Delegation Architecture: A Framework for Analysing Human Authority in Autonomous Decision Systems

This preprint examines the governance of autonomous decision systems by arguing that the widely used concept of human-in-the-loop is insufficient for analysing whether humans retain meaningful authority over consequential decisions. While existing governance frameworks typically require the presence of a human participant, they often fail to distinguish between formal participation and actual decision-making mandate. The paper introduces two original analytical concepts: the Delegation Gradient, which describes the continuous distribution of effective decision authority between humans and autonomous systems, and the Delegation Architecture, a framework for analysing how authority is assigned, exercised, constrained, and revoked within autonomous decision processes. Building on these concepts, the paper develops a taxonomy of mandate failure and demonstrates how authority can be transferred from humans to systems even when formal human oversight remains in place. Drawing on research in human–automation interaction, meaningful human control, AI governance, and documented operational examples, the framework provides a structured approach for evaluating decision authority across military, critical infrastructure, and other high-consequence applications. The paper argues that future AI governance should regulate not only human presence but also the architecture of delegated authority, offering a foundation for more robust policy, regulatory, and system design frameworks.