Approvals and autonomy
How approvals work
The checkpoint between prepared work and executed work.
4 min read
Approvals are the boundary between what Operator may prepare and what a person decides. This guide explains how the checkpoint works and what a reviewer sees.
Principle
Operator prepares. People decide.
Preparation is free. Execution of anything consequential waits for a person.
Operator AI is designed around controlled autonomy. You increase autonomy for specific action types rather than enabling an unrestricted global mode.
What approvals control
An approval sits between prepared work and executed work. It applies to the action types you decided should not proceed on their own:
- communication that leaves the company
- financial activity
- deletion or bulk changes to records
- changes to commercial terms
How it works
Approval is normal product behaviour rather than an exception. Preparation runs freely, risk decides what may continue, and the action types you protect stop for a person before execution.
- 01 / Prepare
Work is prepared
Internal preparation continues without a checkpoint.
- 02 / Evaluate
Risk is assessed
The action type decides whether it can proceed on its own.
- 03 / Approve / Human checkpoint
Human checkpoint
Sensitive actions pause with the evidence behind them attached.
- 04 / Execute
Approved action runs
The decision stays attached to what happened next.
What a reviewer sees
An approval carries the evidence that produced it: the trigger, the context Operator used and the action it proposes. Approving without that context is guessing.
Example
A drafted reply to a customer arrives with the original message, the account record it relates to and the follow-up Operator would schedule after sending.
What stays visible
The decision stays attached to the action: what was proposed, who approved it and what happened afterwards.
Key takeaway
Approvals are the product, not the friction. They are what makes autonomy safe to widen.