Coach Denis

There are five rungs to using AI. Almost nobody gets past the first.

Anthropic published a table of the five steps of AI adoption in July. Microsoft had published the same thing in May, with numbers. At every rung, what blocks you is never motivation - and the most useful part of the document isn't the list of rungs.

A small workshop ladder of brass and pale wood with five rungs, standing alone on a dark workbench under warm light; two open hands are held on either side, at a distance, not touching it.

Let me describe a scene. You hand your AI a task. While it works, you sit there. You read every sentence it produces, you approve every step, you never look away from the screen.

On paper, you’re working with AI. In practice, you’re the driving instructor in the passenger seat, foot hovering over the second brake, while the car drives.

And when it stalls, almost everyone charges it to the same account: their own discipline. “I can’t bring myself to let go.” “I’m too much of a control freak.” “I’ve never been comfortable with tech.” It’s a comfortable explanation, because it can be fixed by gritting your teeth.

It doesn’t survive reading the two documents published this summer.

What the document says, and what it is

On 16 July 2026, Boris Cherny - the creator of Claude Code, at Anthropic - published Steps of AI Adoption. It isn’t an essay, it’s a table: five rows and six columns. The step and your role, the number of agents, what it looks like, what the bottleneck is, the products that help, the guardrails.

Better to say this up front than bury it: of those six columns, two are a catalogue of Anthropic products, and the “guardrails” column talks about SSO, per-seat spend caps and compliance - so it speaks to IT departments, not to you. That leaves four useful columns. They genuinely are.

Two months earlier, in May 2026, Microsoft had published its own in the Work Trend Index: four modes of collaborating with AI, drawn from its own telemetry and a survey. The names differ, the boundaries overlap, and crucially: Microsoft publishes numbers.

Two competitors describing the same climb from two different datasets. At that point, we can stop calling it an opinion.

Here is the ladder, translated out of the terminal. For each rung: what it looks like in your week, what it unlocks (the table calls it the unlock, and it’s the column everyone forgets to copy across), and what blocks you.

Rung 0 - Gated

You have the tool, you don’t use it on your real work.

Inside a company it’s explicit: only older models are approved, access runs through homemade layers of authentication, and there is no path at all for hosting what the AI produces. Cherny sums it up with a line that stings: outputs only exist locally. Work gets made, nothing ships.

When you’re independent it’s quieter, and you’re the one holding the lock: I can’t put client data in there, I’ll get to it when I have time. Same result. The AI ends up rewording emails and brainstorming titles.

What blocks you: legacy approval processes, and an obsession with cost per token rather than outcomes. Hold on to that one, it comes back everywhere: counting what it costs while never counting what it returns.

Rung 1 - The pair

One task, one AI, and you watching all of it. One session at a time, and you review almost everything before it ships.

This is the driving instructor. The table doesn’t say it’s worthless - quite the opposite.

What it unlocks: a change that used to fill an afternoon becomes something you finish between meetings. That’s real, and it’s already good. But it stops there.

What blocks you: your attention. And this is the most important sentence in the whole article. Cherny names the mechanism without hedging: because of low trust in the model’s output and a lack of self-verification, you feel you must read everything, so you never look away.

That isn’t fussiness. It’s rational. Nothing else is checking the work, so it’s you or nobody. He adds the detail that really lands: the work is synchronous. You sit and watch the machine work, rather than moving on to the next task.

At Microsoft, this rung splits in two: Author (you produce, the AI helps a line at a time) then Editor (you set the intent, the AI writes the first draft, you edit and approve). This is where the overwhelming majority of people live.

Rung 2 - The one giving the orders

You write down what you want and what “correct” means, you launch two or three tasks, each in its own space, you go do something else, you come back and judge the result.

The words that change everything in that row: the AI checks its own work before you see it. What you look at has changed in nature - no longer the process, only the output. Microsoft calls this Director: you write a spec and hand off entire tasks, which run in the background.

What it unlocks: a backlog that used to take the team weeks becomes one person’s afternoon of orchestration. Translated for you: the pile of things never done, because nobody ever had the time.

What blocks you: your review throughput. You hand-write less yourself, but you’re checking six streams of it instead of one, and that takes up more of your time than before.

Rung 3 - The supervisor

Routines run without you triggering them. You stop looking at tasks and start looking at exceptions. Microsoft calls this Orchestrator.

The mental shift is captured in one line from the table that holds for any trade: “did you read what came out?” becomes “what context was the machine missing, and how do we solve that for next time?”. You stop correcting outputs and start correcting the system that produces them.

What it unlocks: the maintenance and cleanup that used to wait for someone to find the time now runs continuously, in the background.

What blocks you: trust in the loop, and the speed at which your team decides. With the trap named in black and white - more on that below.

Rung 4 - Intent

The loop is fully closed. You stop handing out tasks and hand out a direction: you steer by intent and monitor by exception.

