Before agentic AI scales in government: the democratic authorization gap

Abstract

Some governments have begun to procure and test artificial intelligence systems that can pursue goals through connected actions. However, publicly documented evidence of mature agentic AI in public administration remains limited, with most initiatives remaining at procurement, pilot, or early deployment stages. This Perspective introduces the democratic authorization gap, defined as a break or attenuation in the demonstrable chain connecting legally and democratically grounded public authority to actions selected, sequenced, or executed by an AI agent. Drawing on democratic delegation, accountability, administrative law, and recent agentic-AI scholarship, the article distinguishes this prospective, authority-based problem from responsibility gaps and technical authorization. It identifies four mechanisms through which the gap may develop: mandate translation, recursive delegation, action diffusion, and contestability lag. Five governance conditions are proposed for the pilot stage: bounded authorization, permission inheritance, action-level traceability, named institutional responsibility, and operational interruption with reversible redress. The argument is anticipatory rather than empirical. Following the Collingridge dilemma, limited evidence before large-scale deployment provides a reason to establish governance conditions while institutional choices remain open.