This paper aims to contribute to the study of legislative politics in the European Union (EU) in three main ways:

First, it introduces a computational approach to analyzing policy design in EU legislative decision-making. By fine-tuning a domain-specific BERT model with a named entity recognition head to map actor involvement in EU legal texts, this approach bridges the gap between manual policy analysis and coarse representations of legislative text, offering a scalable yet nuanced framework for automated policy annotation.

Second, applying the trained model to a corpus of more than 3,000 EU legislative texts nested in more than 1,000 legislative procedures, the paper investigates the evolution of policy design throughout the legislative process. Comparing initial European Commission proposals with final texts adopted by the European Parliament and Council reveals significant variations in actor involvement and other policy features, highlighting the dynamic, multi-institutional nature of EU policymaking.

Third, the paper investigates possible covariates of changes in policy design over the legislative process, focusing on the institutional and procedural factors frequently associated with shifts in policy outcomes.

By leveraging and adapting pre-existing models and an annotated dataset, this study provides novel tools, data, and empirical insights for addressing both existing and emerging questions in the field of EU decision-making and legislative politics, that can help us understand power and legislative outcomes in the EU. Beyond the EU, the presented approach offers a transferable framework for large-scale, automated analysis of policy text, providing new avenues for research on legislative politics and policy design.