Autonomy levels are the three settings a decision inside a workflow can be assigned: runs independently, drafts for approval, or remains human. Every decision in a decision map gets exactly one. The assignment is made from two separate ratings — how much judgment the decision requires, and what the consequences are if it is wrong — and it is agreed in writing before anything is built.
Why judgment and stakes are rated separately
Most autonomy decisions get made on difficulty alone, which produces two predictable mistakes. Work that looks clerical gets automated even though a wrong answer is expensive and hard to reverse. Work that looks expert stays entirely manual even though the AI could safely prepare it for someone to approve in seconds.
Rating the two dimensions separately breaks that pattern. A high-judgment, low-stakes decision — which of these accounts deserves an hour of research — is a good candidate for independent operation. A low-judgment, high-stakes decision — approving a credit above a threshold — is not, however mechanical it looks.
The three levels
Runs independently
The coworker has the context and the permission, and the action is safe or recoverable. Nobody is asked to review it, though everything is still logged and monitored.
Drafts for approval
The coworker prepares the work and a person reviews or authorizes it before anything happens. Most customer-facing output sits here, whatever the coworker is capable of.
Remains human
Too ambiguous, sensitive, consequential, or irreversible to delegate. The coworker may assemble the summary a person decides from, but it does not decide.
How the levels change over time
The initial assignment is deliberately conservative. As the coworker accumulates a record on a given decision, some things move up a level, usually from drafts for approval to runs independently once the approver has stopped changing anything. That promotion is a decision made with the client and recorded in the map rather than something that drifts. We do not chase maximum automation: the objective is autonomy that is useful, measurable, and appropriately governed.
Read more in our blog: Judgment and Stakes Are Two Different Axes and An Escalation Is a Feature, Not a Failure.