Every goal proposal carries a confidence score. The threshold bar on a goal splits that score into three zones, each with a different level of human oversight.
- Advisory — the agent surfaces a suggestion for review. Nothing happens until a human approves it. The agent never acts on its own.
- Supervised — the agent prepares the action (drafts the email, the status change, the task) and sends it to the inbox for one-click approval. Lower friction than Advisory; still human-confirmed.
- Autonomous — the agent acts on its own. The action is logged on the activity feed but does not wait for approval.
How to choose
- Start with Advisory for everything you can't unsend or easily undo (outbound emails, status changes, anything customer-visible).
- Move to Supervised once Advisory proposals are usually correct on first read — the friction is just clicking accept.
- Reserve Autonomous for goals where:
- The action is internal (a task, a tag, a note).
- The action is reversible.
- The track record from Supervised mode is consistently right.
Reading the threshold bar
The bar shows how today's matched records distribute across the three zones at the boundaries you've set. If everything is landing in Advisory, your boundaries may be too strict — you're paying review cost for high-confidence cases. If everything is in Autonomous, you're not catching enough proposals before they fire.
Cold-start
New goals haven't seen enough cases for confidence to be meaningful. Vertiqa marks a goal as collecting or warming until calibration data exists. See Calibration.