New Framework for Agentic AI Autonomy and Agency in Regulated Contexts

A recent paper on arXiv presents a two-dimensional framework for implementing agentic AI within regulated settings, focusing on the interconnectedness of agency (system capabilities) and autonomy (degree of human involvement). This structure categorizes both dimensions into five operational tiers, ranging from human-directed operation (L1) to completely autonomous oversight (L5) for autonomy, alongside reasoning based on provided context for agency. The goal of this methodology is to establish a solid foundation for evaluating oversight, the implications of actions, and error rectification—areas where existing approaches fall short.