What it unlocks: the migration that used to take a quarter becomes a workflow you kick off and check on.

Let’s be honest: for an independent in 2026, this rung is an asymptote, not a target. It’s in the table because reasoning needs an end point, not because you’ll get there this year.

What every rung has in common

Read the blockers again. Permission. Attention. Review throughput. Trust. Scope.

Not one of those words is “motivation”. Top to bottom, same verdict: nobody is blocked by a lack of willpower. They’re blocked by a verification system that doesn’t exist yet.

And Microsoft puts a number on it, probably the single most useful figure in its whole report: organisational factors - how work is carved up, what gets recognised, which practices exist - weigh more than double individual factors like mindset and behaviour. 67% against 32%.

Two more numbers from the same report, to be read side by side:

  • 65% fear falling behind if they don’t adopt AI.
  • 13% are rewarded for reinventing their work with AI.

Everyone is afraid, almost nobody is paid to make the one move that works. In that gap, it’s entirely normal that each person concludes they’ve personally failed. They haven’t, and it’s measured.

One last figure, the one that should reassure two thirds of my readers: asked which human skill becomes most critical, the top answer is quality control of what the AI produces (50%), ahead of critical thinking (46%). Your judgement doesn’t become useless. It moves up a floor.

The most useful part of the document isn’t the list of rungs

It’s what’s written between them.

Between each row of the table sits a discreet band: how to get from step 0 to 1, from 1 to 2, from 2 to 3, from 3 to 4. That’s where all the actionable content lives, and it is invariably the part that the summaries going round drop in order to keep the five neat labels.

But knowing your rung is useless. The five labels are a diagnosis, and a diagnosis cures nobody. What you want is the step. And for you it’s almost always the same one: 1 to 2.

The four jobs that take you from rung 1 to rung 2

The band in the table fits on one line: run more than one agent at a time; a self-verification loop you trust; remove the blocking permission prompts; automate review.

That’s written for developers. Here is the transposition for real desk work. Budget a week of evenings, not a quarter.

1. The contract. Everything you repeat out loud, session after session - your tone, your clients, your rates, what you never touch, how you want sources handled - goes into a file the AI reads at every start. As long as an instruction stays spoken, it’s forgotten by the next conversation. Written down, it stands on its own, and you’ve just deleted half your supervision work.

2. The definition of “good”, written before you launch. This is the heart of the step, and it’s what the table calls the self-verification loop. Not “write me a proposal”, but: one page maximum, price in figures and in words, no promises about deadlines, the client’s name spelled as it is in the contract, and list every point where you had to guess. Then the rule, non-negotiable: the AI has no right to show you work it hasn’t first checked itself against that list, and it tells you what failed.

3. The sandbox, and the end of repeated permission prompts. An isolated working copy: a duplicated folder, a draft file, never the master document. A bad run has to cost zero. And settle once and for all what happens without you and what always goes through you: money, anything sent outside, and deletions stay in your hands, the rest is approved in advance. It’s explicitly on Cherny’s list, for a simple reason: every window asking for permission puts you back in the instructor’s seat.

4. The reviewer who produced nothing. Before the work reaches your desk, have it reviewed by a fresh conversation that wrote none of it, with a single mandate: find what breaks. The maker never judges themselves. You then read only the final version, with the review report next to it.

One clarification that isn’t mine, it’s in the table’s “guardrails” column: the final review holds to the same quality bar whether the work came from a human or a machine. Your standard doesn’t drop a millimetre. What changes is when you step in: at the end, on an outcome, never during, never on keystrokes.

Where you actually are: two tests

The physical test. Launch two tasks this morning, close the screen, go walk for two hours. If the idea is unbearable, your verification loop isn’t finished. It isn’t your courage that needs work, it’s the loop.

The three questions. They also work for sizing up a supplier, a service, or your own management when you’re told “we’ve moved to AI”:

  1. How many tasks run in parallel on a normal day, and which ones?
  2. Show me the check. The one that fires without a human pressing anything.
  3. Who reviews the final deliverable, and against which written criteria?

A rung 0 answers with the name of its subscription. A rung 2 answers with a procedure and criteria. You’ll spot the difference in thirty seconds.

The instructor doesn’t lift off the brake the day he gets brave

He lifts off the day the student can drive alone, safely.

You don’t move up a rung by forcing yourself to look away while the AI works. You move up the day your check exists - the contract, the criteria, the sandbox, the reviewer - and runs without you.

Your attention then goes back to being what it should never have stopped being: the resource you reserve for decisions, not for surveillance.

The challenge, if you want to do it this month: take a task that has been sitting in your list for too long. Build the four jobs around it. Bring it up to rung 2 - launched and finished without you stepping in between the start and the final check. Then write down two things: where you started from, and what you missed most while you let it run alone. That second answer is what tells you what to build next.

The sources

← Back to the blog