1,050 Municipalities, 550,000 Civil Servants: OpenAI Spotlights Japan's QommonsAI as a Blueprint for Public-Sector AI
Polimill's QommonsAI — built on GPT models and already used by ~1,050 Japanese municipalities and ~550,000 public employees — is evolving into a 'public OS', with a Qommons ONE app store and super agent landing this fall.
On Monday, September 1, OpenAI published a customer story that reads less like marketing and more like a preview of how AI could quietly absorb the back office of an entire country. The subject is Polimill, a Japanese startup whose generative AI platform QommonsAI now serves roughly 1,050 municipalities and about 550,000 public employees across Japan — and which is preparing its most ambitious expansion yet: a “public OS” complete with an app store and an orchestrating “super agent,” slated for full rollout in fall 2026.
The numbers alone make this one of the largest public-sector AI deployments anywhere in the world. Japan has 1,718 municipalities in total, meaning QommonsAI has penetrated well over half of them. Behind the adoption curve is a structural story about labor shortage, fragmented data, and a startup that decided the best way to amplify citizen voice in policy was first to make government itself run faster.
From civic forum to government AI layer
Polimill did not start out as a government AI vendor. Its first product, Surfvote, is a social platform where citizens exchange views on political and social issues — a space designed to surface “loose consensus,” the common ground between people with opposing positions, rather than amplify division.
But as the company worked with local governments, it hit a structural wall: public-sector teams were so consumed by daily operations that they had little time to reflect citizens’ voices in policy. If the goal was civic participation, the prerequisite was administrative breathing room. So Polimill pivoted its ambitions inward — into the machinery of government itself — and released QommonsAI in October 2024.
Today the platform provides specialized AI across a roster of distinctly unglamorous but critical domains: drafting assembly responses, public services, social welfare, and legal search. In April 2026, Polimill began offering GPT-5.4 through Azure OpenAI Service from domestic Japanese regions, with 300 million tokens per month offered free to municipalities — and an LGWAN version (Japan’s closed government network) provided at no charge.
The hard part: unifying Japan’s fragmented administrative knowledge
The most instructive part of the OpenAI story is what Polimill says was the biggest obstacle to generative AI in government: not model quality, but data fragmentation. Every Japanese municipality has its own workflows, document formats, and scattered historical records. Preparing a single assembly response can require manually reviewing years of minutes to ensure an answer is consistent with prefectural policy. Even the most capable model produces nothing practical if the underlying knowledge isn’t organized.
Polimill’s answer was to do the unglamorous work first: it collected and standardized assembly minutes from across Japan, then used AI to add metadata and build a high-precision search foundation that works across municipalities and time periods. That structure was then extended beyond assemblies into welfare, laws, and other administrative domains — all surfaced through a common interface, QommonsUI, which also enables cross-municipality collaboration.
It is a point worth dwelling on for anyone building enterprise AI: the moat here is not the model. GPT models are available to everyone. The moat is the standardized corpus of Japanese administrative documents and the cross-organizational search infrastructure built on top of it. Polimill’s bet is that this layer — not any single model — becomes the durable asset.
Why GPT, and why the ChatGPT brand matters in government
QommonsAI’s model core is OpenAI’s GPT series. Polimill’s CAIO, Masahiro Wakabayashi, cites two reasons that go beyond benchmark scores.
First, operational control: public-sector deployments need information management and audit readiness, and QommonsAI includes administrator controls for reviewing feature usage history and restricting which models are available under organizational policy.
Second — and more unusual — brand familiarity. When hundreds of thousands of public employees with no technical background are asked to adopt a new tool, the fact that it is “based on ChatGPT technology” matters. Employees who have never heard of any other model name still know ChatGPT, and that recognition lowers the initial barrier to adoption. According to Polimill, GPT models are the most frequently selected option in QommonsAI’s general conversation feature for everyday work.
The story also details the development side: Polimill adopted OpenAI’s Codex across its workflow — from requirements definition through implementation and testing — and reports development speed increased 3–5x, with engineers shifting to reviewing AI-generated plans and making high-level decisions. OpenAI’s hands-on support, including sharing global best practices and helping design which steps AI should handle and where humans must review, is described as being as valuable as API performance.
The tacit-knowledge experiment — and its honest result
One finding from Polimill’s internal validation deserves attention for its candor. Less-experienced employees used QommonsAI plus the accumulated administrative knowledge base to draft policy proposals — and those proposals received evaluations close to those from veteran officials. Junior staff, augmented by AI and organized data, could nearly match senior output.
But not equal it. The experienced officials’ proposals still scored highest, and Polimill attributes the gap to tacit knowledge: practical judgment not written in any manual — the procedures required to make a proposal real, the concerns residents will actually raise. Rather than papering over this gap, Polimill’s plan is to capture it: recording how veteran officials instruct AI to research and how they revise outputs, converting undocumented judgment into organizational knowledge. The explicit goal is amplifying skilled employees and passing know-how to the next generation, not replacing them — a framing directly relevant to every government staring down a wave of retiring senior civil servants.
Qommons ONE: the app store for government
The next phase arrives this fall: Qommons ONE, a store where outside companies can offer applications for municipalities. At its center will be a “super agent” that orchestrates multiple specialized AI systems and private-sector apps — a user states a goal, and the system calls whatever AI or application is needed to produce a practical deliverable, from research through presentation.
It is essentially an agent-based platform strategy for the public sector: Polimill supplies the substrate (identity, knowledge base, trust, the relationships with 1,050 municipalities), third parties supply vertical functionality, and the super agent stitches it together. The company projects expanding to 1,200 municipalities and is running joint PoCs in areas from disaster-response coordination between municipalities to humanoid robotics.
Why this matters beyond Japan
Three broader signals make this story more than a customer testimonial.
Public-sector AI adoption at scale is now measurable, not theoretical. Half a million civil servants using a generative AI platform in daily work — with audit controls, a closed-network deployment option, and standardized knowledge infrastructure — is a working reference architecture for every government studying AI adoption.
The “public OS” is an emerging category. Polimill’s framing — a common foundation that supports every municipality equally, so that AI doesn’t widen service gaps between rich and poor local governments — turns AI infrastructure into a question of administrative equity. Expect this framing to spread.
OpenAI’s enterprise strategy is getting more concrete. Publishing deep technical customer stories with named CAIOs, development-speed metrics (3–5x), and hands-on support details shows OpenAI competing for institutional accounts the way classic enterprise vendors always have. In Japan — a market where Anthropic and Google are also courting government and enterprise deals — the story is itself a competitive move.
Japan’s demographic arithmetic makes the direction almost inevitable: a shrinking workforce serving an aging population leaves little room for skepticism about automation in public administration. The interesting questions are now about the shape, not the fact, of that adoption — who owns the knowledge layer, how tacit expertise transfers, and whether a common platform really equalizes small towns and big cities. On all three, QommonsAI is currently the world’s largest live experiment.