The AI autonomy ladder
Where's your line with AI? The three levels, the split, and a prompt that maps your own tasks. Built to take into work.
What this is for
Most people have one line with AI, drawn on a feeling: “I don't trust it with customer stuff.” The businesses actually getting value from AI don't have one line. They decide, task by task, how much control to hand over, and they move that line as they build trust.
This page is the version you can put in front of your team. The three levels, the trick most people miss, and a prompt that maps your own work onto them.
The three levels
Take one example, a customer email, and watch it climb. Same task, three levels of control.
Human in the loop
You ask it for help writing the email. It drafts, suggests, improves. You are in the conversation and you send. It is helping you do the job.
Human over the loop
It writes the email itself and you approve each one before it goes out. Spot checks, not every word. It does the job, you check its work.
Human out of the loop
It handles that email start to finish and sends it. No check. You just see the results afterwards.
The trick most people miss
You don't pick one level for the whole job. You split it. Take that same inbox:
And here's the part that makes it work: you can use AI to sort the incoming work into those groups, not just to do it. It reads each message and decides which pile it belongs in, so the safe stuff flows and the risky stuff gets escalated to a person. Sorting is a much easier job to trust than replying, so it is often the first thing worth automating.
Map your own tasks
Paste this into Claude (or any AI you use). It interviews you about the work you actually do, then hands back a table: each task, the level it should sit at, and what would need to be true for it to move up. You just answer the questions.
Where each task sits is just your risk tolerance
There is no correct answer to where a task belongs. It is a judgement about how much it would cost you if the AI got it wrong. High stakes stays with a human. Low stakes earns its freedom. And you move a task up a level when you are comfortable, not before, usually after you have watched it do the job well at the level below for a while.
The line also moves over time. As the tools get better and as you build confidence, work that felt risky last year quietly becomes safe to hand over. The skill is moving it deliberately, task by task, rather than all at once or never.
The honest bit
- -This is slower to set up than “just automate everything.” That's the point. The businesses with the AI horror stories are almost always the ones that skipped the ladder and jumped a task straight to out of the loop.
- -It is a framework for thinking, not a safety guarantee. Something at level three can still get things wrong. The point is that you chose to accept that for that particular task, with your eyes open.
- -Moving a task up a level is not permanent. If it starts making mistakes, move it back down. The ladder goes both ways.
Before you automate anything
The tasks you can trust an AI with are the ones where it has the right context about your business. If you are still figuring out AI, that groundwork comes first: get your business AI ready.