Following a good plan feels like making one.
AI does not only change how much work a manager ships. It changes what kind of manager he becomes.
A manager asks his AI for the go-to-market plan. It comes back good, so he approves it. Then the restructure plan. Good. The post-mortem outline, the board pre-read, the competitor analysis. All good, all approved. He ships more than he ever has and feels sharper for it.
He has stopped making plans. He’s executing his AI assistant’s. He hasn’t noticed it, because following a good plan feels almost the same as making one…
Run that for a year and something happens to him that nobody measured until recently.
Doctors who spent a few months working with an AI assistant got measurably worse at their own job without it. Experienced endoscopists, the kind who catch polyps during a colonoscopy before they turn into cancer. They worked with a tool that flags suspicious tissue on the screen, and after a few months someone measured their detection rate with it switched off. It had dropped below where they started.
Not juniors. Specialists with years behind them, in a field where a missed adenoma can become a missed cancer. A few months of letting the machine watch alongside them was enough to dull a skill they’d spent a career building.
That’s the manager who follows.
He’s the endoscopist who doesn’t know he’s in the study. The skill fades whether or not anyone is measuring it, and it fades silently, because the work still ships and the plans still look fine.
So am I, most weeks. I build AI products, I use the tools for hours a day, and I’m writing this with the team of agents arguing if fixing my non-native commas is against the guidelines I gave or not. This isn’t a sermon from outside the trap. It’s a note from inside it, about the quiet thing it does to you: it stops you growing.
What the manager handed over was not drudgery. It was the practice.
For over twenty years I watched juniors turn into seniors the same way. They wrote, built, shipped, broke things, explained them badly, fixed them, and tried again. Code, strategies, documents, product decisions. Different outputs. Same mechanism.
We thought we were paying them for the outcome. We were also running a training programme nobody had designed on purpose.
The drafts were the curriculum. The judgment was the product.
Be precise about which part taught, because the drudgery taught nothing. The part that mattered was someone senior showing you why your draft was wrong. Not “fix this line,” but “this whole argument assumes the customer already knows they have a problem, and they don’t.” You wrote the next one, still wrong, wrong in a new way. After a hundred of those you didn’t just write a good strategy. You could see, on contact, whether a strategy was any good, and say why.
AI makes the drafts optional.
That is the problem. Not because drafts are sacred. Most drafts are bad. Many are boring. Some are useless.
But the first draft was never only a document. It was a rep. It forced you to decide what mattered before someone else showed you a better answer. It gave a senior something to correct. It gave you a mistake that was yours, which meant the feedback had somewhere to land.
A few years ago I did Seth Godin’s altMBA. One task: ninety-nine business plans, in under three hours.
When you get such an assignment you don’t think - you have to run with it. Fast. Don’t argue - time is ticking… so I did.
The first ten went slowly. I wanted to think each one through, the way you do when you have a few business ideas you’ve been carrying around, the ones that feel cool in your head. They came out exactly as I expected. Solid. Considered. Boring.
By thirty I was speeding up, because there was no time to do anything else. By fifty I’d stopped asking “is this a good idea” and started asking “what’s the next one.” The last ten were the best of the whole set. The most creative.
Today I could generate ninety-nine business plans in ninety seconds. They would probably look better than mine did. And I would learn almost nothing.
The first ten were my thoughts, what I already had before I started. The last ten were judgment under pressure, and it only showed up through the doing. A capacity I did not have at the beginning, grown in three hours across ninety-nine repetitions.
That’s habituation. You don’t get skilled by reading about the skill. You get there by repetition, with feedback, slowly. Competent means I can do it. Skilled means I can adjust it while I’m doing it, I know why it’s good, I notice the moment it goes wrong. Only the reps get you the second one.
There’s a darker version the word deskilling hides. You can only lose a skill you had.
The junior who starts today, points the AI at every task, and never writes the hundred bad drafts doesn’t deskill. He never builds the skill. It’s almost funny, until you remember he’s three promotions from owning the call nobody else can make.
