Autonomy levels are a graded scale describing how independently an AI agent may act, from only assisting a person, through drafting and acting with approval, to acting alone within bounds. They give organisations a shared vocabulary for how much they delegate. Their weakness is that they usually grade the agent as a whole, while risk sits in individual actions.
Where do autonomy levels come from?
Several bodies and vendors have published scales. The Cloud Security Alliance, for example, has published autonomy levels for AI agents running from 0 to 5. Most scales share a shape: the lowest level has the agent suggest and a person act; middle levels have the agent act with approval or within narrow limits; the top level has the agent act on its own across a broad scope.
The scales are useful for describing intent and for comparing systems. How they work in operation is covered in how autonomy levels for AI agents work.
Why do autonomy levels matter once agents act?
Because they turn a vague question, "how much should we trust this agent?", into a decision someone can record and review. Without a scale, teams grant autonomy implicitly, by connecting a tool and seeing what happens.
But a single level per agent hides too much. A servicing agent might be excellent at updating addresses and poor at handling hardship disclosures. Giving it "level 3" overall either holds back the first or over-trusts the second. Whether levels are the right tool at all is weighed in are autonomy levels the right way to govern AI agents?
How should autonomy levels be used in practice?
Apply them per capability, not per agent. Set a ceiling for each capability from the business's risk appetite. Grant actual autonomy within that ceiling only on evidence from the agent's performance on real cases. Let it expire, so evidence is refreshed. And enforce it at the moment of action, not by trusting the agent to remember its level.
Our depth scale follows that approach: 1 read, 2 draft for a person to send, 3 execute routine work within bounds, 4 move money or take hard-to-undo actions, 5 coordinate other agents. Each is granted per capability, per customer.
What does that look like in an example?
A lender's servicing agent has five capabilities. Answering balance questions runs at depth 1. Drafting arrears letters runs at 2. Updating contact details runs at 3. Agreeing payment plans was proposed for 3 but certification showed weak handling of vulnerable customers, so it stays at 2. Releasing refunds is capped at 2 by the lender's ceiling regardless of evidence. A single autonomy level could not express any of that.
How are autonomy levels different from an autonomy ceiling?
A level describes how much independence an agent has. An autonomy ceiling is the most it may ever have for a capability, set by the business in advance. The ceiling caps the level; evidence fills it up to the cap. Both sit inside the wider idea of bounded autonomy. The argument for certifying capabilities rather than grading agents is in autonomy levels are not enough: certify the capability, and the broader case in regulated autonomy.
Questions people ask
Is the highest autonomy level the goal?
No. The goal is the right level for each capability. For many actions in regulated work, the right level stays at draft or staged for release permanently.
Do regulators require a particular autonomy scale?
Not as such. Rules such as the EU AI Act's human oversight requirement for high-risk systems require effective oversight, which a well-applied scale can help show. This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.