The agentic turn in transfer pricing control: From experimentation to traceable governance

Abstract

The integration of Artificial Intelligence in tax administration has reached a qualitative turning point with the emergence of "agentic" systems capable of reasoning, planning, and executing complex control processes autonomously. This research examines the application of agentic AI to transfer pricing, identified as the "last mile" of tax administrations' digital transformation. Through a mixed-method approach combining conceptual analysis of multi-agent architectures, experimental study on simulated data (20 MNE cases across 3 sectors), and comparative analysis of international deployment models (11 countries), we demonstrate that agentic AI can reduce processing time by 98% while increasing the depth of functional analyses. However, our results reveal a structural tension between algorithmic efficiency and legal transparency requirements, with a 15% error rate in complex functional characterizations, 22% irrelevance rate in comparable selection, and 15% hallucination rate in legal citations. We propose a "Tracer-Wire" governance framework based on explainability by design and the maintenance of a "Human-in-the-Loop" for critical decisions. The article concludes with implications for tax treaty interpretation and proposes amendments to the OECD Model Tax Convention Commentary.