Responsible AI (RAI) governance is increasingly understood not as a static checklist of principles, but as a dynamic process embedded in institutional, organisational, and sociotechnical contexts. While several ethical frameworks exist, translating high-level principles into situated organisational practices remains challenging. Empirical studies examining how public sector organisations operationalise RAI remain fragmented, limiting cumulative insights. To address this gap, we conduct a realist synthesis review of 21 empirical studies. Our analysis shows that similar interventions in different contexts activate distinct mechanisms and produce divergent outcomes with varying degrees of alignment to RAI principles. From these variations, we identify three cross-cutting dynamics explaining outcomes: organisational embeddedness, power- expertise tensions, and trust-transparency relationships. Together, we term it the situated dynamics of RAI governance. This approach moves beyond asking whether interventions “work” to explain why similar interventions succeed in some contexts and fail in others.

Public sector organisations are increasingly expected to operationalise Responsible AI (RAI) principles such as transparency and accountability. Yet translating these principles into everyday practice remains challenging and underexplored, particularly in public sector settings. In this study, I explore how accountability practices around transparency requirements are navigated inside a Dutch regulatory agency developing AI systems for risk-based inspections, drawing on my 1.5 years of organisational ethnography. The preliminary findings highlight how public sector practitioners engage in everyday governance work – creating accountability processes through improvised practices when formal structures are absent. This work contributes to a practice-based understanding of RAI and challenges assumptions about organisational maturity as the solution to operationalisation challenges.