Controlling Autonomous AI

Controlling Autonomous AI examines how organisations can retain meaningful control as AI agents are given authority to take consequential actions across enterprise systems. The paper introduces three linked ideas: Control Illusion, where formal controls remain present but no longer constrain the outcome; Control Margin, which tests whether intervention can still take effect before an outcome becomes unavoidable; and global admissibility, which considers whether individually valid actions combine into an unacceptable system state. It proposes IABEI — Identity, Authority, Boundaries, Evidence and Intervention — as an assurance structure for testing whether existing controls form an effective end-to-end control chain. Working paper, Version 1.0.