When Intelligence Got Cheap: What Is Expensive Now

Light you no longer have to work for

In the early nineteenth century, an hour of reading by candlelight cost a European roughly a few hours of labour. Candles were expensive, they smoked, people rationed them. The economist William Nordhaus once did a curious piece of work: he recalculated the price of light in working time — and found that today the same hour of illumination costs us a fraction of a second of work. Not a thousand times cheaper. Tens of thousands.

That, though, is not the interesting part. The interesting part is what happened next.

When light got cheap, people did not simply read more in the evenings. Night shifts appeared, and round-the-clock factories. Hospitals that work at three in the morning. Airports. Nightlife as a culture of its own, with its own music, its own fashion and its own professions. Cities you can see from space. The cheapened resource did not just make old activities cheaper — it created activities that had never existed and that nobody sitting with a candle could have predicted.

Right now, in front of us, the thing all of human history treated as the most expensive and the scarcest is getting cheap: the ability to work something out, put it into words, analyse it, lay it out. Not "intelligence" in the philosophical sense — that is a complicated business — but its working, applied part. The part people were paid a salary for. The part smart people were kept on staff for.

And the main question here is not "who gets replaced". Everybody asks that one, and honestly, it is not the most substantive. The main question is a different one: what becomes expensive when this gets cheap?

The answer I keep arriving at is fairly uncomfortable. Especially for those of us who spent our whole lives being straight-A students.

I. Scarcity has moved: what is scarce now is intent

Picture two lawyers in the same office. Same years of experience, same diplomas, same subscription to the same model, paid for by the firm on the same day.

The first opens a chat window and types: "Draft a statement of claim for recovery of debt." He gets a statement of claim. A good one, as it happens. He edits it, files it, closes the tab.

The second sits and thinks for three days. Then he writes: "We have forty small-business clients. Every year they have receivables stuck somewhere between 300,000 and two million tenge, and they do nothing about it, because a lawyer costs more than the dispute. Let's work out at what unit cost this service becomes worth doing for them and for us, and what it would take to get there."

Both of them "know how to use AI". Both write excellent prompts. The difference between them is not in the prompting.

I think we are misidentifying the main divide. Everyone is discussing the gulf between those who have mastered the new tools and those who have not. But mastering them is a week's work. In two years it will be as unremarkable as knowing how to use a search engine: nobody brags about it.

The real gulf, it seems to me, runs somewhere else — between those who know what they want to achieve and those who are waiting to be told.

Artificial intelligence is an amplifier. It multiplies a person's capacity to act. But an amplifier with no signal at the input puts out nothing but noise. Access to a strong model does not by itself create a goal, or initiative, or a willingness to own a decision. It makes fast the person who already knows where he is going, and it leaves exactly where he was the person waiting for instructions. That difference between them always existed — it was just compensated for. Someone had to execute: someone had to write the text, do the spreadsheets, dig up the case law. Now that compensation is shrinking.

And here comes the unpleasant part. Our entire education system — school, university, corporate — is tuned to produce executors. A good pupil is one who correctly solved the problem he was set. A good employee is one who carried out the assignment well. In many organisations, initiative is still treated as a form of insubordination. For thirty years we have been selecting people for their ability to do well what they were told to do — precisely the trait that is getting cheap fastest.

Hence my first conclusion, and it is not only about business: entrepreneurial thinking is becoming useful even to someone who will never start a business of their own.

I do not mean registering a company or raising venture rounds. I mean a distinct competence: to see a problem others missed or considered inevitable; to propose a solution; to carry it through to a result somebody is willing to pay for — in money, in attention or in trust.

A doctor at a district clinic who notices that forty per cent of his time goes on discharge paperwork and builds himself a template is not an entrepreneur. But he acts like one. A teacher who works out that parents do not read her messages in the group chat and redesigns the format — likewise. A government lawyer who notices that the same drafting error recurs two hundred times a year and writes a one-page guide — likewise. None of them is creating a business. All of them are creating value out of nothing, in a place where before there was only a complaint about circumstances.

That used to be a pleasant addition to professionalism. Now, it seems, it is becoming professionalism itself.

II. The work is moving to the edges

Almost any project — from a lawsuit to building a house, from an article to a product launch — breaks into three parts.

First somebody decides what needs doing at all, and why. Then somebody does it. Then somebody looks at the result and decides whether it is any good.

Setting the task. Doing the work. Accepting the result.

