Japan gave 180,000 civil servants its own AI, Gennai. Half the staff it surveyed barely use it
Japan's Digital Agency built its own generative-AI platform, Gennai, and opened it to about 180,000 government staff. Its first survey estimates that users save about 1.83 hours a week. But 52% of the 6,439 staff who answered rarely or never use it, or have not even registered. We redid the agency's math from its published numbers and found the figure falls to about 53 minutes when spread across everyone who answered. We also look at the AI that drafts answers for parliament, the ¥4.4 billion budget, the monthly usage reports promised in April, and how the UK and France measured their own government AI.

"Are Japan's bureaucrats using AI as much as we do?"
That was the first question DATA TIMES editor-in-chief Kaito Ogasawara asked when he started reading the government's documents on AI. For a newsroom that works with AI and data, generative AI has become a tool we can't do without for research and transcription.
Japan's government has handed that tool to about 180,000 of its own staff. It is called Gennai, a generative-AI platform that the Digital Agency, the ministry in charge of digital government, built in-house. On October 30, 2026, ten days after adding AI agents, the agency also gave all 180,000 users a prototype AI that helps draft answers to questions in parliament.
The same day, the agency published its first usage survey. Of the staff who use Gennai, 71.7% said they were satisfied, and the agency estimates they save about 1.83 hours a week on the tasks they use it for.
But of the 6,439 staff who answered the survey, 3,361, or 52%, rarely or never use Gennai, or have not even registered.
Has the AI changed how Japan's government works? We put the agency's own documents next to the debate in parliament and the experience of the UK and France, then recalculated the survey numbers ourselves.
Gennai源内
An in-house generative-AI platform that Japan's Digital Agency built for government staff
- 2025.05Agency trialThe agency opened it to about 1,200 of its own staff. About 80% used it within three months
- 2025.12PM's top orderPrime Minister Takaichi made full use of Gennai the first of seven orders at the AI Strategy HQ
- 2026.05Government-wide trialA trial for about 180,000 staff at 39 bodies began, starting with about 100,000
- NOWAgents and a Diet AIIn October, AI agents and a prototype for drafting parliamentary answers were added
Gennai lets government staff use generative AI through a chat screen. Digital Agency engineers built it themselves, and it runs on Japan's Government Cloud. Besides drafting, summarizing and translation, it offers about 30 AI apps built for government work, such as one that searches past parliamentary answers and one that researches laws. Staff can choose among several overseas and Japanese AI models.
It covers about 180,000 people at 39 bodies, every ministry and agency including their external bureaus. That is roughly 30% of Japan's 590,000 national public servants. The name plays on "Gen AI" and honors Hiraga Gennai, a famous inventor of the Edo period.
Since April 2026, part of Gennai's code has been published as open source, free for anyone to use, so that local governments and companies can build something similar.
There are three reasons to check the numbers now.
First, real usage data has appeared for the first time. The trial for 180,000 staff began in May 2026, and the first survey results came out on October 30. Second, Gennai is turning from a chat tool into a tool that does work. AI agents arrived in October, and the parliamentary-answer AI is scheduled for full use from the ordinary Diet session in January 2027. Third, the money question has started. On October 8, Prime Minister Sanae Takaichi told the lower house that the government would continuously track and verify Gennai's cost-effectiveness.
Using AI is only half of our job. The other half is measuring what it changed. Is the AI a company handed out actually being used? How much time did it save? Every organization that rolls out AI runs into these questions after the rollout. Here is what we found, in four parts: why the state did it, what the time figure really means, why so many staff don't use it, and how other governments measured it.
Why Japan gave AI to 180,000 staff

It started as a small experiment inside the Digital Agency.
In May 2025, the agency opened its home-built AI platform to its own staff. Within three months, about 950 of roughly 1,200 employees had used it, more than 65,000 times in total. But the same report also showed that half of the division directors had never used it. Young staff and hires from the private sector used it most; managers used it least. In April 2026, Tomoya Hasui, a director-general at the agency, told an upper house committee that one reason was that most apps were general-purpose tools like summarizing and manual search, and did not always fit managers' work of making policy decisions.
The government then decided to roll it out everywhere.
At the AI Strategy Headquarters on December 19, 2025, Prime Minister Takaichi issued seven orders. Gennai was the first.
First, make full use of Government AI Gennai. From May next year, more than 100,000 government staff will be able to use it. Use AI Gennai to work creatively, and show the public what trustworthy AI means. ── Prime Minister Sanae Takaichi, AI Strategy Headquarters (December 19, 2025; our translation)
The Artificial Intelligence Basic Plan, revised in July 2026, names Gennai and describes it as AI use by "as many as 180,000 government staff, the largest scale in Japan."
