The five levels of AI leverage, and why most people stop at two
There are five distinct levels of working with AI, and most people are at level two. The difference between levels is not a better prompt — it is a different relationship with the machine.
Nearly everyone asking ChatGPT questions is at level one or two, and cannot see the shape of what is above them. The people who appear to be getting dramatically more out of AI are not writing better prompts. They are operating from a different position entirely — and the gap between each position is not incremental. It is structural.
There are five levels. This is how I draw them.
Leverage is not a finance word here
Leverage, in this context, is the ratio between the work you put in and the work that comes out. A hammer and nails gets you roughly one hour of output for one hour of effort. A power tool doubles that. A well-briefed crew working overnight while you sleep changes the ratio to something you cannot achieve by working harder.
AI moves that ratio in steps, not smoothly. Each level is a different kind of relationship — a different answer to who is doing the work, and who is checking it. The tool does not change between levels. The relationship does.
It is my favourite way of explaining AI. I use this AI leverage visual, I guess, analogy in a workshop I ran at the consulting firm, to teach how people should be perceiving AI.
Chapter 1.4 · 00:42
Level one: you ask, the machine answers
The most common relationship with AI. You bring a question and take back an answer, and then you do something with it. The work still happens in your hands. The AI is a faster reference book.
In the cubby house analogy: you are building everything by hand with non-electrical tools — hammer, screwdriver, measuring tape. It is slow, it is rewarding, and because it costs you everything, you may never actually finish the cubby house.
There is nothing wrong with level one. Not every task should climb. A quick factual lookup, a gut-check on wording, a translation — these live permanently at level one, and forcing them higher adds cost without benefit. The problem is when level one is the only relationship you have. The ceiling is roughly equal to how fast you can read and act on advice.
Most people never leave level one because they do not know the other levels exist.
Level two: you delegate a task
Here, you hand over a whole job rather than a single question. You are not asking how to write a project brief — you are asking the AI to write one. The work happens on the machine's side; your job becomes reviewing and directing.
The power drill: you are working faster, but you are still the one building. AI drafts, helps with one step, and waits for you to take it from there.
This is where the second-largest group stops. Delegation at this level is genuinely useful — most people who reach level two report real time savings — but it has a ceiling too. The moment you stop asking, it stops working. Nothing is running unless you are running it.
Level three: an agent holds the project
This is where the model changes.
A single bounded agent — not you — is responsible for a project. You brief it on the goal, the constraints, and the standards. It goes off and gets quotes from suppliers, builds a budget in a spreadsheet, drafts a schedule, and checks back at agreed points. You are not directing each step. You are working a review checklist at agreed checkpoints.
The robot helper in the analogy does the screwing and measuring and timber-cutting through the night. You wake up, work through the checklist, approve what is good enough to proceed, redirect what is not.
The key shift here is what you are doing: you have swapped doing for reviewing. Reviewing scales. You cannot do twenty things simultaneously, but you can check twenty things. One person can hold multiple projects moving at once.
The checkpoint is the mechanism. You define what good enough looks like before the agent starts. The checklist is the control — not proximity to the work. Most people who try level three and fail do so because they gave up the checklist and went back to watching every step. That is not level three. It is level two with more complexity.
Level four: a supervisor agent coordinates the crew
At level three, you coordinate the agents. At level four, an agent does that coordination on your behalf.
You direct one supervisor. The supervisor coordinates the specialists. It returns only the decisions that genuinely need a human — a supplier has changed a price, a delivery date has shifted, a choice between two options that the brief did not anticipate. Everything else moves without you.
The analogy: you are inside watching the footy. The agent who used to receive your instructions has become the supervisor. It is inspecting the quality of the work, instructing the crew, and calling you when something needs a call. You are not removed — you set the standards, you hold the final authority — but you are not building.
This is where the work genuinely runs while you are not watching. The system sends notifications rather than questions. Your job becomes holding the standards clearly enough that the supervisor knows when to escalate.
I am not at level five. I am at level four. I am planning level five, but I don't have time, because I am still trying to get this set up.
Chapter 1.4 · 13:04
The honesty matters to the argument. The levels are described from inside the climb, not from the top of it.
Level five: one system manages the portfolio
Level five is not a bigger version of level four. It is a different model.
One AI-native system manages several separate streams of work — different products, different projects, sometimes different companies — each with its own dedicated agent trained specifically on that context. The site D project manager knows everything about how site D operates, and has been fine-tuned for it. It knows about the hill.
The analogy extends to a construction group running four or five sites simultaneously, each with its own supervisor, each reporting into a weekly portfolio review. You are no longer operating a business. You are the owner of the infrastructure that runs several businesses.
Sam Altman and others have made public bets about when the first one-person company worth a billion dollars will exist. Level five is the architecture that makes it structurally conceivable. A handful of people are attempting it. Most of them are figuring it out as they go, because the tools that make it possible became accessible recently enough that there is no established playbook.
Software businesses and AI-native services fit this model. A business that requires physical labour or large inventories does not, at least not yet.
Why most people stop at two
Not for lack of skill.
Level three asks you to give up watching the work happen in real time. That trade feels like losing control long before it feels like output. The natural response is to stay involved — to check each step, to redirect continuously — and that puts you back at level two with more overhead.
The actual mechanism is the review checklist. You define the standard up front, the agent works to it, and you check at the gates. Skipping the checklist is the single most common reason level three fails.
The jump from three to four requires the same shift at a higher level: trusting a supervisor to coordinate specialists rather than coordinating them yourself. That trust gets built by seeing the supervisor work well on smaller, lower-stakes jobs first.
Placing yourself honestly
The useful exercise is not identifying your goal level. It is identifying the level where things actually worked last week — consistently, safely, without constant intervention.
That is your baseline. The next level is usually one specific capability you have not yet built: a review checklist with real quality criteria, a supervisor brief with clear escalation rules, or the habit of leaving the work running overnight rather than checking it every hour.
Not everything should climb. The art is knowing which jobs warrant a higher level, and which ones belong at one permanently.
Drawn from chapter 1.4 of AI Magic 2033.
