The Friction Dividend

Manual data work felt like a chore, and getting rid of it felt like relief. In reality, friction spent years paying analysts a hidden dividend — their most important skill: doubt. And doubt is disappearing along with it.

For the past few months I’ve been watching analysts, product managers, and myself: you open Claude — or your assistant of choice — ask a question, and a minute later you have a chart that explains everything. Then I noticed a strange thing.

Those of us who have worked with data long and seriously — the people who actually understand where the numbers come from — became more careful with AI. They look at the result and ask: is this actually calculated correctly? did anything get duplicated when the tables were joined? is the event even tracked the way we think it is?

Those with less data experience ask almost nothing. There’s an answer, the chart looks clear, moving on.

This is an observation, not statistics. But it’s too consistent to be a coincidence: AI has changed the very nature of analytical work. And to explain how, I first need to describe what a strong analyst is made of.

Two Axes

There’s a type of specialist that everyone who has ever built a product team hunts for. A person at the intersection of two axes — the proverbial “ideal analyst.”

The first axis is product thinking. Understanding user behavior, seeing what a metric actually means, sensing which question is worth asking — and which one merely looks important, though its answer won’t change a single decision.

The second axis is technical craft. Getting the data, cleaning it, computing it in a way that makes the numbers trustworthy.

The intersection is rare because both axes take years. And the technical axis was the real barrier to entry. This is exactly where AI changed everything — but not in the way it seems at first glance.

The Hidden Function of Friction

The obvious logic goes like this: AI lowers the technical barrier. A product manager with strong product thinking now relies on AI, which writes the SQL/R/Python under the hood and builds the dashboard — no analyst required. The technical axis became accessible — therefore, there should be more people at the intersection. The ideal analyst finally stops being a rarity.

I believe this is an illusion. The technical axis performed two functions, but we only ever noticed one.

The first function, the obvious one: being able to do the work technically. Write the query, build the pipeline, get the numbers.

The second function, the hidden one: the very process of mastering the technical side was a teacher. It taught two things at once. When an analyst spends years writing queries by hand, they aren’t just completing tasks. They run into data that doesn’t exist. They notice an event tracked the wrong way. They see numbers that don’t match and have to figure out why. Through constant friction with raw data, they develop a feel for the product — not from a textbook, not from a course, but from contact. That’s the first thing.

The second is the habit of never trusting a number on first sight. What matters is why friction taught it: it made errors impossible to miss. The query failed. The numbers didn’t add up. The dashboard showed nonsense — and there was no way not to see it. You couldn’t walk past an error — it stopped you itself — and after a hundred such stops, doubt became a reflex. The data rapped you on the knuckles, and every rap taught you something.

AI removes the friction. But friction was the school — the school of hard knocks, quite literally.

And note what exactly disappeared: not just the effort — the effort is no loss. What disappeared is the forced visibility of errors. An AI error is quiet: the query ran, the chart rendered, and the fact that rows got duplicated somewhere or a filter never got applied — invisible.

A person who arrived at the intersection of product thinking and technical craft through AI can technically do more — but they arrived without the journey that made the intersection unique and valuable. They’re standing at the point of intersection without the thing that point was supposed to give them. We’re used to counting such people by coordinates: can they do both? But the value of the ideal analyst was never in the coordinates. It was in how they got there.

The Trust Inversion

An experienced product manager next to AI looks less confident. Not because they got worse — if anything, the opposite: their doubt was always there. It’s just that the hard analytical work used to be done by someone else, and now they do it themselves — and they see how many places things could have gone wrong. They know how many ways data can lie. AI gave them speed, but the judgment didn’t go anywhere — it switches on as doubt. “Is this actually right?” — because years of friction taught them: usually not, not on the first try.

A product manager with less data experience became more confident. Not because they’re right, but because they don’t see the layers of complexity. AI produced a clean answer — no questions.

