This paper analyses how artificial intelligence is governed across a diverse set of countries, by mapping policy intervention using a portfolio approach. We collect and classify AI-related policies along two key dimensions: targets, denoting the specific objectives pursued by each policy, and instruments, referring to the regulatory or programmatic tools employed. Building on this dataset, we present a comparative description of how policy portfolios vary across countries along these two dimensions.
Our data collection and classification method relies on a pipeline grounded in text analysis and generative AI. A central methodological contribution of the paper is the systematic comparison of three increasingly complex architectures for policy classification: a standalone Large Language Model (LLM) approach, a Retrieval-Augmented Generation (RAG) pipeline, and an agentic workflow. We evaluate each architecture on classification accuracy, scalability, and robustness to heterogeneity in policy formats and language, offering practical guidance for researchers seeking to deploy generative AI in large-scale policy analysis.
The resulting method is designed to be both scalable to other policy sectors and transferable across constituencies, whether countries, regions, or local entities. We detail the classification scheme and pipeline steps that enable systematic cross-country comparison. While our primary focus is on data collection and classification, we highlight patterns of convergence and divergence in AI regulation that provide an empirical foundation for understanding the politics of AI governance and lay the groundwork for subsequent analyses of how policy portfolios shape trajectories of technological development and innovation.