Governing Enterprise AI Agents: A Framework for Permissions, Human Oversight, and Recovery

Enterprise artificial intelligence agents can select tools and change business records. Organizations must define who may authorize an action, when human review is required, and how to recover from failures or uncertain outcomes. This paper develops a conceptual framework through a targeted synthesis of AI risk guidance, agent security research, human oversight studies, and reliable execution practices. A fictional customer service workflow illustrates the proposed controls; no deployment or experimental evaluation is reported. The framework integrates an action contract, independent permission enforcement, approvals bound to specific operations, durable execution records, and authoritative outcome verification. It distinguishes a completed action from its unfinished workflow and an unknown outcome from confirmed failure. A refund example demonstrates why recovery must preserve completed effects while resolving remaining work. Enterprise autonomy should expand only when authority, review quality, and recovery can be verified. The framework provides a basis for implementation and evaluation. Its effectiveness depends on correct policies, complete enforcement, trustworthy interfaces, and accountable operators.

Artificial Intelligence