Coding was never the real work
Late on a Tuesday, I was reviewing what the AI had built that afternoon. The code compiled. The tests passed. The feature worked. And buried in the middle of it, two components were quietly drifting toward each other in a way that would have been genuinely painful to untangle in three months. Not a bug. A posture. The kind of thing that doesn't show up in any metric and only becomes visible to someone who has seen it happen before.
I caught it. I asked the AI why it was doing that — and then I told it to write a note, in the code, so it would not do it again.
That is the job now. Not the building. The knowing.
The pressure to skip it
There is a particular pressure right now to treat this kind of review as overhead. The executive logic is clean: if the output looks right, the tests pass, and the app works, why put a human in the loop at all? AI is faster. The backlog isn't getting shorter. Pick your battles.
The problem with that logic is it mistakes what the review is actually for.
The human in the loop isn't checking the output. The human in the loop is building the judgment that knows whether the output is any good — and transmitting it, in real time, to everyone watching.
Demanding because it matters
I learned the shape of this from a person, not a framework.
Asko Känsälä was deputy CEO of Elisa when I was there. He has since written a book about leadership as mastery — which is, not coincidentally, exactly what he practiced. The thing I took from him and have never been able to shake: kindness and accountability are not in conflict. They are the same gesture.
A carpenter shapes the wood carefully because the carpenter cares about the piece being good. Not because the carpenter doesn't trust the tools. The tools are fine. The tools don't know when the piece is wrong.
Demanding something of someone is an act of care. The only people you demand nothing from are the ones you've already written off.
The mechanism nobody names
A paper published in Harvard Business Review in June 2026 put the clearest language I've seen to something I've believed for years. Chengwei Liu and Balázs Kovács were studying capability pipelines in professional knowledge work, and they made one observation that stopped me:
"The sign-off doubles as teaching. Residents read scans under a signing radiologist. The accountability mechanism is also the apprenticeship mechanism. Cut one, you cut the other."
— Liu & Kovács, Harvard Business Review, June 2026
This is what gets lost in the efficiency argument. The review is not a check. It is the lesson.
What happens in an AI-native workflow
In an AI-first team, this dynamic accelerates in both directions. AI compresses output. More ships. More gets reviewed. The review workload goes up at the exact moment the model says there's no need for it.
I feel the temptation. The code looks right. The tests pass. I've reviewed six things today already. Can I let this one go?
No. Because the next person who builds on top of this will inherit the posture, not just the code. And they won't know what to do with the weird accumulation three months from now — because nobody was in the room when the direction quietly went wrong.
This is not hypothetical. I catch this kind of drift constantly. The AI is remarkably capable and remarkably confident. It will accumulate small wrong assumptions across a hundred files and present you with something that looks coherent until you understand the architecture it quietly dismantled. It is a constant battle against entropy, and the only person who can fight it is someone who knows what it looked like before.
The lesson, in person
My son Tommi finished school last week. He is eighteen, loves computers, knows some Python and JavaScript, can build a simple script — but has never built a real application, never opened a pull request, never written a test. He asked me if there was something he could work on.
There was. We have an internal tool at Black Belts, the consulting firm I co-founded, that handles hour collection from our consultants. I handed Tommi a Claude subscription and told him I wanted him to automate "the hell out of the invoicing process." Those were my exact words.
The catch — deliberate — was that the tool had been built by one of our most senior engineers, who would review everything Tommi produced. Not to quality-gate it. To show him what a senior engineer actually thinks about: the workflows, the pipelines, the architectural choices that never appear in a tutorial, the reasons behind decisions that look arbitrary until you understand the system underneath.
Tommi went at it like an inexperienced boxer with a lot of energy — punching left and right, endlessly, without ever getting tired. He — or rather, the AI he was directing — produced a huge amount of code, a huge number of changes. And then, one pull request.
He got some feedback. He learned.
Not from a course. Not from documentation. From the gap between what he submitted and what a senior engineer knew it should be.
Tommi is learning to be a software engineer the way his generation will — AI-native, from day one. The AI generates the code. He directs it, reviews it, argues with it, owns the output. That is fine. That is the future.
But here is the thing we have confused for thirty years, and AI is finally making undeniable: coding was never the real work. Anyone who spent a career "just coding" was doing about ten percent of the job. Software engineering is solving problems in ways that are not obvious. It is knowing, from experience, what will go wrong before anyone has thought to specify that it can. It is understanding what happens to your architecture when a hundred thousand users show up simultaneously. It is recognizing the library that will be deprecated in eighteen months and building accordingly.
We confused the craft with the output. Now the output is cheap. Which means the craft is more valuable than it has ever been — and the only way to transmit it is still a senior engineer in a room with someone younger, asking the questions that reveal the gap.
Hinton predicted AI would replace radiologists by 2025. He was wrong in the way only very smart people are wrong: he described the output of the job and mistook it for the job itself. We did exactly the same thing with coding.
Tommi isn't learning to code. He is learning to engineer. The review is the difference between those two things. Pull the review, you pull the whole pipeline.
The debt nobody is counting
Liu and Kovács describe two invisible liabilities compounding on every tech company's balance sheet right now: capability debt and judgment debt. Both are invisible on the income statement. Both compound. The bill arrives when the next genuinely complex problem lands at a firm with neither builders nor judges.
The A-players who look like they don't need managing were developed somewhere. Someone reviewed their work. Someone asked the uncomfortable question. Someone was in the room.
It is individually rational, right now, for every company to skip that work. Hire experienced people. Let AI do the rest. Don't invest in developing anyone — that takes years and the trained person can leave. So no one trains. The industry extracts from a pipeline it is simultaneously running dry.
I should be transparent about my own position here. Black Belts, the firm I co-founded, places senior engineers in exactly these roles — embedded in companies that need the judgment they haven't had time to develop internally. We exist because the gap is real and the demand is immediate. There is nothing wrong with bringing in senior expertise from outside. It is a legitimate answer to a real problem.
But it is not the only answer, and it cannot be the only one. The consultant reviews the work, transmits judgment, catches the drift — and then leaves. The engineer you develop internally stays. Mentors. Becomes the person who reviews the next Tommi. External and internal are not competing strategies. They are the same discipline at different timescales. The industry needs both. Right now, it is only investing in one.
You cannot memo your way out of this. The pipeline doesn't care about your org structure.
The review you're tempted to skip — because the AI built it and it looks right — is not a check. It is a teaching moment you are quietly handing to no one.