Autonomy is a dial, not a switch.
Hendl starts human-first: agents observe, draft and ask. As their work proves out against your reviewers, you grant more autonomy — one task, one notch at a time. And the dial turns both ways.
From observing to operating — in earned steps.
Every level keeps a human in the picture. What changes is when they're involved — before the action, or reviewing a sample after it.
Agent — watches the case and suggests
You — do the work yourself
Agent — prepares the work in full
You — review and send
Agent — acts once you sign off
You — approve each action
Agent — acts, then reports back
You — are informed, can reverse
Agent — acts within hard limits
You — sample and audit
Dialled per task, not per platform.
A chaser email is not a settlement decision. Every step in a playbook carries its own level — so routine work runs free while judgement stays gated.
Where you set it.
The dial reads the case, not just the task — value, risk, sector rules and team readiness all pull it up or down.
A chaser email is not a settlement decision. Every step in the playbook carries its own level.
Thresholds on case value, exposure or fraud signals pull the dial down automatically.
Regulated decision points are pinned to human approval — the dial won't go past them.
Start new teams at observe-and-draft; open up as reviewers build trust in the output.
Every notch is earned with evidence.
The platform recommends a level change and shows its working. A named owner approves it — levels never move on their own.
Every suggestion is compared with what your team actually did. Sustained agreement is the case for the next notch.
When reviewers stop correcting an action, it's a candidate for more autonomy. When overrides rise, the dial steps back.
At higher levels, a fixed sample of every agent's work still lands in a human queue — autonomy never means unwatched.
Questions teams ask about the dial.
Can autonomy increase on its own?
No. The platform recommends a change with the evidence behind it; a named owner approves it. Levels never move without a human decision.
What happens when we turn the dial down?
It takes effect immediately. In-flight work finishes at the stricter setting, and pending actions return to the approval queue.
Is there anything agents can never do?
Yes. Hard limits — payments over threshold, declines, anything you pin — sit outside the dial and always require a person, at every level.
How long until a task earns L3 or L4?
Typically weeks, not days — enough volume for the agreement and override evidence to be meaningful for that task in your book.