Autonomy levels

Feng, McDonald and Zhang, in a paper revised on 28 July 2025, define autonomy as the extent to which an AI agent is designed to operate without user involvement². The user can be a person or another agent². They treat the level as a design choice separate from capability: a capable agent can still be required to consult the user before each action².

The five levels are user roles, with no level 0². As operator, the user stays in charge and the agent helps only on demand². As collaborator, the user and agent plan, delegate and execute together, and the user can take control at any point². As consultant, the agent plans and executes most of the work, and the user is asked for expertise and preferences². As approver, the agent involves the user only at a failure or high-risk state, or when approval is required². As observer, the agent plans and executes over long time horizons and makes the decisions². The user can read activity logs but cannot provide input or change the trajectory, apart from an emergency off-switch that stops all agent activity².

A document published on 18 March 2026 by the Cloud Security Alliance AI Safety Initiative uses another scale¹. Its PDF is marked unofficial AI-assisted research and has not had official CSA review³. It defines AI autonomy as the degree to which a system can decide and act without human approval, oversight or intervention¹. Capability is what an agent can do; autonomy is what it is allowed to do without human intervention¹. Level 0 only supplies information, such as a chatbot giving advice¹. Level 1, such as a code assistant waiting for a run confirmation, needs approval of each action¹. Level 2 executes a plan a person has approved¹. Level 3 acts inside boundaries and escalates outside them, and is the first level without per-action or per-plan approval¹. Level 4 uses monitoring and exception handling rather than decision approval¹. Level 5 includes goal-setting and self-modification with minimal human involvement, and is not recommended for current enterprise deployments¹. Scripted robotic process automation, with no AI decision, is outside this scale¹.