The autonomy ladder

I do not hand an AI full autonomy on anything that matters, and I did not arrive at that by being timid. I arrived at it by watching trust get earned in stages, the same way you would with a new hire, and building the stages into the system on purpose. I call it the autonomy ladder.

The bottom rung is read-only. The AI can look, analyze, and report, but it cannot change anything. Next rung, draft-only: it can produce the work, an email, a post, a code change, but the output sits in a queue and does nothing until I approve it. Next, approve-then-send: it can take the action, but only after I click yes on that specific instance. Higher still, scoped autonomy: it can act on its own within a narrow, well-defined class of actions where I have watched it succeed enough times to trust it. Full autonomy sits at the top, and most surfaces never need to climb that high.

Why build a ladder instead of just picking a level?

Because trust is not a setting you choose once, it is evidence you accumulate. Every rung is a place to watch the model's behavior on real work with the safety still on. You promote a surface up the ladder when it has earned it, and here is the part people forget: promotion is a decision you make, but demotion can be automatic. If the model starts drifting, the system can drop it back a rung without waiting for you to notice.

My newsletter tooling lives on this ladder right now. The draft writer sits at draft-only, it writes the weekly wrap-up and stops, because it has not yet earned auto-send. When it has produced enough good drafts that I trust it, I promote that one surface to auto-send. Not the whole system, just the surface that earned it. Each capability climbs on its own record.

The failure mode is granting full autonomy on day one because the demo looked clean. A demo is a handful of runs on happy-path inputs. Autonomy is thousands of runs on inputs you never imagined. The ladder is how you find the gap between those two safely, one rung at a time, instead of discovering it in production.

The takeaway: Autonomy is earned in stages, not granted at once. Start read-only, climb to draft-only and approve-then-send, and only promote a surface when its own track record says it is ready.