Three goals: fewer hands, a safe place, and the state as first customer
The government had three main goals.
1. Use AI to make up for shrinking staffGovernment work is not shrinking, but the workforce is. In April 2026, then-Minister for Digital Transformation Hisashi Matsumoto told the upper house budget committee that the government would "make up for this staff shortage with AI," starting with drafting answers for parliament. The Digital Agency's documents call that one of the heaviest tasks for staff, demanding "high accuracy and consistency within limited time." Staff must check the question, search past answers and supporting documents, and write a draft, all in a short window.
2. Give staff a safe place to use AIPublic consumer AI services can't be fed information that must stay inside government. Under government rules, services you can use just by accepting the terms of service generally can't handle confidential information. Gennai runs on a government-managed cloud, and Digital Agency staff can enter internal information up to a level called "Confidentiality 2."
3. Make the state the first customerGennai offers Japanese AI models alongside overseas ones. The agency picked seven of 15 applicants and signed contracts with five. From September to November 2026, it has been running a blind test in which staff pick the better answer without knowing which model wrote it. We covered how this connects to the trade ministry's program for building Japanese AI in our piece on GENIAC.
The money: ¥4.4 billion in a supplementary budget
The government set aside ¥4.4 billion for a "Government AI development project" (roughly $29 million at ¥150 to the dollar) in its supplementary budget for fiscal 2025. The Digital Agency's list of no-bid contracts shows the main Gennai contracts signed in April and May 2026.
| Project | Contractor | Amount |
|---|---|---|
| Large-scale rollout trial for 180,000 staff | KDDI | about ¥500 million |
| Parliamentary-answer AI development | NTT DOCOMO BUSINESS | about ¥250 million |
| Government AI app development | NTT DATA | about ¥250 million |
| Research, collection and processing of a large government dataset | TOPPAN | ¥670 million |
| Support for lead divisions on Government AI | Gartner Japan | ¥550 million |
Source: Compiled by DATA TIMES from the Digital Agency's no-bid contract disclosures for April 2026 and the disclosures for May 2026. All were awarded after a proposal competition. Other contracts include test-environment design and development (NTT DATA, ¥390 million) and collecting official gazette data (National Printing Bureau, ¥53 million).
Before that, Gennai was an in-house test project. According to the government's administrative project review, its predecessor spent about ¥220 million in fiscal 2023, ¥360 million in 2024 and ¥50 million in 2025. The ¥4.4 billion is an order of magnitude more.
Divide ¥4.4 billion by 180,000 and you get about ¥24,000 per person. But that is money for building things, including data preparation and the parliament AI. It is not the yearly cost of using AI. Toshiharu Furukawa, Minister for Digital Transformation, told the lower house budget committee on October 14 that the cost of AI usage, billed per token, was included in the fiscal 2027 budget request. Yet the budget request document lists Gennai's costs only as "IT system costs plus an unspecified request," with no figure for Gennai alone. How much it costs per person per year can't be calculated from public documents.
We redid the math on “1.83 hours a week”

So how much easier did Gennai make the work? The Digital Agency's answer: about 1.83 hours (110 minutes) a week per user.
The survey ran from August 21 to September 18, 2026, among staff at ministries with access to Gennai. It drew 6,439 answers, 2,467 from central ministries and 3,972 from regional offices. Only the 3,078 who use it at least once a week answered the questions about time.
How the 1.83 hours is calculated
Staff chose a range for how their weekly hours changed on the tasks where they used Gennai, such as "1-3 hours saved" or "3-5 hours saved." According to the agency's footnote, the 1.83 hours rests on these rules:
- "1-3," "3-5" and "5-10 hours" are counted at their midpoints: 2, 4 and 7.5 hours
- "10 hours or more saved" is counted, conservatively, as 10 hours
- "Almost no change (under one hour either way)" and "don't know" count as zero
We applied those rules to the published head counts. The total time saved comes to 5,650 hours a week. Divided by 3,078 users, that is 1.836 hours, almost exactly the agency's figure. The method is public, and anyone can check the number from outside.
Depending on the assumptions: 1.24 to 2.43 hours
Still, some choices in the calculation move the result.
