AI agents for public services
Sovereign, open-source, built for governments and nonprofits.
AI agents for public services are software agents that do real work for a government or nonprofit, grounded in the organization's own validated documents, protocols and rules, not in whatever a model picked up from the open internet. The Bayes Platform is the open-source, sovereign platform for building and running them: your data stays in-house, every answer cites its source, and the code is yours, with no vendor lock-in.
What is an AI agent for public services?
It is not a chatbot with search bolted on. It is an agent that uses tools to carry out a task: it searches the organization's knowledge base, fills a form field by field, cites the source of every answer, or delegates to a specialized sub-agent. The same orchestration runs all of these capabilities.
What makes it fit for public services is grounding. The agent answers from your handbooks, protocols and legal references, follows your rules, and stays inside guardrails you define. When it does not know, it says so, rather than improvising exactly where it matters most.
What makes the Bayes Platform sovereign and open-source?
Sovereign means you keep control. The platform is self-hostable, and it runs on open models like Gemma 4 and Mistral on your own infrastructure, so citizen data never has to leave your walls. Some of our government partners cannot send data to a cloud API at all, and this is what makes the platform viable for them.
Open-source means the platform is yours. Everything is MIT-licensed: deploy it, modify it, take it over whenever you want. No lock-in, and every use case joins an open library shared with similar organizations worldwide, a digital commons rather than a captive product.
How does an agent stay trustworthy?
Every answer is grounded in your validated documents and cites its source, so a caseworker or a clinician can check it. The goal is zero hallucination, not a confident guess.
Nothing reaches the public before a human validates it, and once in production the agents are monitored over time. The platform is built to sector standards, with EU AI Act and GDPR compliance in mind from the start.
What can these agents actually do?
They are already deployed with real public institutions, and the needs come from the ground up rather than from a feature roadmap. The examples below are live, not mockups.
A conversational agent guiding citizens toward care and social-service careers: profile analysis, mobility, job matching, training guidance.
Clinical insights, step-by-step protocol support, research screening and patient education, deployed and monitored in public hospitals.
OCR agents on sovereign infrastructure turn millions of pages of paper archives into a structured base, ready for applications and AI agents alike.
How do you get started?
There are two ways in. The first is targeted use cases: we take one or two of your real cases from idea to production, fast, then your teams experiment for real. The second is the Impulse program, a cohort that moves from experimentation to production together across a sector.
Either way, the platform manages the full lifecycle, from rapid prototyping to experimentation, then production and monitoring over time.
Why a nonprofit, not a vendor?
Bayes Impact is a nonprofit. Our success is measured by the impact we spread, not by your dependence on us. We optimize for reach and outcomes, not billable hours.
That is why the platform is a partnership, not a product you rent: the code is yours, the data is yours, and what we build with you strengthens a shared commons for public services everywhere.
Generic AI assistant vs a Bayes Platform agent
| Generic AI assistant | Bayes Platform agent | |
|---|---|---|
| Grounding | Open internet | Your validated documents |
| Data location | Vendor cloud | Your infrastructure, self-hostable |
| Sourcing | Unverified | Every answer cites its source |
| Lock-in | Proprietary | MIT, the code is yours |
| Compliance | Generic | EU AI Act & GDPR, to sector standards |
| Model | Fixed vendor model | Your choice, including open self-hosted |
| Economics | Per-seat SaaS | Nonprofit, impact-focused |
Frequently asked questions
- Is the Bayes Platform really open-source and free?
- Yes. The platform is MIT-licensed and built by a nonprofit. You deploy it on a platform you own, modify it, and take it over whenever you want.
- Can agents run without sending citizen data to a cloud API?
- Yes. The platform runs on open models such as Gemma 4 and Mistral, self-hosted on your own infrastructure, so citizen data never has to leave your walls.
- Does it work in French and meet EU rules?
- Yes. The platform is multilingual and built with EU AI Act and GDPR compliance in mind, to public-sector standards.
- How do you prevent hallucinations?
- Agents answer only from your validated documents and cite their source. A human validates before anything reaches the public, and agents are monitored in production.
- Who already uses it?
- Governments and nonprofits across Europe and beyond, including France Travail for employment, AP-HP for healthcare, and justice archives in Togo. Bayes Impact has served more than 10 million people since 2014.
- How is Bayes different from a software vendor?
- Bayes Impact is a nonprofit and works as a partner, not a vendor. There is no lock-in, the code and data are yours, and every use case joins a shared digital commons.
Build public-service agents you can trust.
Content you control, answers your users can trust. Let's talk about your use cases.