Between Adoption and Governance : How AI Systems Are Translated in a Plural Public Organization

Sammanfattning

This thesis investigates how Region Uppsala, a Swedish regional public sector organization, translates governance templates for Artificial Intelligence (AI) in an environment characterized by institutional pluralism. Through an exploratory qualitative embedded case study design, twelve in-depth interviews were conducted across three analytical units within the organization: the administrative AI (Microsoft 365 Copilot); the clinical AI at Akademiska Sjukhuset, and the pre-adoption AI in welfare and an AI specialized unit. Translation theory and institutional logics provide the analytical framework. The study shows how the translation outcomes vary systematically between the units. Three contextual conditions were abductively identified as an explanation for this variation: risk proximity, logic dominance and governance ownership. Where proximity to risk was low and governance unassigned, the result was a ceremonial stabilization of the template. In the clinical unit, where the risk proximity was highest, the dominant professional logic led to a negotiated stabilization of the template through a process of layering. Finally, when the governance of AI use was designed as a prospective, based on a bureaucratic-legal logic of AI governance, an anticipatory compliance was adopted to the template before any specific system had been deployed. The responsibility gap observed across all three sites is therefore not simply a failure of implementation, it is a gap that has been produced structurally.