Use the bottom of each range instead of the middle (one hour for "1-3 hours") and the average is 1.24 hours a week. Use the top (three hours) and it is 2.43 hours. Drop the 313 who answered "don't know" from the count and it is 2.04 hours. None of these is wrong. We think the 1.83 hours is best read as an estimate in the middle of a range of about 1.2 to 2.4 hours.
A small group lifts the average. The 8.4% of users (259 people) who said they saved five hours or more account for 39% of all time saved. The user in the exact middle falls in the "1-3 hours saved" range, but 45.5% of users saved less than an hour. For a typical user, a little over an hour a week seems closer to reality.
One more small discrepancy: the agency's announcement says 54.4% of users "felt a saving of an hour or more," but the head counts give 1,677 of 3,078, or 54.5%. Both numbers appear in its documents. The difference comes from adding up percentages that were already rounded.
Spread across everyone who answered: 53 minutes
The biggest swing comes from a simple question: who do you divide by?

The 1.83 hours is divided by the 3,078 people who use Gennai at least weekly. The other 3,361 respondents rarely or never use it, or have not registered. Count their savings as zero and divide by all respondents, and the figure is about 53 minutes a week (our estimate).
The respondents are just 3.6% of the 180,000. People who answered may use the tool differently from those who didn't, so the 53 minutes won't necessarily hold for all 180,000 either. What is clear is that "per user" and "per employee" differ by about a factor of two.
Saved time goes into other work
There is something else the time figure can't show.
When asked what share of their time changed on Gennai-assisted tasks, 59.1% of users said "almost no change (within 10%)." Yet when asked what improved, the top answer, chosen by 1,498 people, was that they could produce first drafts quickly and no longer had to start from scratch. Next, chosen by 1,088, was getting up to speed faster in unfamiliar fields.
One comment in the survey stood out.
It didn't cut my working hours, but we drafted guidelines in-house in about three months, without contractors, work we used to outsource for six months to a year. But this is work we used to outsource and can now do ourselves, so my own workload went up, and my working hours did not go down.
The agency's summary page quotes only the first half of this comment and leaves out "my working hours did not go down." We think the second half shows Gennai's real effect. Bringing outsourced work in-house may have saved contracting money. But staff time didn't shrink. AI's payoff may show up less in hours saved than in the range of work staff can take on.
The promised "monthly report" is nowhere to be found
How to measure the effect has also come up in parliament.
On April 1, 2026, in an upper house committee, Takahiro Anno of Team Mirai warned that the large trial would be "meaningless if it ends as a mere trial," and asked whether the government would publish cross-government KPIs and a results dashboard. Then-minister Matsumoto replied that the KPIs and the dashboard would be made public.
At a press conference on April 17, Matsumoto gave a concrete plan. Starting around the summer, the agency would publish a monthly report for the whole government and for each ministry, with figures such as the share of account holders who had used Gennai at least once, the share who used it in a given month, and the number of uses per person.
But when we checked the Digital Agency's Gennai pages and its list of announcements on November 14, we found no such monthly report. The only usage figures available are from the first survey, published October 30. Usage rates counted from system logs are still not visible from outside. That is the biggest blank in judging Gennai's results.
The government's project review holds another number. In the review of fiscal 2025, the Digital Agency rated the "share of time saved" on tasks using its AI apps at 58%, against a target of 20%. But that year's tests were mostly inside the Digital Agency, and the subjects and method differ from the staff survey. The same tool can look very different depending on which number is presented as the result.
Why half don't use it, and the parliament AI

Next, the staff who don't use Gennai.

Of respondents, 52% rarely or never use Gennai, or have not registered. Only 10.4% use it almost every day, 11.0% three or four times a week, and 26.4% once or twice a week.
The gap between job grades is small. From senior executives to junior staff, about half of each grade uses it at least weekly. In 2025, half the Digital Agency's division directors had never used it, so the pattern of managers lagging behind seems to have faded, at least among respondents.
Note that this is a share of the people who answered. It doesn't mean half of the 180,000 don't use Gennai. If those who answer surveys tend to be more interested in the tool, usage among those who didn't answer may be even lower.
The top reason: "I don't know how to use it"
When non-users were asked why (3,361 people, multiple answers allowed):
- I don't know how to use it: 1,166
- I regularly use a different AI service: 994
- I don't see the need for Gennai in my work: 925
- My work doesn't call for AI: 547
- I don't have a Gennai account: 271
Worries about output quality (230) and information security (207) were less common than we expected. The main barriers are know-how and need, not distrust. Among the unregistered, "I don't know how to use it" was the top reason at 34.1%. Among those who hardly use it, the top reason was regularly using a different AI service.
