What is Calibrated autonomy?
Calibrated autonomy is the design pattern where an AI agent auto-executes low-risk actions, asks for owner approval on medium-risk actions, and always escalates high-risk actions — with the threshold configurable per action type.
Calibrated autonomy is the design pattern that lets an AI agent do real work on your behalf without making catastrophic decisions. Instead of "fully autonomous" or "always-ask", the AI runs every action through a policy filter that classifies its risk and routes it accordingly.
In practice, a calibrated-autonomy AI agent operates at three levels:
- Auto-execute — low-risk actions like drafting copy, editing one paragraph of an About page, listing recent orders, or sending tracking links. These happen immediately with no approval needed.
- Approval queue — medium-risk actions like posting a journal entry, sending a broadcast to your full customer list, deleting a product, or promoting an A/B test winner. These wait in your inbox for review.
- Always escalate — high-risk actions like refunds, disputes, custom-order pricing, or anything you've specifically flagged. These always come to a human with full context attached.
The owner sets the threshold per action type. A first-week user typically keeps most things in "ask first" mode; a year-three user lets the AI auto-run more operations as trust builds. Every action is logged and reversible — most calibrated-autonomy systems include a one-click undo so even auto-executed actions can be rolled back.
The pattern emerged from the recognition that fully autonomous AI is too risky for production small-business use (one bad bulk-update can wipe out a catalog) while fully manual AI defeats the productivity gain. Calibrated autonomy is the middle path that compounds trust over time.
Common questions
Can I configure calibrated autonomy per skill?
Yes. In WoBooks Companion you can set each of the 68 skills to "manual" (always ask), "approval queue" (run after my approval), or "auto" (run immediately). Most owners start with the safe defaults and relax specific skills as they build confidence.
What happens to actions in the approval queue?
They wait in your dashboard with full context — what the AI proposed, why, and what data it used. You can approve as-is, edit and approve, or reject. Approved actions execute immediately; rejected actions are dropped and Companion learns from the rejection.