The one confusion. People think the ladder is about access — that once you have the subscription you are on it. You are not. Access is rung zero. Every rung above it is about agency: what you ask for, what you check, what you delegate, and what you refuse to hand over. It takes intelligence to use intelligence, and the most common failure is not refusing the tool — it is accepting whatever it hands back.
The refusal (1–4)
| 1 | Hostile. It is theft, it is hype, it is beneath me. The position is moral and it is fixed. |
| 2 | Avoidant. Accepts it exists. Arranges not to touch it. |
| 3 | Forced. Uses it because work put it there. Resents it. Does the minimum. |
| 4 | Novelty. Tried it, got a party trick, concluded that is what it is. |
The opening (5–8)
| 5 | Search box. Asks it questions the way you ask Google. Short, cold, one shot. |
| 6 | First real answer. Notices the reply got better when the question did. This is the hinge of the whole ladder. |
| 7 | First saved hour. Uses it on something that actually mattered and gets the time back. |
| 8 | Habit. Opens it daily without deciding to. |
Competence (9–12)
| 9 | Instructs instead of asking. Context, audience, format, constraints. |
| 10 | Iterates. Sends it back with what was wrong instead of starting over or giving up. |
| 11 | Knows what it is bad at. Checks the confident ones. Has been burned once and changed behaviour. |
| 12 | Keeps what worked. Saves prompts, reuses them, notices which ones keep earning. |
Fluency (13–16)
| 13 | Brings its own material. Feeds it documents, data, a real corpus — stops relying on what it happens to know. |
| 14 | Picks the tool. Different models for different jobs, and can say why. |
| 15 | Builds a workflow. A repeatable path, not a fresh improvisation each time. |
| 16 | Teaches someone. Explains it without jargon and watches them get their first saved hour. |
Where evangelism actually lives. Not at the top — at 16. You can teach the moment you are about four rungs above the person in front of you, and a 16 teaching a 5 is often better than a 24 doing it, because a 16 still remembers being stuck. That is what makes a million people reachable: the teaching capacity sits in the middle of the ladder, where most people can get to.
Leverage (17–20)
| 17 | Builds the context layer. Puts what they know into files the machine can read — one source of truth instead of retelling it every session. |
| 18 | Delegates whole jobs. Hands over work, reviews the result, sends it back with a reason. Judgment moves from doing to deciding. |
| 19 | Lets it touch real systems. Files, data, money-adjacent things — with backups, an undo, and a way to see what happened. |
| 20 | Designs the harness. Loops, stop conditions, permissions, gates: what may run unattended and what must stop and ask. Decides the failure it can live with. |
Building (21–23)
| 21 | Writes the agent. The loop, the tools, the drivers — not just the instructions. |
| 22 | Measures it. Evals against known-good cases, traces, cost per run. Can prove it got better rather than feeling it. |
| 23 | Ships it to other people. And maintains it when it breaks for someone who is not them. |
The frontier (24–25)
| 24 | AI engineer. Serves, tunes and operates models in production. Full-time technical specialism. |
| 25 | Researcher. Advances what the models themselves can do. |
How to use it
Find the highest rung you do habitually, not the highest you have ever managed — the ladder is about defaults, not peaks. Then climb exactly one. Nobody goes from 4 to 14, and nobody needs to reach 24: rungs 1 to 20 are available to anyone, and none of them require writing code. The break between 20 and 21 is where it becomes a technical trade.
The rungs that move the most are 6 (a better question gets a better answer), 11 (getting burned and changing behaviour), 16 (teaching someone) and 18 (delegating a whole job rather than a task). If you are stuck, you are almost always stuck below one of those four.
Words you will see
Context
Everything you hand it before the question. The lever most people never pull.
Agent
A model in a loop with tools it can use. Rungs 19–21.
Gate
A place it must stop and ask a human. What makes unattended work safe.
Eval
A fixed set of tasks with known-good answers. The difference between feeling improved and being improved.
Deliberately not here
Which model to buy, which app to use, and any prompt you could copy — all three date within months and none of them move you a rung. The ethics and the economics are real and are their own conversation; this ladder is only about capability. And there is no rung for reading about AI. Reading is how you find the rung; doing is how you leave it.
The AI fluency ladder · one pager · September 21, 2026 · For the moonshot: agency at scale. Access is not the bottleneck.