What are levels of agent autonomy?
Levels of agent autonomy are a graded scale that describes how much an AI agent may decide and do without a person. There is no single standard. Many published scales borrow from the levels used for self-driving cars, and the five below are typical.
- Suggest. The agent drafts or recommends, and a person makes every decision and every change.
- Act with approval. The agent prepares each action, and nothing runs until a person approves it.
- Act within limits. The agent acts alone inside set limits on amount, record type, or risk, and a person handles the rest.
- Act and report. The agent runs the whole process, and people review samples, logs, and exceptions afterward.
- Full autonomy. The agent sets its own subgoals and acts without routine review, which few businesses allow on a system of record.
Why levels of agent autonomy matter
The scale gives process owners, IT, and compliance staff shared terms for deciding how far to trust an agent. The first two levels are what most teams mean by human-in-the-loop. Autonomy is set for each type of action. One agent might read data freely, update a customer record only with approval, and never delete anything. The right level depends on whether the action can be reversed, what an error costs, case volume, and what regulators expect. The EU AI Act requires effective human oversight of high-risk systems. The General Data Protection Regulation (GDPR) restricts significant decisions about people made by automated means alone.
Levels of agent autonomy in practice
A property and casualty insurer introduces an agent for auto glass claims. At first, the agent assembles each claim file and recommends a payment, and an adjuster decides. After several months in which adjusters rarely change its recommendations, claims from in-network shops move to the approval level. Later, claims under a set amount on policies with no recent losses move to acting within limits. Claims with a fraud flag or a coverage question stay with an adjuster at every stage.
In Kirk Tech Solutions engagements, every action that changes a system of record starts at the approval level. The approval logs then become the evidence for any move up.
What to watch for
- Record the level for each action type where auditors and process owners can find it.
- Enforce the level through permissions and approval gates in the platform rather than through instructions in the prompt.
- Watch for automation bias at the approval level, where reviewers start approving without reading.
- Return to a lower level after a model change, a prompt change, or a change in the underlying process.