The ladder
Shadow. Advisory. Assisted. Autonomous.
Autonomy is earned, not granted. Most people and most companies are either stuck on the first rung, or have jumped to the fourth without earning it. Both cost you: the first in time, the second in trust.
I did not invent this to sell workshops. It is how I decide what my own agents are allowed to do in production, and it is how I taught six people to go from nothing to using AI every working day without getting burned. The ladder is the same whether the thing climbing it is a person, a team, or a piece of software.
The rule that makes it work: you do not move up because you feel ready. You move up because you can show the thing the next rung requires. That is what stops people jumping straight to rung four and finding out in front of a customer.
Rung 1
Shadow
AI watches and drafts. You do the work.
You keep doing the job the way you always have, with the model alongside. It drafts, summarises and suggests, and you compare its output with your own. This is where you learn what it is good at, and where it is confidently wrong in your field.
What it looks like
- A second window open while you work, not a replacement for the work.
- A running list of wins and failures, dated.
- No output goes anywhere without you rewriting it.
You can name three tasks it does well for you and three it gets wrong, from your own notes rather than someone else's blog post.
Rung 2
Advisory
AI proposes. You decide.
The model starts producing first drafts you actually use. Your prompts stop being one-offs and become assets you keep, refine and share. A review habit forms because you have seen enough failures to know what to check.
What it looks like
- Saved, versioned prompts for the five things you do most.
- A short review checklist per task type.
- You can explain to a colleague why a given output is safe to send.
A prompt library you own, a checklist you apply to output before it leaves your hands, and a clear rule for what never goes into a chat window.
Rung 3
Assisted
AI does. You check.
The model does whole tasks end to end and you review the result rather than the steps. Guardrails and checks catch the failure modes you catalogued on the rungs below, so review is fast because it is targeted.
What it looks like
- Agents or workflows that run a task start to finish.
- A deterministic check after the model, not a plea inside the prompt.
- Review time per task measured, and falling.
For each task you hand over, a written failure list and a check that catches each failure. If you cannot write the check, the task stays on rung two.
Rung 4
Autonomous
AI runs. Evidence checks.
The model runs without a human in the loop, because the evidence says it can. Evaluations run on every change, a regression fails the build, and a task is promoted to this rung only when its scorecard has been green for long enough to trust.
What it looks like
- Evals as a CI gate, red on regression.
- Golden datasets built from real failures, not synthetic ones.
- A promotion and demotion record per task, so autonomy is earned and can be lost.
An evaluation suite with real examples, a threshold you agreed before you saw the results, and a way to demote the task the moment the numbers slip.
What the ladder cannot do
It will not tell you which rung is right.
Plenty of tasks should live on rung two for ever. Anything with a legal, financial or safety consequence usually should. The ladder tells you what it costs to move up and whether you have paid it. Whether the climb is worth it is your call, and I will tell you when I think it is not.
Not sure which rung you are on?
That is the most common starting point, and the easiest one to fix. Send a few lines about what you do and what is not working.