Summary
- About 1,050 local governments across Japan and roughly 550,000 civil servants are using QommonsAI, built by Japanese startup Polimill
- Close collaboration with Codex and OpenAI has sped up QommonsAI's development by three to five times compared to before
- Polimill plans to fully launch Qommons ONE, a marketplace for outside apps, along with a super-agent, in fall 2026
Public-sector AI now used by 1,050 Japanese local governments
According to a customer case study published by OpenAI, the platform gives local governments AI features tailored to specific domains — council responses, citizen services, social welfare, and legal research among them.
A company that started as a civic-engagement platform
Polimill originally built Surfvote, a platform where citizens could weigh in on political and social issues. While working with local governments, the company noticed a structural problem: civil servants were so buried in day-to-day tasks that they had little time to actually fold citizen input into policy. Polimill concluded that expanding civic participation first required making administrative work itself more efficient — and that's what led to the October 2024 launch of QommonsAI.
Unifying data that differed from one local government to the next
The biggest obstacle to bringing generative AI into government work was fragmented data. Every local government had its own way of doing things and its own document formats, with historical records scattered across different systems. Preparing a council response, for instance, meant manually digging through years of meeting minutes just to check consistency with regional policy. No matter how capable the AI model, it couldn't produce useful answers if the underlying data wasn't organized.
Polimill collected and standardized council meeting records from local governments nationwide, then used AI to tag them with metadata, building a foundation that could be searched across jurisdictions and time periods. It extended that same structure to other administrative domains like welfare and legal affairs, making it possible to search scattered administrative information through a single common interface.
Why GPT models were the choice
GPT models sit at the core of QommonsAI. Because the public sector places heavy weight on information management and audit readiness, QommonsAI also includes operational controls that let administrators check usage history by feature and restrict which models can be used under organizational policy.
Masahiro Wakabayashi, Polimill's Chief AI Officer, points to both broad capability and familiarity as reasons for choosing GPT models. Even when civil servants find a new tool unfamiliar, the fact that they already know the name ChatGPT lowers the initial barrier to adoption, he explains. According to Polimill, GPT models are the most frequently selected option within QommonsAI's general conversation feature, and their ability to handle everything from reading files to supporting everyday work conversations through a single model has helped the tool take hold on the ground.
Development speed up three to five times after adopting Codex
OpenAI's technology has changed not just how QommonsAI is used, but how it's built. Polimill applies Codex across the entire development process — from defining requirements, to checking consistency with existing GitHub code, to implementation and testing. Engineers focus on reviewing AI-generated plans and making higher-level decisions, while a substantial share of the actual implementation work is handled autonomously by AI.
Wakabayashi says the feedback loop of building a prototype, showing it to local government staff, and incorporating their input has noticeably sped up. On top of that, OpenAI has provided close, hands-on support — sharing best practices from other countries and helping refine the development process itself — which together pushed development speed up three to five times compared to before.
Results at a glance
| Item | Figure |
|---|---|
| QommonsAI launch | October 2024 |
| Local governments using it | About 1,050 (nationwide in Japan) |
| Civil servants using it | About 550,000 |
| Development speed gain | 3–5x after adopting Codex |
| Next step | Full launch of Qommons ONE planned for fall 2026 |
Trying to capture the instincts of veteran civil servants
QommonsAI's value doesn't stop at faster searches. In tests Polimill ran, policy proposals written by less experienced staff using AI and accumulated administrative information came close to the quality of proposals from seasoned civil servants. The highest ratings, though, still went to the veterans — and Polimill traces that gap to tacit knowledge that isn't written down anywhere in a manual: a feel for the procedures that actually get policies through, and for the points that might worry residents.
"The parts that can't be captured as knowledge or data still affected the quality of the output," Wakabayashi says. Polimill plans to record how experienced civil servants instruct the AI and how they revise its output, turning previously undocumented judgment calls into organizational knowledge.
The next step toward a super-agent
Polimill plans to fully launch Qommons ONE, a marketplace where outside companies offer applications for local governments, in fall 2026. At its center will be a super-agent that ties together multiple specialized AIs and third-party apps — the idea being that a user states a goal, and the system calls on the AI or app needed to produce practical outputs, from research to presentation materials. Polimill sees the broad input-output capability of GPT models and OpenAI's accumulated agent-development know-how as important building blocks for making that vision work.
OpenAI previously said it had expanded ChatGPT for Teachers to an additional 55 school districts across 20 U.S. states, bringing the number of newly covered educators and staff past 100,000. As with that case — OpenAI Expands ChatGPT for Teachers to 55 U.S. School Districts — QommonsAI's rollout fits a broader pattern of large-scale, organization-wide AI adoption spreading across the public and education sectors.
Editor's take
What makes the Polimill case interesting is that the numbers OpenAI is highlighting this time aren't about model performance — they're about scale of adoption and speed of development. Rather than the model's edge over competitors, it's the fact that hundreds of thousands of people already use it every day that becomes the pitch for the next customer.
Set against last year, OpenAI's enterprise case studies used to center on revenue and performance metrics. Lately the emphasis has shifted toward stories of Codex reshaping the development organization itself. A story about a startup building a service at government scale three to five times faster without dramatically growing its engineering headcount carries particular weight in the B2G market, where small teams routinely have to deal with much larger organizations.
The part of this case most worth studying domestically is data standardization. The task that consumed most of Polimill's time wasn't tuning AI models — it was collecting and reformatting meeting records from local governments across the country. For any public institution or company here considering a similar project, the lesson is that budget and time need to go toward organizing scattered documents before AI adoption, if the effort is going to produce real results.
Over the coming months, the next thing worth watching will be how far Qommons ONE actually reaches when it launches in fall 2026, and how many outside companies' apps end up plugged into it.
Comments