Svea

Every Swedish public organisation trying to adopt generative AI faces the same challenges: no in-house expertise, no compute, and no clarity on the legal landscape surrounding the technology. On top of that, each one runs its own procurement process, its own legal review, its own training effort — and carries its own uncertainty about whether AI can ever touch the sensitive personal data and confidential information that make up most public-sector work.

That fragmentation is expensive on its own terms: more than 600 organisations solving the same problem separately instead of once, together. But it also masks a harder constraint. Sweden's public sector faces a workforce shortfall of hundreds of thousands of employees over the next decade, driven by an ageing population, and there is no realistic route to absorbing that gap without giving caseworkers tools that let them do more with their time. Generic consumer AI tools can't clear the legal bar for sensitive data. Building sector-specific AI is beyond what any single organisation can justify or resource alone.

Identifying this problem in 2023, I saw the opportunity for an alternative: a single, trusted AI assistant that Swedish public organisations would develop jointly rather than build in parallel. That became Svea, which I have led from the first idea into a national collaboration now spanning 123 fully onboarded municipalities, regions and government agencies.

What I envisioned was a web-based assistant, built on open models and open software, that any public employee could open in a browser. It supports public servants in writing text, retrieving information, and automating repetitive and arduous tasks. No new IT investment, no new hires, no local infrastructure — just an account. Behind it, a shared layer of verified Swedish knowledge sources, fine-tuned AI-models, and a software architecture and infrastructure that meet the legal requirements for handling of sensitive data.

The technology was the easier half of the problem. The harder work was institutional: selling the idea to leaders and earning the trust of the people who would actually use it, creating the necessary training, onboarding and change-management programmes, and establishing the legal and security frameworks that would let Svea handle sensitive personal data.

I raised roughly €9 million in combined funding and in-kind contributions to build the programme, and now lead a multidisciplinary team of around 20 across AI engineering, software development, product, law, change management and education. A distributed annotation programme of more than 900 public-sector contributors — caseworkers, administrators, communicators — has produced over 500,000 annotated data points, used to train the small AI models that support the system, from document retrieval to interpreting what a user is asking for.

Svea now serves 123 fully onboarded organisations, with more than 30,000 registered accounts and approximately 5,000 weekly active users. Active users report saving around two hours a week, and 90% report an improvement in the quality of their work — together representing an estimated €4.5 million in annual efficiency gains across participating organisations. More than 3,000 public-sector employees have been trained in applied, responsible AI use through Svea's education programme.

Beyond the numbers, Svea has become a reference point for what national, collaborative AI adoption can look like in a regulated environment. Presented to Swedish government ministers and under consideration by the agencies responsible for national digitalisation, it is on track to become an established national service in 2027.

The underlying lesson generalises beyond Sweden. The hardest part of adopting AI in a high-stakes, regulated setting was never the application — it was bringing stakeholders with us, establishing legal trust, building shared infrastructure and developing the organisational capability that adoption depends on. That is the problem Svea solved, and it is the part of this work that travels furthest.