Genai and GovTech Tokyo: getting a whole workforce to use AI
遠くまで行くために、組織全員がAI活用できる未来へ - 源内、生成AIプラットフォームの歩み
B · Useful, limited First-hand Code for Japan Summit panel with the builders of Japan's Digital Agency platform Genai and GovTech Tokyo's staff app platform; rare public-sector rollout detail, but all figures are spoken and self-reported.
Summary based on the full transcript·Original: about 45 minutes to watch
In short
A 45-minute panel at Code for Japan Summit 2026 (published 22 October 2026). Naoya Osugi leads the Digital Agency's Genai platform; Junichi Hashimoto leads GovTech Tokyo's platform on which staff build their own AI apps; Yuki Kawabe of Code for Japan moderates and describes his own organisation's use. They cover what they built, how use spread, and where staff-built apps must hand over to engineers. The strongest finding from GovTech Tokyo: municipalities progress when one to three people have official work time for AI and organisational backing.
Key numbers
- ~180,000central government staff the Prime Minister ordered to use Genai
- 175,000 / 40,000+staff able to use Genai versus staff actually using it
- ~4,800staff-built AI apps on GovTech Tokyo's platform by end of March 2026, across 55 municipalities and about 30 Tokyo government workspaces
- ~95%active share of Code for Japan's fewer than 20 staff after standardising on Claude
As reported by the source.
Japanese·Video·opens youtube.com
What it reports
- Genai's design aim is replaceability: accept lock-in at the interface, keep applications, inference (the main cost) and data separable. OSS v1.0 shipped around May 2026 but is still tied to specific AWS services; v2.0 adds agent features.
- New Genai products: "Agent chat", a near copy of the Claude desktop app with model switching and a file drive, already widely used in the Digital Agency; and "Genai Kobo", a Claude Code-style agent shipped inside an isolated, rule-hardened environment, planned for distribution as early as this year.
- The Prime Minister ordered about 180,000 government staff to use Genai, so distribution and permission were settled from the top. About 175,000 can use it; more than 40,000 do. The team prioritises product appeal before training, following a retention-before-activation model.
- GovTech Tokyo's platform began bottom-up in 2024 as a side role, with unclear release criteria and too many stakeholders. About 4,800 apps existed by end of March 2026. Copilot covers personal productivity; the platform targets shared, explicit apps from tacit know-how.
- Code for Japan wrapped every contracted SaaS as an MCP server. Left alone, use drifted: one person consumed 18 times more than the next. Usage analysis, workshops and background workflows corrected course over about three months; team-level use remains weak.
- Osugi: production quality requirements for staff-built web apps have risen, not fallen; a Claude Code security review flagged an unfixable risk for an app open to the public. Engineers and possibly lawyers are needed early. A police SNS-monitoring prototype he built has run for over 1.5 years and is moving to a vendor.
- Neither platform has a way to inventory or retire abandoned apps. Single-year budgets mean the OSS cannot promise maintenance; Osugi says he does not know if his own role continues next year.
- A temporary Genai for municipal staff went live the day after the Kumamoto earthquake because the disaster rationale had been agreed in advance.
How they got it
Spoken accounts from the platform leads, with some figures read from GovTech Tokyo's annual report slide. The transcript is from auto-generated Japanese captions with garbled proper nouns.
Limits
- All numbers are self-reported and spoken; "using" is not defined.
- App counts include apps of very different quality and say nothing about retained use or value.
- No measurement of time saved; Code for Japan names this as an open gap.
- Auto-caption transcript; some passages are hard to parse.
Why it matters for companies
The panel gives the usual enterprise rollout lessons from an organisation larger than most firms. A top-down mandate solved access but not use; usage depended on a product people wanted and on named people with protected time. Citizen-built apps hit a hard line at public exposure, where security, accessibility and legal review return. Any company running a staff app platform should plan inventory and retirement before the app count reaches thousands.
Key passages in translation
Japanese original
その担当の方1人でも2人でも3人でもいいんですけど このAIを活用するっていうことに業務として当たれる時間があるっていうことですね
Whether it is one, two or three people in charge, it's that they have time they can spend on using AI as part of their actual job.
Japanese original
プロダクションの品質のハードルって別に下がってないしむしろ上がる一方じゃないすか
The bar for production quality hasn't come down at all; if anything it keeps rising.
Japanese original
放置されちゃってるものを何か棚卸するとか メンテナンスするみたいなことは まだあまりできる仕組み的にはできてない
We don't yet have a mechanism to inventory or maintain things that have been left abandoned.
Summary based on the full transcript. Read the original for the full argument and data.Watch the original on youtube.com