The promise is seductive: with AI you’re ten times faster. And yes — the machine gets through, in next to no time, a mountain of work that used to cost me days or weeks. That’s real. It’s only half the truth.
The other half sells worse, which is why almost no one says it out loud. I live it every day: a large part of my work hasn’t disappeared because of AI. It has shifted.
Where the work goes
Two things eat up the time you gain.
The first is the infrastructure around the tool. Bare AI is an input line that pauses after a short while and comes back with a question of its own. To turn that into a dependable working process, you need a whole ecosystem around it — tools that guide the machine, safeguard it and keep it in check. What mine looks like I’ll save for another time; only that this ecosystem exists and costs time. A noticeable share of my work — and of my token consumption — goes not into the result but into what makes the result possible in the first place. In my experience that’s roughly 30 percent — a tangible part of the effort that no one sees, because it stays invisible in the finished product. (*)
The second is reviewing. And that’s the part you can’t optimize away.
Where does AI stand compared to human developers?
When someone asks me what level I can expect from AI, I say: roughly that of a junior developer. The comparison breaks down in one place — and that’s exactly the telling part. A human keeps developing every day, task by task. The machine doesn’t. And because I deliberately refuse to give this thing a human “she” or “he,” I treat it as an “it” — in German I even write a deliberately capitalized pronoun, a small, intentional break with the rules of the language, to mark how alien the machine is. It stays exactly at its level until a new version of the model appears — then the level jumps abruptly upward and stands still again. Development in leaps, not along a curve.
One important addition, so this doesn’t sound like junior-bashing: when it comes to the sheer breadth of retrievable knowledge — the knowledge base — no human can keep up with the AI. It commands dozens of programming languages, frameworks, techniques. It has read billions of lines of code, and the important principles are imprinted in the neural network. In that breadth, no human beats it. But what AI can’t do: work intuitively. There the senior developer is clearly ahead, and the junior developer is learning it right now.
A junior on steroids
I give a junior developer a task, get a result, and have to look closely: is this what I wanted — and built the way it should be? That costs time. But a junior is a human; our speeds are similar. The machine is tireless: in next to no time, mountains of code appear. “How on earth am I supposed to review this?”
This junior on steroids takes a task, falls into tunnel vision, shuts itself away and comes back hours later with “Done, works perfectly.” Then you look inside: yes, it’s multilingual — but the texts are hard-wired, right in the middle of the code. “You could have asked.” It doesn’t care. New session, from scratch, as if we’d never spoken.
That’s not a problem as long as you know it. It becomes a problem when you trust the speed and skip the looking. We humans must not confuse speed and volume of output with quality. Volume is no proof of quality.
Why small errors grow large
I explain it to myself with an image — and I explicitly call it an image, not a proof. The AI doesn’t build the whole at once, but detail by detail, each time the most probable one. Picture a tree: from the function to the blocks to the individual instructions. At each level the result is maybe ninety percent right. But ninety percent times ninety percent times ninety percent — that shrinks quickly toward the bottom. At a hundred thousand lines, even ten percent is a lot of code that isn’t right. And a single wrong sign can tip an entire calculation.
Now the caveat the image often hides: reality doesn’t calculate that cleanly. Errors aren’t independent of one another, tests catch a lot, iteration corrects. As mathematics the image is wrong. But as intuition it captures something true: quality that feels good is not the same as quality that is right. And the more code is produced, the more has to be reviewed — not less.
The real shift
That makes clear what really happens. The work doesn’t disappear, it migrates: from writing to steering and reviewing. That’s more demanding than it sounds — and it’s a different activity from the one most of us were trained for.
How to organize that reviewing I’ll save for a text of its own — that’s the real craft. (*) But the direction is unmistakable, and it’s a warning to everyone who plans AI in as a pure accelerator: whoever budgets only for the fast half and not the second one pays the second anyway — just unplanned, and usually more expensive.
Productivity through AI is real. But it’s only half the truth. The other half is the price you should plan for deliberately, instead of being surprised by it.
(*) A separate article on this is in the works.
Disclaimer: This article reflects my personal opinion, based partly on observation and partly on research. I welcome corrections where I have demonstrably got something wrong. I have no time for scaremongering; my aim is an analytical, philosophical look at my own working environment.