AI Agent Autonomy Levels

AI agent autonomy levels describe how much an agent is permitted to do without a person's involvement, from recommending an action for a human to take, to acting and notifying, to acting with no review. In enterprise HR the level is set per task type: deterministic steps such as calculating a balance, verifying a document, or routing an approval can run unattended from day one, while judgment-based steps route to a person.

The common ladder has three rungs. Assisted: the agent prepares, a human reviews before anything is committed. Semi-autonomous: the agent acts and the human is notified. Fully autonomous: the agent acts with no review. A second lens is the type of work rather than the level of review: routine flows that follow a known template, novel situations where the system constructs a path and a person validates it, and one-off dynamic cases at the edge. Most of the value in HR is in the first two, not the edge. The governance principle is that the organization sets the ceiling, users set preferences within it, and autonomy expands only where the system has demonstrated correctness, measured by how often its actions are reversed or overridden. The misconception is that autonomy is binary and configured once at implementation. Darwinbox Cortex is designed around a capped-autonomy model in which the organization sets the limit and the scope of unattended action grows beneath it as the system earns trust.