Delegation of Authority to AI: Three Conditions for Defensible Automated Decisions

Abstract

Delegation of Authority to AI states the conditions under which a licensed professional or regulated institution may route a bounded decision function through an AI system without treating the system as a licensed or independently accountable professional. The paper identifies three independent conditions that must hold: Legal automatability: the governing legal and regulatory regime permits the function to be performed through a system. Domain validity and deployment fitness: the system has been empirically validated for the defined function, population, operating conditions, and failure modes. Runtime authorization integrity: every effect-bearing action passes through a pre-execution authorization boundary designed to be non-bypassable. The first condition belongs to law and the second to empirical validation. Only the third is a governance-architecture problem. The paper explains how the first two conditions define the permitted decision space and how the third enforces that space at runtime. It further distinguishes ABSTAIN plus authorized human override from guardrails plus human-in-the-loop, and maps the evidentiary requirements of runtime authorization integrity to the Five Tests Standard: Stop, Ownership, Replay, Escalation, and Provenance. The paper is explicit about the limits of the framework. Runtime authorization integrity does not replace professional accountability, establish decision quality, confer legal authority on the system, or cure a failure of legal automatability or domain validation. Its contribution is narrower: it makes retained authority operationally enforceable and its exercise independently reconstructable through a tamper-evident authorization artifact. This working paper is the affirmative companion to a research series on runtime authorization, including Authority versus Authorization (DOI: 10.5281/zenodo.21341907) and The Override Asymmetry (DOI: 10.5281/zenodo.19772248).