As artificial intelligence systems transition from advisory tools to agents capable of initiating actions with real-world consequences, existing approaches to AI governance have proven insufficient. Most contemporary frameworks emphasize model behavior, ethical principles, access controls, or post-hoc oversight. However, operational failures in AI systems consistently arise not from reasoning quality, but from unauthorized execution. This paper argues that governable AI requires architectural constraints that enforce authority before action, rather than relying on probabilistic controls, policy artifacts, or retrospective audits. It introduces three minimal and complete architectural invariants that define a governability threshold for AI systems operating in regulated, safety-critical, and liability-bearing environments. The paper examines why current standards frameworks such as NIST, ISO, and CMMC implicitly assume these constraints without structurally enforcing them, and it shows how agentic AI amplifies authority failures when these invariants are absent. By reframing AI governance as an architectural discipline centered on execution control, structural non-existence of unauthorized states, and temporal binding of authority, this work provides a foundation for predictable, traceable, and defensible AI deployment at scale.