AI's real payoff is not productivity, it is the comfort of the person using it

In February 2026, two economists at the Bank of Korea published a representative survey of 5,512 workers. They went looking for productivity. They found something else, and they said so plainly in their conclusion. What I want to argue here is that this something else is not a disappointing side result. It is the main one, and it repairs something very old.

What industrialization did to office work

We did not only industrialize factories. We industrialized the head.

Start there or none of the rest holds. In 2006, the sociologist Marie-Anne Dujarier published L'ideal au travail, drawn from a doctoral thesis supervised by Vincent de Gaulejac. She studied two settings with nothing in common: a public geriatric facility and a private restaurant chain. What she found in both was the same movement. Service organizations were applying to relational work the logic of throughput, yield and mass production inherited from industry.^[1]

The trouble is that this kind of work resists. A service happens inside an uncertain interaction, and its relational core opposes standardization by nature. Nobody standardizes the way a care worker reassures an elderly patient. Organizations try anyway. And to try, they add criteria, procedures, indicators, reporting lines, traceability.

Dujarier then establishes two things that are rarely quoted together. First, the Taylorist split reappears inside white-collar work, between those who execute and those who, at headquarters, design other people's jobs. Second, and this is the heart of it, in services working means thinking, which is precisely what the Taylorist apparatus is built to remove from the shop floor. Industrial Taylorism set out to take thought out of the worker's hands. Applied to the service sector, it tries to take thought out of work that consists of nothing else.

What follows is not overload in the sense of volume. It is a conflict. The norm becomes ideal at the same moment the ideal becomes the norm. Prescriptions contradict one another, sometimes impossible to satisfy at once. Permanent self-monitoring generates a rising mental load, and the work becomes hard to bear.^[1] Nine years later, in Le management desincarne, she names the people who manufacture that layer: the planners, those who design the organization at a distance from the real work, roughly 40% of French middle and senior management.^[2]

Industrialization did not add work. It added proof of work.

That formulation carries everything that follows. What the industrialization of white-collar work deposited on these jobs is not mainly extra tasks. It is a layer of justification: the meeting minutes, the status report, the compliance filing, the summary note, the tracking sheet, the standardized document certifying that the work happened and happened as prescribed. An entire layer whose function is not to deliver the service but to make it legible and controllable to people who do not perform it.

The cost of that layer is measured, and the measurement is official. The Sumer survey run by the French labour statistics office, covering 48,000 employees interviewed by 2,400 occupational physicians, finds that 66% of workers must frequently interrupt one task to handle another, unplanned one. Among managers and professionals it reaches 73%.^[3] Work intensity has risen over twenty years. We are describing a workforce in a state of chronic interruption, spending a meaningful share of its day documenting the work rather than doing it.

The layer the machine eats first

A language model is not good everywhere. It is excellent exactly where we hurt most.

Consider what a language model does best. It produces standardized writing out of raw material. It reformats, condenses, brings into compliance, turns a spreadsheet into a memo and a memo into a spreadsheet. This is not one skill among others: it is its native skill, the one it was trained on. And it is also, word for word, a description of the proof layer that industrialization laid over service work.

The coincidence is too convenient to accept without a test, and the Korean survey provides one. The authors asked each respondent, task by task, where the time was saved. The result is sharp: savings are highly concentrated. The single most-improved task accounts for roughly 70% of a user's total time reduction.^[4] Among health and social professionals that share reaches 80%. Among administrative and accounting clerks, 77%. Among culture and arts professionals, 84%.

The authors draw a conclusion I will take as it stands: generative AI works as a targeted tool aimed at specific bottlenecks, not as a general accelerator of work. It does not make people broadly faster. It removes one blockage, almost always the same one, for each person.

Which leaves the question of which one. The survey does not name the tasks one by one, and I will not make it say what it does not say. But in the rooms where I have been training people for three years, the answer barely varies. What people automate first, on their own, without being prompted to, is the meeting summary, the briefing note, the rewrite of a document for a different reader, the filling-in of a mandated template. The proof layer. Not the craft.

The number the economy cannot read

Zero point zero zero eight. That is the number nobody knows what to do with.

To the headline results. Among 5,512 working people aged 15 to 64 surveyed in May and June 2025, 63,5% report having used generative AI, and 51,8% use it for work. Korean adoption runs at roughly twice the American rate, and the technology spread faster there than the internet did in its time.^[5] Among users, working time falls by 3,8%, about an hour and a half in a forty-hour week.

