Japan doesn't do things halfway. When Polimill decided to build AI infrastructure for municipal government, they didn't target a few pilot cities. They built QommonsAI, released it in October 2024, and now serve approximately 1,050 municipalities and 550,000 public employees across the country.
This is what shipping AI infrastructure at national scale actually looks like. And the architectural choices are fascinating.
The problem: fragmented data, distributed knowledge
Polimill started with Surfvote, a civic engagement platform. Working with local governments revealed the core bottleneck: public-sector teams were drowning in operational work and had no bandwidth to incorporate citizen input into policy.
The insight was structural. You can't improve civic participation if government employees are spending their days manually searching through decades of assembly minutes to ensure policy consistency.
Every municipality in Japan maintains its own workflows, document formats, and scattered historical records. Assembly response preparation could require hand-reviewing years of transcripts. Even the best LLM is useless if the underlying data is a mess.
Polimill's move: collect and standardize assembly minutes from across Japan, use AI to add metadata, and build a cross-municipality search foundation. Then expand that structure beyond assemblies into welfare, legal domains, and other administrative areas.
The result is a unified knowledge base that turns distributed administrative information into something usable in daily work. That's the foundation for what Polimill calls a "public OS" for municipal government.
Architecture: GPT models, government-grade controls
OpenAI's GPT models are the core inference layer. The choice wasn't just about capability—it was about adoption friction.
Masahiro Wakabayashi, Polimill's CAIO, points to GPT's "broad capability and familiarity" as the deciding factor. Public employees don't know model names or technical jargon. But they know ChatGPT. That brand recognition dramatically lowers the initial barrier when rolling out a new tool across hundreds of municipalities.
According to Polimill, GPT models are the most frequently selected option in QommonsAI's general conversation feature. The ability to handle a wide range of work requests—file reading, dialogue support, research—in a single model matters when you're trying to get adoption across diverse use cases.
But public sector means strict operational controls. QommonsAI includes:
- Feature usage history review for administrators
- Model availability controls based on organizational policy
- Audit-ready information management
This is the balance you need when shipping AI into government: familiar interface, enterprise-grade governance.
Development velocity: Codex and 3-5x speedup
Polimill adopted Codex across their entire development workflow—requirements definition, consistency checking against existing GitHub code, implementation, testing. Engineers focus on reviewing AI-generated plans and high-level decisions while AI handles more of the autonomous implementation work.
The result: development speed increased 3-5x. Wakabayashi notes that the validation cycle—build prototypes, show municipal teams, gather feedback—has "accelerated dramatically."
But here's the interesting part: OpenAI provided hands-on support beyond API access. That included:
- Sharing best practices from global examples
- Tailored information for Polimill's specific use cases
- Help designing which development steps AI should handle versus where humans review and take responsibility
Wakabayashi sees this ongoing collaboration as "valuable as the API's performance itself." As models and capabilities evolve, AI-assisted development practices keep changing. Individual companies can't track advanced global examples and incorporate them into engineering orgs alone.
This is the unsexy reality of enterprise AI adoption: the scaffolding around the model often matters as much as the model.
Results: adoption and quality benchmarks
The numbers:
- ~1,050 municipalities using QommonsAI
- ~550,000 public employees
- Unified infrastructure for assembly minutes and administrative information searchable across organizations
- 3-5x faster development cycles with Codex and OpenAI support
But the quality metric is more interesting. In Polimill's validation, less-experienced employees used AI and accumulated administrative information to draft policy proposals. Those proposals received evaluations close to proposals from veteran officials.
Close, but not equal. The experienced officials' proposals still scored highest.
Polimill's analysis: the difference is tacit knowledge. Practical judgment that isn't written in manuals—procedures required to make a policy proposal real, resident concerns that might arise, political context.
I think the parts that are not captured in knowledge or data contributed to the quality of the outcome. Put another way, when people with that kind of experience use AI, they can create even better work.
This is the right framing. The goal isn't replacing skilled employees with AI. It's amplifying their abilities and capturing their know-how for the next generation.
Polimill plans to record how veteran officials instruct AI and revise outputs, turning undocumented judgment into organizational knowledge. QommonsAI becomes a knowledge transfer platform, not just a productivity tool.
The roadmap: super agent and app ecosystem
Fall 2026 target: full rollout of Qommons ONE, an app store where outside companies can provide applications for municipalities.
At the center: a "super agent" that brings together multiple specialized AI systems and private-sector apps. Users state a goal, and the system calls the necessary AI or apps to produce deliverables—research, presentations, whatever's needed.
Polimill sees GPT models' broad input/output capabilities and OpenAI's agent development know-how as critical foundations. OpenAI is supporting development through technical validation and individual samples.
The ambition extends beyond administrative efficiency. The time created by AI should let public servants listen more closely to citizens' voices and support them better. That's the "public OS" vision—government as a platform where AI creates capacity for human connection.
Why this matters
This is a case study in what national-scale AI infrastructure deployment actually requires:
- Data infrastructure first. You can't skip the unglamorous work of standardizing and organizing distributed information.
- Adoption psychology matters. Familiar branding (ChatGPT) lowers barriers more than you'd think.
- Governance and usability aren't opposed. You need both audit controls and broad capability.
- Development velocity compounds. 3-5x speedup means faster validation cycles, which means better product-market fit.
- Capture tacit knowledge, don't ignore it. The gap between junior + AI and senior + AI is the insight, not the problem.
Polimill's work is just beginning. But watching how they're building public infrastructure—where governments, companies, and citizens connect through AI—is instructive for anyone shipping AI systems at scale.
The public sector is often treated as a slow follower in tech adoption. QommonsAI suggests that framing might be backwards. When you get the fundamentals right—data infrastructure, governance, and adoption design—government can move faster than you expect.