Organizing AI oversight in digital government: A multidimensional analysis of enforcement architectures for the European Union Artificial Intelligence Act

How do governments translate a harmonized artificial-intelligence regulation into national enforcement arrangements? This article examines the organizational infrastructure being constructed to implement the European Union Artificial Intelligence Act. We conduct a comparative documentary study of all 27 member states using a 65-record source register frozen on 15 August 2026, a five-tier evidence hierarchy, a multidimensional codebook, and independent full-sample recoding by a second author. The frozen corpus identifies nine formally adopted arrangements with at least one public activation marker, five adopted arrangements without such a marker, nine proposed or officially announced designs, and four unclassified states. A graded activation measure shows that the nine M4 cases displayed interface or procedural/organizational activation; none met the strongest A3 criterion of documented supervisory action at the data freeze. A pre-adjudication sensitivity analysis changes marginal counts but not the principal pattern: hub-and-spoke remains the modal topology, inherited regulators remain the largest institutional locus, no centralized case emerges, and the combined number of adopted or publicly activated arrangements remains 14 under both coders and the consensus dataset. Documentary evidence establishes formal allocations and public activation markers, not enforcement performance. We therefore theorize three forms of formal-operational decoupling and a six-stage regulatory learning loop. The article contributes a replicable framework for studying digital-government oversight infrastructures and testable propositions concerning political incentives, coordination authority, rights oversight, and adaptive learning under multi-level regulation.