Some comments read: "I didn't even know Gennai existed," and "I'm busy with deadlines every day and got no instruction at all, so I would have to learn it from scratch on my own." One person handling parliamentary work said they can't "wait 10 to 15 minutes every time I run a prompt" when they need to find past answers fast.
Asked what support they want, respondents overall split almost evenly between a "collection of use cases from my ministry and others" (2,372) and "basic training" (2,276). Among those who hardly use it or have not registered, basic training was the top request.
"Other AI" and the shadow AI rule
The second reason, "I regularly use a different AI service," needs some explanation.
The Gennai Standards, published October 20, list "no use of shadow AI" among staff rules: staff must "not use generative-AI services that have not been approved." The examples include asking a personal chat AI to analyze non-public work documents.
But the survey doesn't say the "different AI service" is unapproved. Comments include "I'm used to Copilot, so it's hard to switch" and a request for "material that explains, with concrete cases, when to use Copilot and when to use Gennai." Many seem to use another AI their ministry has approved. The 994 can't be read as a count of shadow-AI users.
Users' worries: accuracy, and what they may enter
Among the 3,078 users, the top problem, cited by 949, was "answers that were inaccurate or unreliable at times." Next came "had to edit the output before using it" (627), "slow or error-prone" (428) and "unsure what information I was allowed to enter" (304). The most common answer, though, was "no particular problems" (1,136).
The comments are specific.
If you doubt the answers and fact-check everything, it ends up taking time anyway. And if it's something I can fact-check myself, I could often solve it without Gennai in the first place. I'd like to know a practical way around this.
Others wrote that "some things I couldn't tell were confidential or not, so I sometimes gave up on Gennai," and that its training data seemed to stop "around 2024."
The agency says the AI-agent environment added on October 20 answers some of these complaints. It shows the URLs or files behind each answer, handles files of several hundred megabytes and formats like Word and Excel, and can search the web for current information. A second survey around January 2027 will check what changed.
What the parliamentary-answer AI does
The prototype released on October 30 is the most closely watched app in Gennai.
According to the Digital Agency, it helps staff check a question, find and organize past answers and supporting documents, draft an answer, and verify the content. Development began in April 2026 and was contracted to NTT DOCOMO BUSINESS. A panel of 102 experienced staff recruited from 16 ministries tested it, and about 80% of them said they wanted to use it.
The agency's document states that the system "does not hand staff's judgment over to AI." Staff decide how to interpret policy and what the answer says, and ministers deliver the answers.
Parliament has voiced both hopes and conditions. In an upper house plenary session on May 27, 2026, Mizuho Umemura of the Sanseito party backed the rollout, quoting the proverb "start with yourself," but asked that AI not make answers more formulaic. Answering her, then-minister Matsumoto revealed that his answer that day had been drafted with Gennai.
For this answer of mine, staff created the draft in Gennai, staff checked the facts, and then I did the final review and approval before answering. ── Hisashi Matsumoto, then Minister for Digital Transformation, House of Councillors plenary session (May 27, 2026; our translation)
Earlier, on May 12, Akira Nagatsuma raised a different concern in a lower house special committee. He said using AI to check past answers for consistency was fine, but worried that training it on government answers would teach it to "learn nothing but evasive answers".
In the same committee on May 8, Matsumoto said that if you enter a question into Gennai, a two-page draft answer comes out in three minutes, but parts still differ from the facts, and staff check it before it becomes an answer. The draft takes minutes; the checking is human. How much this AI helps will depend on how much it cuts that checking.
The real time sink may be the waiting
Some outside voices are sharply critical. Yukio Noguchi, professor emeritus at Hitotsubashi University, wrote in January 2026 that what makes parliamentary work so grueling is not writing the answers but "long waits caused by late notice of questions and coordination between departments". Making the writing faster, he argued, won't fix the root problem unless the way of working changes.
The government is aware of this. On June 3, Matsumoto told the lower house cabinet committee that notice of questions had become much earlier than before, and asked lawmakers to cooperate further. At the October 30 press conference where the prototype was released, Minister Furukawa addressed the fear of relying too much on AI.
It's important that people who use AI always keep a critical eye on whether its answers are right, and build a relationship in which both sides get better. ── Toshiharu Furukawa, Minister for Digital Transformation, press conference (October 30, 2026; our translation)
How the UK and France measured government AI

Other governments are also handing AI to their staff and trying to measure the effect. Comparing them shows how to read Gennai's numbers.