Agents and models amplify the middle above all. Execution — gathering material, the first draft, the calculation, the assembly, the formatting — is what compresses fastest. And here is the arithmetic: if the middle gets cheap while the ends stay human, then all the value drains to the edges. To knowing how to choose what to do. And to knowing whether it has actually been done.

I have a favourite thought experiment which, sadly, looks less and less like a thought experiment.

An agent produces a flawless motion in four minutes. The statutory citations are precise, the case law is fresh, the structure is exemplary, not a single typo. One problem: in this case, no motion should have been filed at all. What was needed was to sit down that same day with opposing counsel and settle, because the client needed money this quarter, not a victory next year.

The system executed a badly set task perfectly. And it did it fast. Which is to say: it got to the wrong place sooner.

Speed without direction is just a longer distance from your goal.

Something very practical follows, and it is currently being misunderstood en masse. Managing agents is not a contest in how many processes you can run at once. It is not the sport of "I have twelve tabs open and you have four." Working with agents means working with goals, constraints, quality criteria and feedback.

Which is to say it is management, actually. Except that now everyone has to do it — including a huge number of people who have never managed anyone and never intended to. The ability to explain what counts as a good result has always been rare: anyone who has ever hired a contractor for a renovation knows this. Now it is becoming a mass requirement.

The second part — accepting the result — is more interesting still, and there is a landmine buried there that I will come back to.

To accept a piece of work, you have to know what good work looks like. Not "like it / don't like it", but on the merits: where the weak point is, what the author passed over in silence, which citation looks persuasive but will not survive checking. That knowledge does not come from reading. It comes from having done the thing yourself — from exactly those years when you sat there doing it by hand, badly, then better, then well enough.

Those are precisely the years we are about to cut. More on that below.

III. Between a strong model and a strong business lies an enormous market

A story economists have been telling each other for a hundred years.

At the end of the nineteenth century, electricity arrived in the factories. You would logically expect an explosion of productivity. There was not one — for about thirty years. Factory owners did the sensible thing from their point of view: they threw out the steam engine, put a big electric motor in its place, and left everything else as it was — the line shafts under the ceiling, the belts, the machines arranged around the central drive. It worked. Almost no gain.

The gain came when the next generation of engineers realised that the motor could go into each machine separately. Which means the shop floor no longer has to be built around a shaft. Which means the equipment can be laid out according to the logic of the process. Which means the assembly line, a different layout, different workshops, different logistics, different trades and a different system of training workers.

The technology was ready immediately. Organisational readiness lagged by a generation.

The economists Erik Brynjolfsson, Daniel Rock and Chad Syverson described this pattern as the productivity J-curve. First a company invests in the invisible: retraining people, rebuilding processes, new data, new procedures, new ways of checking. The statistics barely see those investments — they are "intangible". The costs, on the other hand, are perfectly visible. So for the first few years it looks as though the technology is not paying off. And then, if the organisation really has rebuilt itself, the curve turns upward.

Now let us translate that into today's language.

A law firm buys subscriptions for everyone. Six months pass. The partner looks at revenue and sees nothing. The conversation begins about how "this whole AI thing is overhyped."

But let us look at where the time in that firm actually goes. The client spends two weeks sending documents in pieces and loses half of them. Then a junior lawyer spends a day hunting through email for the latest version of the contract and finds three files with identical names. Then the draft sits with the partner for a week, because the partner is in hearings. Then three rounds of edits, two of them about style. Then accounting cannot issue the invoice, because nobody knows which matter it belongs to.

In that chain, actually writing the text took maybe fifteen per cent of the time. The firm radically sped up fifteen per cent and is sincerely baffled that the deal cycle has not changed.

Hence what I consider the most interesting entrepreneurial hypothesis of the moment: the biggest opportunity is probably not in building even smarter AI, but in rebuilding work around the AI that already exists.

And that is excellent news for a great many people who cannot train neural networks and never will. Between "there is a strong model" and "there is a strong business" lies a gigantic field that calls for completely different competences: understanding how work actually happens in a specific industry; the ability to talk to the people who do that work; the patience to take a process apart and put it back together.

A practical conclusion for anyone who sells something. The offer "we generate documents ten times faster" is a bad offer. It addresses fifteen per cent of the problem, and the client feels that even if he cannot articulate it. The offer "a contract from request to signature in two days instead of three weeks, and here is how we guarantee it" is an entirely different conversation. The first sells typing speed. The second sells the removal of the whole bottleneck.

The difference between them is the difference between a vendor and a partner. And, as a rule, an order-of-magnitude difference in price.

