Pre-Action Authorization in High-Risk AI Systems Minimum Structural Requirements for Legitimate Decision Execution

Many governance failures in high-risk AI systems are commonly attributed to model inaccuracy, insufficient monitoring, inadequate testing, or a lack of transparency. While these factors matter, a substantial class of real-world failures originates at a deeper structural layer: the absence of legitimate authorization prior to action execution. In such systems, decisions produced by AI models (or hybrid human-AI workflows) transition into real-world actions without a distinct, enforceable legitimacy gate at the action level. This creates an architecture in which oversight becomes primarily observational and post-hoc, rather than preventive.This scientific opinion argues that effective governance in high-risk contexts requires an explicit separation between decision generation and action execution, and the establishment of pre-action authorization as a distinct governance layer. It provides implementation-agnostic structural requirements that can be independently examined and audited: clear responsibility allocation prior to execution, enforceable veto capability, non-bypassable authorization gating, audit-stable authorization traces, and the ability to prove both the presence and absence of authorization for specific actions.Rather than proposing a new regulatory regime or a particular technical implementation, this document defines minimum structural requirements for legitimate decision execution. It is intended as a reference for auditors, risk owners, governance teams, system architects, and regulators seeking criteria to determine whether an AI system is structurally capable of executing actions legitimately in high-risk environments.