Autonomy levels
Autonomy describes the highest action boundary the workflow may cross: none, draft, recommend, decide with approval, or bounded action.
Definition
Autonomy describes the highest action boundary the workflow may cross: none, draft, recommend, decide with approval, or bounded action.
V1 rule
Use the most restrictive cap produced by any applicable rule. A cap is a maximum, not a target. Autonomy is about what the workflow may do, not how capable the model appears. The final cap is the strictest limit produced by the task rules.
Operational test
- None: AI must not perform or influence the assessed task.
- Draft: AI may prepare material that a person fully reviews and owns before use.
- Recommend: AI may propose an option or interpretation, but a qualified person makes the decision.
- Decide with approval: AI may select a proposed decision, but it cannot take effect until the accountable approver reviews the exact proposal.
- Bounded action: AI may perform an allow-listed, observable, and reversible action inside externally enforced limits.
Worked example
A CRM assistant may draft field changes and, after exact preview and approval, perform a bounded write. It may not expand the record set, add fields, or treat a prior approval as permission for another action.
Record and reassess
- Record the intended autonomy, every rule-imposed cap, and the strictest final cap.
- Reassess if permissions, approval binding, reversibility, monitoring, or stop behavior changes.
Scope
The Task Fit Check applies autonomy levels from categorical answers and deterministic house rules. It does not inspect the real environment, verify professional credentials, determine applicable law, or certify residual risk.