IV. Automation does not necessarily shrink a market. Sometimes it opens up demand that was never there

There is a very durable intuition: if a machine can do a person's work, there will be less work for people. It seems so obvious that it is rarely tested.

History refutes it regularly — not always, but often enough to take seriously.

The classic case is the ATM. They spread across the United States in the late 1980s and the 1990s. A teller in a branch did exactly what an ATM does: handed out cash. Logic says tellers should have disappeared. The economist James Bessen showed that something else happened: the number of tellers per branch fell, the cost of running a branch dropped, banks started opening branches far more aggressively — and for a while the total number of tellers actually grew. What changed was the content of the job: from dispensing money to sales and customer service.

Another case is photography. Film cost money, developing cost money, there were thirty-six frames in a roll, and you thought before you pressed the shutter. Today several billion photographs are taken every day around the world. The profession of photographer has not disappeared — what disappeared was the part tied to the technical difficulty of taking a picture, and what grew was the part tied to direction, taste and knowing how to work with people.

The logic is simple: when a service gets cheaper, people start buying it more often. And then everything depends on the arithmetic — how much demand grows compared with how much productivity grew.

Hence a useful question for anyone looking for a product: what do people not do at all today, because it is too expensive?

Not "what do we do and how do we make it cheaper", but "what never gets done, though it should".

I will apply the question to my own industry, because I understand it better than any other.

Somebody buys a flat in Almaty. It is probably the largest transaction of their life. Do they order a legal check on the property's history — every transfer of title, the encumbrances, the litigation, the bankruptcies of previous owners, the legality of the alterations, third-party rights? In the overwhelming majority of cases, no. Not because they think it does not matter. Because the check costs real money, takes time, and anyway, "the notary looks at all that." People buy homes on trust, on luck and on an agent's word of honour.

Now imagine that check costs five thousand tenge and takes an hour. How many people order it? I suspect almost everyone — because at that price, saying no is irrational even for someone convinced that nothing bad ever happens to him.

A small business does not go to court over a debt of four hundred thousand tenge. Not because it does not want the money — because the lawyer, the court fee and six months of aggravation cost more than the dispute. That is not an absence of demand. That is demand crushed by price. It has not gone anywhere; it simply does not show up in the statistics, and so, in effect, it does not exist.

Nobody shows a lawyer an employment contract before signing it. The contract with the renovation crew gets read after the crew vanishes. The commercial lease gets discussed with a lawyer at the eviction stage. In every one of these cases, prevention is dozens of times cheaper than the cure — and still nobody buys it, because at the moment of signing it feels like money thrown away.

My hypothesis: the opportunities in legal work lie not so much in making cheaper what people already order as in the checks nobody has ever ordered. It is a vast layer of non-consumption hidden underneath a price.

But here I am obliged to be honest all the way, or this turns into advertising rather than analysis.

First. The technical replaceability of a task and the economic redundancy of a worker are entirely different things. That a machine can perform a step does not mean the person who used to perform it is surplus: he may perform ten other steps, and his value may even rise. Surveys about "what percentage of tasks can be automated" systematically confuse the two questions.

Second, and this one is nastier. Growth in the number of orders does not by itself guarantee growth in employment. If there are ten times more checks but each can be done fifty times faster, then in total there is less work for people, not more. Demand has to grow faster than productivity — otherwise the market's expansion does not save the worker, it merely makes the service cheaper.

Nobody knows these coefficients in advance. Not me, not the authors of the cheerful forecasts, not the authors of the gloomy ones. Anyone who claims to know is probably selling something.

V. By cutting its juniors, a company eats its own future

Now for that landmine.

In the summer of 2026, researchers at the Stanford Digital Economy Lab updated a paper with a telling title: "Canaries in the Coal Mine". They look at data from the American payroll processor ADP — real pay stubs, not surveys — from November 2022 through June 2026.

The central finding is not that everyone was fired. It is the shape of the change.

Young workers aged 22 to 25 in occupations highly exposed to AI show an employment shortfall of roughly 19% relative to their peers in less affected fields. A year earlier, in the July 2025 data, that gap was 15%. And the roles that sag are precisely the ones where AI is used to substitute for human tasks; where it is used to augment the worker, employment holds up or grows — especially among experienced specialists.

The authors, to their credit, are careful: they state outright that these are descriptive patterns, not an estimate of a causal effect; that some of the differences are explained by education; that part of the divergence began before generative models spread widely; and that they observe no widespread displacement of workers across the economy.