The UK's "26 minutes a day" also used range midpoints
From September to December 2024, the UK government gave Microsoft 365 Copilot to about 20,000 staff across departments. Its June 2025 report found users saved an average of 26 minutes a day. That figure is also self-reported and was calculated using "the middle point of each range." It is built the same way as Gennai's 1.83 hours.
Individual departments then ran their own, more careful evaluations.
- Department for Work and Pensions: From October 2024 to March 2025, it gave licenses to 3,549 staff, compared users with non-users, and estimated 19 minutes saved a day on eight routine tasks. The report notes that licenses were not allocated randomly but largely to volunteers and nominees, which may lead to overestimating the benefits.
- HM Revenue and Customs: From September to December 2024, it allocated 3,000 licenses at random. Because not every license holder used Copilot, it cut self-reported savings by about 20% and reported 2-3% of the working week, about 60 minutes.

Side by side, Gennai's 110 minutes a week is in the same range as the UK figures. But all three UK numbers are also self-reported. With the same AI, differences in how licenses were handed out and how answers were counted produced results from 60 to 130 minutes a week, about a twofold spread.
France got out a stopwatch
France went further.
From December 2025 to June 2026, the French government trialed an AI assistant for state employees with 10,000 licenses. In the evaluation by the Interministerial Directorate for Public Transformation (DITP), 66% of survey respondents said it saved them time.
DITP then asked about 30 volunteers to do the same tasks with and without AI, and timed them. The result: average savings of 2.5 and 5.4 minutes per task on two tasks. The report warns that these results can't be generalized to all civil servants and should be read with caution.
The 66% is a share of people who felt they saved time; the 2.5 and 5.4 minutes are savings per task. They measure different things. But France ran a survey of perceptions and a timed test as separate exercises.
In Japanese companies, too, only one in four cut their hours
Private-sector surveys show a similar picture.
A February 2026 survey by Persol Research and Consulting of full-time employees found that tasks done with generative AI took 16.7% less time on average. But only one in four users had actually cut their working hours, and only about one in ten used it as a daily tool.
Tasks get faster, but total working hours are slow to fall. It is the same thing the Gennai guidelines comment describes.
Handing it out: pass. Next: measure, then delegate
Put the Digital Agency's documents, the debate in parliament, the UK and French cases and our own recalculation side by side, and Gennai's results split cleanly into two: handing it out, and actually changing the work.
Handing it out: pass. Changing the work: one self-reported survey so far
In little over a year after the internal trial, the agency opened Gennai to about 180,000 staff at 39 bodies and added AI agents. For a government IT system, I think that is remarkably fast. Publishing the code so local governments and companies can build their own is also a sensible way to give back what taxpayers paid for. On rollout, I'd give it a pass.
But the only evidence on whether it has changed government work is one self-reported survey. The 1.83 hours is a midpoint estimate, and spread across all respondents it is about 53 minutes. UK departments' figures varied twofold with how licenses were handed out and answers counted, and France's stopwatch found a few minutes per task. I'm not saying Gennai's number is inflated. But feeling that time was saved and actually saving time are not the same. The cost-effectiveness review the prime minister promised will need both usage logs, like the monthly report promised in April, and tests that actually time the work.
The top reason for not using it is know-how. The upside lies in colleagues' methods
The top reason for not using Gennai was not knowing how, not distrust of its performance. In 2025, half the Digital Agency's division directors had never used it; among this year's respondents, the gap between job grades has almost disappeared. There is still plenty of room to grow.
Users most wanted a collection of other ministries' use cases; non-users wanted basic training. Before the end of the fiscal year, the agency plans to add the Gennai Library, which lets staff share procedures they built in a form AI can reuse. If one person among 180,000 finds a good way to use it, the next department can use it as is. Alongside training, I think that's the fastest way to bring more people in.
The parliament AI will succeed or fail outside the answer itself
Drafting answers for parliament is one of the heaviest jobs staff do, and an AI that produces a draft in minutes is a real help. But as Matsumoto himself acknowledged, staff check what comes out. And as Noguchi pointed out, the burden of parliamentary work comes not only from writing answers but from waiting for questions and coordinating between departments.
So the parliament AI should be judged in the January 2027 Diet session by how long it takes to finish one answer, waiting time included, before and after. The more work we hand to AI, the more we need to measure it. Having given AI to 180,000 staff, can the government show that next? We'll check against the second survey and the full launch of the parliament AI.
