Organizations often treat the person listed as an AI agent’s owner—typically its creator—as the person accountable for it. This approach is inadequate because an owner field identifies whom to call when an agent fails, not who has authority over its purpose, risk, permissions, or termination. Responsibility and accountability must be distinguished: responsibility concerns performing work and maintaining operations, while accountability concerns answering for outcomes and possessing the authority to make decisions.
Accountability per Decision assigns exactly one accountable party to every significant decision in an agent’s lifecycle, including its purpose, authorization, operating bounds, residual risk, and shutdown. Responsibility may be shared among builders, operators, platforms, and automation, but accountability cannot be delegated to a tool, another agent, or a diffuse group. Control planes can document an agent’s activity, access, and execution, yet they cannot establish who authorized those actions; that authority is an organizational matter.
Accountability should generally sit outside the technical team operating the agent. Business owners answer for purpose and residual risk, risk functions independently validate risk, and a separate intervention authority decides when an agent must be stopped. The same principles apply to workload identities, whose accountability follows business purpose rather than technology ownership. Identical agents can share an accountable owner only when their purpose, authority, and risk are identical.
Organizations should extend this model to existing agents through access reviews, prioritizing high-impact systems. Before any agent receives credentials, it must have a documented purpose, a named accountable business owner, and approved operational bounds. When an agent’s approved purpose ends, its credentials, delegated authority, and child agents must also be revoked.
See All Locations
See All Locations