Levels of autonomy

What it means

Levels of autonomy are a way to say how much an AI system does by itself before a person has to step in. Picture a new assistant. At first you check every email before it goes out. Later you approve only the important ones. One day the assistant may send them all, and you only watch. Each step is a level. The European Union's AI law builds the idea into its definition of an AI system, calling it "a machine-based system that is designed to operate with varying levels of autonomy."¹

Where the word comes from

The idea that machines can work at different degrees of automation is older than today's AI. A 1978 report by T. B. Sheridan and W. L. Verplank on remote-controlled undersea vehicles put the "Emphasis ... on human operator control of such teleoperator systems as a function of degree of automation."² The best-known numbered scale comes from cars. SAE International's J3016 standard "defines six levels of driving automation, from SAE Level Zero (no automation) to SAE Level 5 (full vehicle autonomy)."³ In 2026 the Cloud Security Alliance, an industry group, said J3016 "has become the lingua franca for discussing vehicle automation."⁴ As AI agents arrived, meaning AI systems that take actions as well as answer questions, researchers carried the idea from cars over to software.

How others define it

• Law (EU AI Act): an AI system is "a machine-based system that is designed to operate with varying levels of autonomy."¹
• Researchers Kevin Feng, David W. McDonald, and Amy X. Zhang: "We define an AI agent's autonomy as the extent to which it is designed to act without user involvement."⁵ Their five levels are named for the role the person plays: operator, collaborator, consultant, approver, and observer. At the top level, the agent "does not require, and comes with no means for, user involvement."⁵
• Margaret Mitchell and colleagues: "AI agents may be said to be more or less autonomous (or agentic) along a sliding scale."⁶
• Cloud Security Alliance: six levels, from Level 0, where "humans perform all actions," to Level 5, which "represents full autonomy, including the ability to set goals and potentially modify its own behavior."⁴

Where they agree and differ

They all treat autonomy as a dial, not an on-off switch, and most borrow the car scale's shape of numbered steps. They disagree on two things.

The first is whether autonomy comes with capability. Feng and colleagues argue that "an agent's level of autonomy can be treated as a deliberate design decision, separate from its capability."⁵ A paper by Meredith Ringel Morris and colleagues instead sets out "levels of AGI performance, generality, and autonomy" side by side and discusses how they "interact with deployment considerations such as autonomy and risk."⁷

The second is whether the top level is a goal at all. Mitchell and colleagues found that "risks to people increase with the autonomy of a system,"⁶ and their paper's title states their view: Fully Autonomous AI Agents Should Not be Developed. This is the strongest view against treating Level 5 as the finish line, and we take it seriously. The fair reply is that a scale describes systems. It does not recommend any of them, and naming the top level clearly helps a leader decide not to use it.

In the news

In January 2026, Cloud Security Alliance chief executive Jim Reavis published a six-level scale for agentic AI. He described it as "deliberately echoing the structure of SAE J3016 while adapting it to the unique characteristics of AI systems."⁴ At Level 1, "each action requires explicit human approval before execution"; at Level 3, "the AI makes decisions and takes actions autonomously within defined boundaries."⁴ It is one industry group's proposal, not a law or a formal standard.

What it means for your organization

Our view so far, which we are still testing: the most useful thing about a level is that someone has to choose it. If autonomy is a design decision, as Feng and colleagues argue,⁵ then every agent's level is a leadership choice. That fits the rule in The Knowledge Core that a person stays in the middle of every chain of command; see [a person in the middle](/definitions/a-person-in-the-middle). We would change our minds if evidence showed that agents' levels mostly end up wherever their technical ability allows, whatever leaders decide.

One question to ask your team: For each AI agent we use, what level of autonomy did we choose, who chose it, and who approves the actions it takes?