But notice the character of the shift. The change is happening not where it is noisy — not in layoffs — but where it is quiet: in hiring. Nobody is being thrown out. They have simply stopped taking new people on.

This is an exceptionally insidious form of change. A person who is fired is an event. He goes to the union, to the courts, to social media; he gets written about. A person who is not hired is a non-event. Nobody counts him, he never turns up in the news, his absence is invisible to everyone except himself.

A career ladder can be dismantled from the bottom entirely unnoticed by those already standing on the upper rungs. You come to work, everything is fine, you have even become more efficient. You do not see that there are no longer any rungs beneath you. That becomes clear in about seven years.

And here is the trick. An entry-level position is always two things at once. It is a job. And it is a way of accumulating experience — that is, a school.

The junior lawyer who spends two years digging up case law, compiling spreadsheets, proofreading contracts and writing drafts that a senior then goes over with a red pen is not merely doing cheap work. He is, in that moment, being manufactured as an expert. Through those boring two thousand hours the instinct forms: where in a contract the problem is usually buried, why this particular clause looks fine but will fall apart in court, what it actually means when a client says "everything there is clean."

If you remove the bottom rung to save money this quarter, a question arises that nobody today has an answer to: where, in seven years, will the people capable of judging complex decisions come from?

Go back to the section on accepting the result. Value shifts to setting the task and judging the outcome. But the ability to judge an outcome grows only out of the experience of producing one. A judge who has never written a claim will judge worse. A partner who has never sat in an archive will not feel that something is missing from the file. An editor who has never written will correct commas.

We risk optimising away exactly the stretch of road on which experts were made, and discovering a few years later that we have fantastically fast tools — and nobody able to say whether the result is any good.

The conclusion for any organisation that intends to exist longer than three years: if routine assignments no longer work as a school for the profession, training will have to be designed separately and paid for deliberately.

Training used to be a by-product of routine, and so it seemed free. Nobody set aside a budget for "producing experts" — it happened by itself, like heat from a running engine. Now the engine has become more efficient and runs cooler. Which means you will have to install separate heating: real matters from day one, mentoring as a paid obligation rather than a favour, systematic review of mistakes, a gradual handover of responsibility — and, most important of all, the right to make mistakes under controlled conditions.

This is expensive. It does not pay for itself this quarter. Which is exactly why most will not do it — and exactly why those who do will, in a few years, be the only ones with people.

VI. The unexpected revenge of the humanities

For the past twenty years, parents and schoolchildren have been told something simple: philosophy is not a profession, history is a hobby, go and study programming. In Kazakhstan this was said with particular insistence, and broadly not without reason: the market really did pay for engineering skills.

And now a strange turn is possible.

When options were few, what counted was the ability to produce an option. When options become unlimited and are produced in seconds, what starts to count is the ability to choose — to work out which of them deserve attention, and why.

Philosophy, history, literature, art, music are — if you strip away the lofty words — gyms for judgment. They do not supply knowledge you can apply on Monday morning. They supply trained discrimination: how an argument differs from rhetoric, how form differs from substance, why one decision is called great a hundred years later and another merely fashionable.

I would put the task of education today roughly like this: teach people to tell the beautiful from the persuasive, the persuasive from the true, and the useful from the merely impressive.

This is not an abstraction. A language model, by its very design, is magnificent at producing the persuasive. That is at once its greatest strength and its central danger. A text that sounds like an expert opinion and a text that is an expert opinion differ by signs you have to be able to see. A person without trained judgment will not see the difference — and, worse, will not know that he does not.

But an immediate caveat is required here, or this turns into a pleasant fairy tale for humanities graduates.

Taste without a technical foundation is just opinion with airs. A lawyer who "senses" that a citation is wrong but cannot open the code and check it is useless and dangerous. A doctor with wonderful clinical instincts and no knowledge of pharmacology is a catastrophe with a human face. Judgment that does not rest on the ability to verify degenerates into subjective impression, and a subjective impression is easy to fake — which, as it happens, is exactly what machines have learned to do best.

So this is not a story about the humanities defeating the technical disciplines. It is that the division was always false, and now it is becoming expensive as well. You need both halves: the ability to verify and the ability to judge. The first without the second gives you an executor who will soon be cheap. The second without the first gives you an expert who can be fooled by a well-turned paragraph.

VII. "A human made this" is a feature, not an excuse

In 1997 Deep Blue beat Garry Kasparov, and everyone wrote that chess was finished. Why watch people play worse than a machine?