Now let’s imagine another manager. Same AI, same week, going the other way.
He isn’t refusing the tool. Refusing is dishonest, because the tool does real work. He’s doing two things the follower manager isn’t.
First, he keeps the reps. Before he asks for the plan, he spends fifteen minutes writing his own, or at least writing down what a good one would have to do and what a bad one would look like. Then he reads the AI generated version against his. He’s still producing bad first drafts, in his head now instead of on paper, and still getting them corrected, by the gap between his answer and the machine’s instead of by a senior’s red pen. The fifteen minutes is the rep. It’s the only part the follower skipped.
Second, he points the leverage at execution, not at thinking. He sees more strategies, more structures, more drafts in a week than the follower sees in a quarter. And because he judges every one against his own frame, the volume sharpens his taste instead of replacing it. The leverage multiplies a skill he’s still building. For the follower, the same leverage multiplies a skill he’s quietly losing.
Same AI. The follower aims it at the thinking and hands the thinking over. The one leveling up aims it at the execution and keeps the thinking for himself. That’s the entire difference, and the research is starting to back it. Use it procedurally — ask, accept, move on — and your understanding erodes. Use it constructively — argue, test, compare — and it holds. The pattern is consistent enough to take seriously. It isn’t the tool. It’s whether you did the difficult part or gave it away.
The optimist says AI clears the drudgery and frees the time to do the reps. Maybe. But freed time has a long history of becoming more output, not more practice, because nobody designs for the practice. Bainbridge noticed this about automation back in 1983 and the pattern hasn’t moved. So why doesn’t every manager just choose to be the second one and keep his reps?
Because the gradient runs the other way. Accepting the output is effortless. Forming your own answer first is effortful. Put effortless and effortful side by side and effortless wins, slowly, and you don’t feel it winning. Ferdman calls the result a capacity-hostile environment: one that removes the chances to build a skill and quietly discourages you from looking for them. You don’t beat that with willpower. Most of us aren’t virtuous superheroes. You beat it by changing your own gradient, by making the difficult part the default instead of the thing you have to force.
Write the frame before you read the answer. Before you hand a task to an AI, or to a person, spend fifteen minutes writing down what a good answer looks like. Your criteria. What “wrong” would look like.
Then judge what comes back against your frame, instead of letting the output become the frame. Sounds like more work. It is. The fifteen minutes is where the judgment gets built. Skip it and the AI did the thinking, not you.
My bet: two weeks of this teaches you more than many months of approving good-looking output.
Learning often feels worst while it is working.
Recalling before rereading. Generating your own answer before seeing the model’s. Writing the bad first draft before reading the good one. These are the conditions that feel slow and inefficient, and they are the ones that build the skill.
The smooth version feels better. It also teaches less. There’s a name in learning research for why all of this is so easy to miss. Robert Bjork calls them desirable difficulties. You finish learning sure you’ve got it, and you haven’t. They call the gap the fluency trap.
That’s the AI assistance in one sentence. Its gift is removing the effort, and the effort was the part that was building you.
So the difficulty was never in the way of the work. For twenty years it was the work, the thing that turned drafts into judgment. AI lets you skip it now, and most days skipping it will feel like a favour. The two managers are two answers to the same question: when the friction becomes optional, do you keep enough of it to stay the person who can still tell when the output is wrong?
I know which one I’m trying to be. Some days I even manage it.
References
Endoscopist deskilling after AI exposure: Budzyń et al., 2025.
AI deskilling is a structural problem: Ferdman, AI & Society, 2025.
Your Brain on ChatGPT: Kosmyna et al., 2025.
How AI Impacts Skill Formation: Shen & Tamkin, 2026.
Impact of generative AI on student learning outcomes: Pallant et al., 2025.
Ironies of automation: Bainbridge, 1983.
Making things hard on yourself: Bjork & Bjork, 2011.
Brainrot: Deskilling and Addiction are Overlooked AI Risks: Chalkidis & Søgaard, 2026.