The more experienced, the more careful; the less experienced, the more confident. Confidence has become inversely proportional to competence. The person with real judgment hesitates; the person without it doesn’t. And in the room where decisions get made, confidence reads as competence. That’s how we’re wired — we trust the one who speaks without pauses, without doubts.

So AI doesn’t merely allow people to appear ideal without being ideal. It makes them more convincing — because it removes the single external signal that used to give a novice away: hesitation. An inexperienced analyst used to stumble over technical complexity, and everyone could see it.

You’ll say — novice overconfidence is nothing new; it was always like this. True. But it used to be temporary: reality punished it fast — the query failed, the numbers didn’t add up — and the novice learned. What’s new is not that overconfidence exists. What’s new is that the correction is deferred: an error that a failing query used to catch in five minutes will now be caught by the market a quarter later — when the decision built on wrong numbers is already live.

What’s Happening

You could say: fine, AI is still imperfect; in time it will handle both data quality and interpretation better. It will. The only question is how.

The first process. AI closes the technical axis entirely. It stops being an axis — it becomes commodity infrastructure. It stops being an advantage. Only one thing remains rare: product judgment. When everyone has perfect data, the winner is whoever asks the right question. Companies understand this, by the way — senior roles have long been hired for product thinking. But how was it verified? Through track record: where the person grew, what problems they went through, what experience they gained. In other words, even when hiring for judgment, companies relied on the candidate having gone through the old school of friction. Now look at who gets laid off first: juniors. Why keep a junior when AI does their work? A decision that’s rational today and expensive tomorrow: a senior with judgment isn’t hired out of thin air — they’re grown. A company cutting junior positions today will, five years from now, be searching for seniors that nobody grew.

The second. AI is getting good at interpreting the data it’s given. But it doesn’t know what the data is missing — it works with what’s there. And the rarest judgment of an analyst is precisely about that: noticing that the needed metric isn’t computed at all; that the product tracks the wrong thing; that the question — or the query — everyone is asking is the wrong one. And here’s the paradox: the better AI interprets, the more valuable this ability becomes — and the fewer places remain where it could emerge. The price is rising, and the school that nurtured it is already closed.

The third — and it worries me most. What if AI becomes so convincing that doubt disappears even in the experienced? The experienced hesitate because they know data can lie when it’s prepared wrong or queried wrong. But if AI delivers accurate conclusions for years — even they will stop checking. Why bother, if the last hundred times everything was right? That’s exactly what we do with a calculator, and it’s reasonable. The difference is in how the system fails. A calculator either works or it doesn’t.

AI fails plausibly — its error looks exactly like a correct answer, with the same confident tone and the same beautiful chart.

And at the moment it does fail, there will be no one left with the instinct to notice. We will collectively unlearn doubting at precisely the moment doubt becomes most expensive.

What to Do About It

All three processes lead to the same point. However AI develops, the value shifts from answers to questions — and to the ability not to believe an answer too soon.

Doubt has always been part of an analyst’s work — no revelation there. What changed is something else: it stopped being free. Doubt used to come on its own: you learned to doubt without noticing it — the data caught you making mistakes every day. It was a dividend that friction paid out daily, and nobody thought of it as income until the payments stopped. Now there’s nothing to catch you, and doubt has to be held deliberately. It’s no longer a reflex. It’s a discipline.

Practically, this means the following. If your team has that rare person at the intersection — hold on to them: the school isn’t producing new ones. If you’re hiring — remember: candidates who went through friction will only get scarcer, and confidence guarantees nothing. A candidate’s doubt may turn out to be more valuable. And if you run a team that works with data through AI every day, ask yourself one question:

When was the last time someone on your team said, “I’m not sure these numbers — or this AI conclusion — are right”? And what did you do about it?

If the answer is “long ago” or “nothing,” it doesn’t mean there are no errors. It means nobody is looking for them. Friction used to sustain the doubt. Now it’s on you.