Almost thirty years have passed. Chess is enjoying the best period in its history: millions play online every day, tournaments draw bigger audiences than ever, grandmasters have become media figures with streams and millions of subscribers. A machine that plays incomparably better than any human is available free on a phone — and it has done nothing to reduce interest in how humans play.

Cars have been faster than people for well over a century. Tens of thousands sign up for marathons.

From which follows a hypothesis I consider underrated: "a human made this" may become a meaningful feature of a product rather than an apology for its imperfections.

But caution is needed here, or it turns into a comforting mantra for the frightened. This effect is far from universal.

Nobody pays a premium because an X-ray was read by a human rather than an algorithm. Nobody takes pride in a balance sheet reconciled by hand. Where the value lies exclusively in the correctness of the result, the origin of the result interests no one — and rightly so.

The effect works where the value lies in the relationship, not only in the result.

And here, for me, is the most important dividing line in this whole essay. There is the quality of a result — and there is the value of a relationship. They are different things, and they are constantly confused.

You can get a technically stronger answer from a machine and still prefer to talk to a particular human being. Because of trust built up over years. Because of a shared history in which you do not have to explain the context. Because of the simple wish to be understood by this person, not understood in general.

For professions built on trust this is critical, and it is poorly understood inside those professions themselves. A client who comes to a lawyer is not coming only for a document. He is coming so that someone will take on part of his anxiety, so that someone will say: "I understand your situation, I have done this a hundred times, here is what we are going to do." A patient who goes to a doctor wants more than a diagnosis. A parent choosing a school is not choosing a curriculum.

The professional who thinks he is selling a text or a diagnosis will be displaced by whoever produces texts and diagnoses more cheaply. The professional who understands that he is selling the removal of uncertainty and shared responsibility is in a fundamentally different position.

And still — a caveat, without which this section would be dishonest. This is a direction in which to look for value, not a guaranteed refuge. I am not claiming that human contact will lie forever beyond the reach of machines; I am claiming only that today there is value there which very few people are deliberately looking for.

VIII. Even abundance will not abolish competition — it will change what we compete for

Suppose for a moment that the optimists are right. Suppose basic goods and services get radically cheaper: a decent medical consultation, respectable legal support, personal tutoring, workable software for any purpose become available to almost everyone.

Will competition disappear?

No. It will relocate.

Because there is a scarcity that cannot be eliminated by production. Recognition is relative by its nature. You cannot make everyone the most respected — that is exactly the same logical impossibility as making everyone taller than average. However many goods an economy produces, the number of places at the top of any hierarchy stays limited by the very definition of a top.

Which raises a question I consider genuinely important and almost never discussed: where will a society channel the competition for significance once the competition for survival weakens?

History gives different answers, and not all of them are encouraging. Classes freed from the need to earn a living directed the released energy in various ways. Sometimes into science, collecting, patronage and travel; that is how the European science of the modern era came about, made by people who did not need a salary. Sometimes into duels, intrigue, conspicuous consumption and elaborate boredom. Freedom from want does not by itself make a person either happy or useful. It merely releases a resource that has to go somewhere.

Hence a conclusion of mine that goes beyond economics: the future cannot be described through production and consumption alone.

Discussion of the future is now conducted almost entirely in terms of GDP, jobs and incomes. As though the only question were how much people will produce and how much they will consume. But a person needs more than to consume. He needs to be needed. He needs grounds for self-respect that he himself finds respectable.

Which means a society will need worthy ways to earn recognition — through research, mastery, care for others, public service, mentorship. And ways that genuinely count as achievement in that society, rather than being handed down from above as a consolation prize for surplus people.

Otherwise material abundance will sit very comfortably alongside a mass sense of one's own uselessness. And that, unfortunately, is not a hypothesis. Societies where basic needs are met better than at any point in human history report rates of loneliness, anxiety and meaninglessness that in no way follow from their level of wealth.

Being well fed is not a synonym for being well. For some reason we have to learn this again every time.

IX. Price tells us less and less about value

What does a search engine cost? Zero.

What is it worth? In one study by Erik Brynjolfsson and his colleagues, people were asked how much they would have to be paid to give up search engines for a year. The median answer ran into thousands of dollars — orders of magnitude more than the zero this service contributes to consumer spending.

This is a gap that traditional statistics cannot see, by construction. GDP counts turnover: how much money changed hands. A service that brings a person enormous benefit and is free contributes essentially nothing to GDP. What is more, if a new tool lets you do without a paid service, GDP formally falls — even though the person is better off.