SOURCES ── References
- Digital Agency, "Results of the first Government AI Gennai usage survey" (October 30, 2026) (in Japanese)
- Digital Agency, "Recent developments in Government AI Gennai," Material 5, Study Group on National Public Servants in the AI Era, 3rd meeting (November 9, 2026) (in Japanese)
- Digital Agency, "Prototype of the parliamentary-answer drafting AI app released on Government AI Gennai" (October 30, 2026) (in Japanese)
- Digital Agency, "AI agent environment launched on Government AI Gennai and Gennai Standards published" (October 20, 2026) (in Japanese)
- Digital Agency, "Government AI Gennai Standards on Governance, Security and Safety," 1st edition (October 20, 2026) (in Japanese)
- Digital Agency, "Generative AI use by Digital Agency staff" (August 29, 2025) (in Japanese)
- Prime Minister's Office, "AI Strategy Headquarters, December 19, 2025" (in Japanese)
- Cabinet Office, "Artificial Intelligence Basic Plan" (Cabinet decision, July 14, 2026) (in Japanese)
- Digital Agency, "Large-scale trial of Government AI Gennai" (March 6, 2026) (in Japanese)
- Digital Agency, "Large-scale trial of Government AI Gennai for about 180,000 staff at all ministries begins," overview (May 28, 2026) (in Japanese)
- Digital Agency, "Outline of the FY2025 supplementary budget" (December 17, 2025) (in Japanese)
- Digital Agency, "Disclosure of no-bid contracts (April 2026)" (in Japanese)
- Digital Agency, "Disclosure of no-bid contracts (May 2026)" (in Japanese)
- Digital Agency, "Outline of the FY2027 budget request (main projects)" (August 31, 2026) (in Japanese)
- Digital Agency, "Results of the call for Japanese large language models to trial in Government AI" (March 6, 2026) (in Japanese)
- Digital Agency, "Trial of Japanese foundation models on a domestic cloud in Government AI Gennai" (July 10, 2026) (in Japanese)
- Digital Agency, "Government AI Gennai released as open-source software" (April 24, 2026) (in Japanese)
- Digital Agency, "Guidelines on the procurement and use of generative AI for the advancement and innovation of public administration," version 2.0 (June 12, 2026) (in Japanese)
- Digital Agency News, "What is Government AI? The government's AI strategy and Gennai" (December 11, 2025) (in Japanese)
- Diet record, House of Councillors plenary session, question by Mizuho Umemura and answer by Hisashi Matsumoto (May 27, 2026) (in Japanese)
- Diet record, House of Representatives special committee on regional revitalization, children's policy and digital society (May 8, 2026) (in Japanese)
- Diet record, same committee, remarks by Akira Nagatsuma (May 12, 2026) (in Japanese)
- Diet record, House of Councillors Audit Committee (June 15, 2026) (in Japanese)
- Diet record, House of Representatives plenary session, answer by Prime Minister Sanae Takaichi (October 8, 2026) (in Japanese)
- Diet record, House of Representatives Budget Committee, answer by Toshiharu Furukawa (October 14, 2026) (in Japanese)
- Digital Agency, "Press conference by Minister Furukawa (October 30, 2026)" (in Japanese)
- Diet record, House of Representatives Cabinet Committee, answer by Hisashi Matsumoto (June 3, 2026) (in Japanese)
- Yukio Noguchi, "Why the Gennai AI for drafting parliamentary answers is missing the mark," Gendai Business (January 15, 2026) (in Japanese)
- Diet record, House of Councillors special committee on digital society and AI, question by Takahiro Anno, answers by Hisashi Matsumoto and Tomoya Hasui (April 1, 2026) (in Japanese)
- Digital Agency, "Press conference by Minister Matsumoto (April 17, 2026)" (in Japanese)
- Administrative Project Review, "Government AI (Gennai)," project no. 007632 (published September 1, 2026) (in Japanese)
- UK Government, "Microsoft 365 Copilot Experiment: Cross-Government Findings Report" (June 2, 2025)
- UK Department for Work and Pensions, "An evaluation of DWP's Microsoft 365 Copilot trial" (January 29, 2026)
- HM Revenue and Customs, "Evaluating the impact of Microsoft Copilot in HMRC" (July 9, 2026)
- DITP (France), "Synthèse de l'évaluation de l'expérimentation de l'Assistant IA" (June 2026) (in French)
- Persol Research and Consulting, "Survey on generative AI and work" (February 3, 2026) (in Japanese)
- DATA TIMES recalculation based on the Digital Agency's first usage survey (November 14, 2026)