Then comes the measurement that caused the noise. The authors had the rare idea of asking, in the same questionnaire, both the time saved and the change in output. Then they correlated the two.

The saved time does not go into output. The coefficient is indistinguishable from zero, controls included. Had every saved hour been reinvested in production, the productivity gain would have been about 1,0%. It was not.^[6]

The standard reading of this result is the Solow paradox: the technology is everywhere except in the statistics, so be patient. The Bank of Korea itself takes a cautious line in its June policy note, suggesting AI has entered an efficiency stage without reaching a productivity stage.^[6] That is reasonable. It is not enough, because the academic paper goes looking for where the time actually went. And it finds it.

First channel, direct: the share of working time spent on what the authors call, without euphemism, on-the-job leisure rises by 1,3 point among users.^[4] This is not time theft. It is a drop in intensity. The same work, with less effort and less strain.

Second channel, and this is the one that matters here. The authors regressed the change in job satisfaction on time saved and on change in output. Time saved carries a coefficient of 0,039. Output carries 0,007. The link between time saved and satisfaction holds even when controlling for output. Satisfaction tracks the time you get back, not what you do more of with it.

Their conclusion sits in the abstract, and it deserves slow reading by anyone running an AI rollout.

These findings suggest that standard productivity measures may understate AI's impact by missing non-pecuniary welfare channels.

Why she does not produce more

The economists measured the silence. They could not ask why.

A survey of 5,512 respondents gives you the scale of a phenomenon. It does not give you its reasons. That is where fieldwork takes over, and I have one case that explains the 0.008 on its own.

A finance director I met in a training session says she has recovered 75% of her working time by driving her spreadsheets through a conversational assistant. Seventy-five per cent. I ask the obvious question: what do you do with it?

Her answer was not what I expected. She does not produce more, and she knows exactly why she does not. Because the surplus would be absorbed immediately by her management, and she would see neither the fruit nor the gain of it. So she listens to podcasts. She reads articles. She spends time on social media. Meanwhile she keeps delivering the same work, praised for its quality, produced four times faster. Her employer is satisfied. Her stress has dropped sharply. Asked whether she uses AI, she answers without hedging that you should assume everything she produces has passed through it.

This woman is not cheating. She is running a calculation, and the calculation is correct. She assessed the individual return on extra effort inside her organization, found zero, and allocated accordingly. I have described elsewhere why that return is nil in most structures, and why it is not a management flaw but a failure of alignment between the person supplying the effort and the person collecting the fruit (Your organization captures nothing from AI. It was built against it.). What interests me here is the other end of the chain. Not what the organization loses. What the person recovers.

Hers is not an isolated case, and I have started measuring it. When managers are out of the room, I ask two questions. First: during your working day, do you ever handle personal matters? Second, for AI users: has that increased since you started? The answer to the second is overwhelmingly yes. And when I ask what those matters are, the answer does not point to a second job or an exit plan. It points to leisure, personal projects, community and volunteer commitments.

I do not yet have a publishable statistic, and I am not going to invent one for the sake of the argument. The diagnostic tool I am rolling out has only just begun measuring this systematically. But the convergence between what I see in French training rooms and what the Bank of Korea measures at national scale is too clean to ignore. The economists' on-the-job leisure and my finance director's podcasts are the same phenomenon.

Two artificial intelligences, two directions

Fair objection. Most of what we have proven about AI at work is that it harms. Let us take it head on.

There is a substantial and serious body of research documenting the exact opposite of what I have just written. Ignoring it would make this essay collapse under a single sentence. So let us look at it, because looking at it produces the distinction the debate is missing.

The one that comes down

When the machine arrives from above, it extends Taylorism. It does not repair it.

The literature on algorithmic management converges: systems that direct, score, pace and monitor work degrade health. It documents heightened time pressure, loss of autonomy, erosion of meaning, damaged cooperation, fear of replacement.^[7] Recent work links perceived algorithmic management to workers' physical and mental health, with frustrated autonomy and replacement anxiety as the mediating mechanisms.^[8]

None of this contradicts Dujarier. It extends her. Algorithmic management is the proof layer pushed to its technical limit: total traceability, real-time prescription, control without a visible controller. It is the automated planner. The organization installs the tool on the worker, and the worker adapts or disappears. That is the classic vector of automation, unchanged for two centuries.

The one that comes up

Generative AI at work, in its majority form, was installed by nobody. It arrived from below.

