Published
Autonomy belongs to actions, not systems
The question is rarely whether to trust AI in general. It is which actions it may take alone. Reading a document is low risk; sending a price to a customer is not.
Setting autonomy action by action lets one system be fast where that is safe and careful where it matters.
Assist: AI prepares, a person acts
In Assist, AI drafts the reply, fills in the form or prepares the work order, and a person reviews and sends it. Nothing leaves your systems unless someone chooses to send it.
It is the right starting point for any new workflow, because every draft shows you how the AI works.
Approve: AI acts, a person signs off
In Approve, AI does the work and stops at defined points for a decision: an offer, a refund, a job that touches safety. The person sees what the AI proposes and why, then approves, edits or rejects it.
Most of the work moves quickly; the decisions stay with people.
Execute: AI acts within limits
In Execute, AI completes routine actions on its own, within limits you set, and passes exceptions to a person. It suits frequent, well-understood work such as booking confirmations or status updates.
Every action is still recorded, so you can always review what was done.
How to decide
Four questions settle most actions:
- What happens if this action is wrong, and can it be undone?
- Does it reach a customer, a payment or a safety decision?
- How often does it happen, and how alike are the cases?
- Has the AI's work on it been reviewed for long enough?
Change it as you learn
Start new actions in Assist. Move an action up when its reviewed work shows it is ready, and keep the sensitive ones in Approve.
Autonomy is a setting, not a decision made once. Review the activity record regularly, widen what works and narrow anything that surprises you.