What practical hurdles arise when public agencies collaborate to build their own sovereign generative AI infrastructure?

Host: Welcome to A.I.S. Insights — powered by Living Knowledge. I'm Anna Ivy Summers. Today, we are examining a study titled "SVEA-in-the-Making: A Sociomaterial Genealogy of Public-Sector LLM Development in Sweden," authored by Shengnan Han, Gideon Mekonnen Jonathan, and Hongxiu Li. The study analyzes the development of SVEA, an initiative to build a sovereign public-sector AI assistant in Sweden. Joining me to discuss it is our analyst, Alex Ian Sutherland. Alex, welcome.

Expert: Thank you, Anna. It is a pleasure to be here.

Host: Alex, what was the primary problem this study set out to address?

Expert: Many public administrations are working to build sovereign language models to reduce dependence on commercial providers and align with strict privacy regulations. However, research often treats AI systems as finished, stable products. That perspective overlooks the coordination challenges, institutional tensions, and operational hurdles that arise when independent public organizations attempt to co-develop shared AI infrastructure.

Host: How did the researchers examine this development process?

Expert: The authors used a sociomaterial genealogical approach across two distinct phases of the initiative. They collected qualitative data through interviews with technical leads, project managers, and legal specialists, as well as a focus group and internal technical and governance documents. This allowed them to trace how technical architecture and institutional environments influenced each other over time.

Host: Let us talk about what they found. What were the core outcomes of the study?

Expert: First, the study identified four coexisting versions of the SVEA initiative: a technical model, an object of governance, a political project for digital sovereignty, and a workplace tool for public servants. These versions often competed with one another. Second, creating training data required extensive manual annotation and curation by municipal employees. Because this was added on top of their standard daily workloads, it resulted in notable staff exhaustion.

Host: There were also structural and legal obstacles identified, correct?

Expert: Yes. The technical team successfully built the capability to handle sensitive data, but it could not be deployed. The multi-agency innovation consortium lacked legal personhood, meaning it could not legally serve as a data controller under Swedish law. Additionally, political directives prioritizing data sovereignty triggered a localization cascade. These mandates restricted hardware hosting choices and prevented the integration of external search services. Because there was no single central authority over the participating municipalities, the project had to be coordinated through evolving legal interpretations rather than standard managerial directives.

Host: What does this mean in practice for business leaders and public sector managers?

Expert: There are three main takeaways. First, legal architecture must precede technical development. Building advanced data capabilities provides little practical value if the organizational entity lacks the legal authority to process that data. Second, data curation cannot be treated as informal side work for operational staff. Without dedicated roles and resources for data preparation, initiatives risk project delays and employee burnout. Finally, digital sovereignty involves operational trade-offs. Mandating strict local hosting directly impacts technical options, infrastructure costs, and functionality, all of which need to be accounted for early in project planning.

Host: In short, building sovereign AI infrastructure requires solving legal, organizational, and labor questions alongside technical engineering. Alex, thank you for breaking down the study.

Expert: You are welcome, Anna.

Host: Thank you to our listeners for tuning in to A.I.S. Insights — powered by Living Knowledge. Join us again next time.