The other movement is structurally different, and I have devoted a full essay to establishing what it is (BYOAI, when the worker brings their own AI). The worker brings the tool, uses it quietly at first, and the organization runs to catch up. This is neither a technical workaround nor a transfer of hardware: it is a reappropriation, by the employee, of their own cognitive means of production.

Hold the two together and the contradiction dissolves. The studies that find degradation are observing systems imposed on the worker. The studies that find improved wellbeing, the Korean survey among them, are observing use that the worker mostly brought in. These are not the same objects, not the same vectors, and there is no reason for them to produce the same effects.

I put this asymmetry forward as a thesis, not a demonstrated law. Two separate literatures pointing in opposite directions form a converging body of evidence, not a controlled experiment. Nobody, as far as I know, has compared the two vectors on a single population under one protocol. Until that work is done, honesty requires saying that the direction of the relationship is highly likely, and that it is not established.

The restitution is not automatic

The machine gives nothing back on its own. It gives back what you feed it.

The Korean paper sets the guardrail itself, which is what makes it a good paper. Its authors do not stop at measuring time: they ask each respondent, for each task, whether it provides personal fulfilment or a sense of achievement. Then they track how AI shifts the composition of work between fulfilling and unfulfilling tasks.

The reasoning is explicit in their framework. If AI automates drudgery, the worker shifts time toward what carries meaning, and welfare rises even at constant output. But if AI automates the fulfilling part, the efficiency gains cancel out, because the worker ends up spending proportionally more time on what wears them down.

That is precisely what their results show. Depending on the occupation, the share of fulfilling tasks rises or falls. Some gain, others lose. The authors conclude that relatively few workers report unambiguous satisfaction when judged jointly on the time dimension and the task dimension, and that AI's welfare impact remains highly heterogeneous at current adoption levels.^[4]

This does not weaken the thesis. It conditions it, and it names the condition. The restitution depends entirely on what the user chooses to delegate. I set that rule down elsewhere from fieldwork, before I had this study in front of me: AI on the tasks that grind you down, never on the ones you love (Tipping into AI). What I was presenting as a trainer's intuition has just been validated at national scale by a representative survey of 5,512 people.

Which moves the question of training. Learning to operate AI, in the sense of the technical gesture, takes a few hours and already happens without us. Learning to decide what you will not hand over is a skill of another order, it is not acquired through use, and it decides whether the time you save repairs you or hollows you out.

What no dashboard counts

Productivity does not measure work. It measures what you hand to your employer.

Now the instrument defect has to be named, because the entire public debate on AI rests on it. Productivity is output over time. The numerator and the denominator each carry a problem, and we only ever argue about the first.

The numerator first. By construction, it counts only what travels upward: what the worker hands to whoever employs them. The time they keep, the tension that subsides, the drudgery that disappears, the interruption that does not happen, none of it appears in the indicator. Not overlooked: invisible, in the strict sense, because the indicator has no slot for it.

The denominator next, and this one is more vertiginous. The time being measured is human time, and the ratio only means something if you assume a roughly stable proportionality between the hours a person puts in and what comes out. That approximation held while the tool merely assisted the gesture. It breaks the moment the tool substitutes machine time for human time, and a few minutes on one side produce what took dozens on the other. An indicator whose denominator no longer counts the resource actually manufacturing the result is not measuring badly: it is measuring something else.

So when we announce that AI is producing no productivity gains, we state an accounting truth and miss the event. The gains exist. They are measured, dated, quantified. They simply did not go where the instrument is pointed.

And we should resist the easy version of this story, the one where technology comes to repair what management broke. It is not AI that gives anything back. No machine restores anything by itself, and an AI imposed from above gives back nothing at all, as we have just seen. What is happening is more interesting and less comforting: the machine opens a breach in the proof layer, and it is the worker who takes back. Taking back is an act, it belongs to whoever performs it, and it happens precisely because they brought the tool in themselves.

Still, the loop closes, and it closes on an irony Dujarier could not have anticipated in 2006. The industrialization of white-collar work set out to take thinking out of the office, replacing it with procedure and proof. Twenty years on, a machine swallows the procedure and the proof. What it leaves on the desk is the work that was left over once everything else had been standardized. In service occupations, that was the one thing that never could be.

My finance director did not gain seventy-five per cent of her time. She recovered seventy-five per cent of a layer that should never have weighed on her. And if her employer sees no gain in it, perhaps that is because he never measured what the layer was costing him.

Jean-Jérôme DANTON