GENAI: Japan's government-built AI platform for its civil service, from drafting parliamentary answers to everyday paperwork

The problem

Japan's public service is being asked to do more with fewer people. A declining birthrate and an ageing population are steadily shrinking the workforce, and the Digital Agency has warned that sustaining public services is becoming harder as labour shortages deepen.

That pressure falls hardest on work that is essential but heavy on skilled staff time, and one task stands out. Ahead of sittings of the Diet, Japan's national parliament, members submit their questions to the government in advance. Officials across the ministries must then draft an answer that is accurate and consistent with the positions the government has taken before, usually against a tight deadline, while also preparing for the follow-up questions that may come in the chamber. The volume and the deadlines have made it a long-standing source of heavy overtime for public servants.

The solution

Japan's Digital Agency, the body responsible for the country's digital transformation, built a generative AI platform for civil servants called GENAI, an English rendering of "generative AI" pronounced "Gennai", after Hiraga Gennai, an inventor and polymath of Japan's Edo period, 1603 to 1867. The platform is the working expression of a national policy: Japan's 2025 AI law and its December 2025 AI Basic Plan established that the government should lead by example in adopting AI, both to build civil servants' skills and to encourage private investment by demonstrating the technology at work inside government.

The agency developed the platform in-house and did not procure a commercial product. It had run generative AI workshops for central government staff since 2023, and those had shown two things: testing AI in real administrative settings required considerable adjustment, and its usefulness could not be verified in any environment where confidential information could not be handled. Building the platform itself let the agency meet government-wide security standards while still allowing officials to work with sensitive material, which was the condition that made the tool usable for genuine government work.

GENAI gives staff a single interface that connects to the Government Cloud and draws on government-wide datasets, including legislation and the government's record of official notices, alongside knowledge held by individual ministries. It offers more than 30 applications in two groups: general tools for everyday work, such as chat, drafting, proofreading, summarising and translation, and applications built for administrative practice. Reported functions across the platform include document generation, meeting minutes, transcription, and searches of past parliamentary answers.

The flagship application prepares the government's answers to questions from the Diet, Japan's parliament. It analyses the submitted questions, searches past answers and supporting documents, drafts a reply consistent with the government's earlier positions, anticipates likely follow-up questions and checks the draft for contradictions. To keep those answers dependable, this and other administrative applications use retrieval-augmented generation, meaning the system first retrieves the relevant official documents and then generates its answer from their contents, which keeps the output tied to the current wording of a law or a past response. In a government setting, where the exact current text is what matters, this is the feature that makes the output reliable enough to use.

Two further choices shape how GENAI manages risk. The whole platform runs in a closed, government-only environment that meets government-wide security standards, so officials' inputs and the documents they work with stay inside government systems. The agency presents this as the way to protect confidential material, with access controlled through the government's single sign-on and the environment cleared for a defined tier of confidential information.

On the models themselves, GENAI lets users select from several. It has used foreign systems, including through a collaboration with OpenAI, while the agency has issued a public call for domestically developed large language models for trial use. Its reasoning is that models attuned to the Japanese language and values suit government work, and that developing domestic capability is a matter of national AI autonomy as well as of safety and performance.

The Digital Agency’s guiding principle is to redesign processes and data around AI, so the technology reshapes how a task is done. It notes that limited use of chat-style tools on their own has produced little fundamental improvement. Adoption is governed accordingly: each ministry is expected to establish a framework to promote use and strengthen oversight, to follow the government's procurement and use guidelines, and to place management under a Chief AI Officer, while the agency builds a dashboard to show how AI is being used across government.

The agency deployed GENAI in steps. It ran the platform within its own offices from May 2025, opened it to a few hundred staff in selected ministries from January 2026, and began a large-scale pilot in May 2026 that reached around 100,000 employees at the outset, with the aim of covering roughly 180,000 across all 39 national agencies by the time it closes in March 2027. A full, government-wide launch is planned for the 2027 fiscal year. In April 2026, the agency also released part of the code as free open-source software under a licence permitting commercial use, so that local governments can stand up a comparable environment and private firms, including smaller companies, can build services on the same base. For another public servant, this is the most directly reusable part of the story: the code itself is available to adopt, so a similar capability can be built without starting from scratch.

The impact

GENAI is at the large-scale pilot stage, so the evidence so far is about how widely it has been taken up and how early users rate it, rather than measured gains across government.

1. It is being rolled out across all 39 national agencies, and early users rate it as useful
From late May 2026, about 100,000 employees had access, with the aim of reaching roughly 180,000 across the agencies before it closes in March 2027; the agency describes extending it to all staff, including regional and local offices, only as a longer-term goal. In the earlier phase inside the Digital Agency, which began in May 2025, a survey of participating staff found that about 80% felt the platform had helped their work.

Lessons learned

• Building in-house can satisfy security requirements that commercial tools cannot. Because the platform runs in a closed government environment, officials can process confidential material through it without that material leaving government systems, which the agency treated as a precondition for using AI on sensitive work at all.
• Ground the system in current official documents rather than the model alone. Since laws and prior positions change, a government assistant needs to retrieve and answer from live documents, which is what keeps its output dependable for administrative tasks.
• Design the work around the AI rather than attaching it to existing processes. The agency's stated position is that rethinking processes and data is where the benefit lies, and that adding tools to unchanged workflows has underperformed elsewhere.
• Roll out in stages and confirm value before committing. Progressing from an in-house start to selected ministries, then a large-scale pilot, then a full launch allowed the agency to gather evidence at each step rather than deploying government-wide at once.
• Publishing the build lowers the cost for others to follow. Releasing part of the code as open-source software allows other governments and firms to establish a comparable capability without building from scratch.

This case study was written with assistance from artificial intelligence.