how to write a prompt.
- 304 were unsure which information they were allowed to enter.
- 1,136 reported no problems at all.
The people who rarely or never use Gennai were asked why. The most common answer, from 1,166 people, was that they don’t know how to use it. The third most common is the one I would underline: 994 said they regularly use a different AI service instead.
The agency also notes friction it did not expect. Some staff reported that colleagues who use AI end up taking work from those who don’t, and that AI output can create more to review.
Which AI is Gennai actually running?
Most of what staff described sounds like a chat assistant. The top benefits were faster first drafts, cited by 1,498 people, followed by catching up on unfamiliar subjects and pulling what they need from long documents.
According to the agency’s May 2026 overview, the chat tool let staff choose between four models as of March 2026: Amazon’s Nova Lite and Anthropic’s Claude Haiku 4.5, Claude Sonnet 4.5 and Claude Sonnet 4.6. Translation runs on Preferred Networks’ PLaMo. Five Japanese-made models were contracted in May for trial inside Gennai. The list may have grown since, and the survey doesn’t say which models respondents used.
That matters for a government that talks about sovereign AI. The chat tool behind most of these time savings offered staff a choice of American models.
Does this tell us Gennai’s own apps work?
Not directly. Gennai is a chat interface plus about 30 task-specific apps that the agency builds itself, such as legal research and help desks for internal systems. The survey doesn’t break results down by app, so it can’t show whether any particular one is reliable.
In April the agency published three of those apps as open source and invited local governments to reuse them. I reviewed that code. The legal research app can quote the wrong law and label it as official text. Set beside the 31% of weekly users who hit unreliable answers, that is the finding I would most want a local government to know before copying it.
What should a business take from this?
The survey is worth reading as a template, not just as news. A second round is due around January 2027, after the agency added an AI agent feature on October 20. If the next report publishes results by app and by model, it will be far more useful than a single average.
Any organisation rolling out AI can run the same check on itself:
- Count everyone, including people who never started. A satisfaction or time-saving figure from regular users says nothing about the half who stopped or never began.
- Ask what went wrong, not only what improved. The most useful lines in this report are the 949 unreliable answers and the 994 people using something else.
- Find out what people use instead. Staff who skip the approved tool are often using another AI service with your data in it. That is a governance problem before it is a productivity one.
- Separate perceived time from measured time. Self-reported savings are a fair first signal, but a decision about staffing or budgets needs more than a show of hands.
If you are weighing which AI tools your team should use, and what they should never be given, a Diagnostics review starts with the processes and data the tools would touch.