The approach known as GDP-B proposes measuring not only monetary turnover but the benefits received — how much people would be willing to pay in order not to lose what they already have for free. It is a supplement to GDP, not a replacement: you still have to count money, because wages are paid in money, not in utility.

For anyone building a product, a very practical research technique follows.

The usual question: "How much would you pay for this?" People answer it badly. They are polite, they do not know, they name figures out of courtesy or out of a wish to haggle.

A different question: "What would you lose if this disappeared tomorrow?"

Call it the shutdown test. The answer "well, it would be inconvenient" and the answer "I could not work, I would have to hire someone" describe two completely different businesses, even if both products cost the same today. The first is a nice feature. The second is part of the infrastructure of somebody's life or work.

Again, an honest caveat: this does not remove the need to test for paying demand. Enormous benefit with zero willingness to pay is charity, not a business, and a great many wonderful products died exactly that way. But the shutdown test helps distinguish what people like from what they can no longer do without. And that, as a rule, is the difference between a product you will have to sell and a product people start buying on their own.

X. A rich economy does not guarantee well-being to the people living in it

Now for the most important thing — most important to me, at any rate, given what I do for a living.

Automation can simultaneously increase the total volume of value created and weaken the position of specific people whose labour becomes replaceable. These two processes do not contradict each other. The economy as a whole grows richer; the individual worker's bargaining power falls.

And here it is important to understand something usually presented as a technical detail, though it is deeply political.

There are two different ways to apply a technology. You can augment a person — give them a tool that makes them three times more effective, and thereby raise the value of their labour. Or you can replace a person — build a system that lets you do without them.

Technically, these are often the same algorithms. Economically and socially they are two completely different machines, because they distribute income and bargaining power differently.

A tool that makes a lawyer three times more effective strengthens his market position: he is needed, and he is now worth more. A service that lets the client do without a lawyer weakens it. Both can be written on the same framework by the same team in the same sprint.

And the choice between them is determined, as a rule, not by the technology. It is determined by who is paying for the development and what they are optimising for. A customer who pays the salaries more often funds replacement. A customer who does the work himself more often funds augmentation. This is not a conspiracy — it is the arithmetic of interests, and it works silently.

Hence my main social conclusion: helping people through the transition is part of deploying the technology, not charity after the deployment is over.

The current habit is to reason like this: first we will deploy it, get the effect, and then, if social problems come up, we will think of something — retraining, benefits, support programmes. The trouble is that "then" arrives years later, and by that time the person who fell out of the profession at forty-five has already lost his qualifications, his connections, his confidence and his time. Retraining works as prevention and hardly works at all as resuscitation.

And one more thing: it is not enough to show that the economy as a whole has won. "On average" is a poor unit of measurement for a human life. The average temperature across the hospital is no comfort to a particular patient.

So to any deployment — in a company, in an industry, in a state programme — I would attach three questions.

Who gained new opportunities?

Who lost the ones they had?

And — the most important and the most rarely asked — how can a person move from the second group into the first?

If there is no concrete answer to the third question, with names, deadlines and money, then you do not have a deployment strategy. You have half of one, and the other half will be paid for by somebody else, usually by those least able to afford it.

What not to get confused about here

I am obliged to say a few things plainly, because texts like this are easily read as a promise.

Nothing listed here is a guaranteed refuge. Setting the task, taste, judgment, human relationships — these are directions in which it is worth looking for value today. They are not fortresses machines will never take. Every time someone has declared a particular human quality fundamentally impossible to reproduce, the shelf life of that claim has turned out to be shorter than expected.

The employment data calls for care. The researchers I cite say so explicitly: the differences observed do not prove a causal effect of AI, part of them is explained by other factors, and no widespread displacement of workers across the economy is visible yet. A warning signal is not the same thing as an established fact. Anyone who tells you confidently today what percentage of jobs will vanish by such-and-such a year is either mistaken or selling consulting.

And finally. I am writing this from Almaty, not from Silicon Valley, and the difference matters. In an economy where a substantial share of employment sits in the resource sector, the civil service and small businesses with low digital maturity, the curves will look different — flatter, stretched out over time. But stretched out is not cancelled. It is, if you like, a gift: time to prepare. It would be a shame to spend that time arguing about whether any of it will happen at all.

What to do about all this on Monday

Light got cheap — and we did not simply read more in the evenings. We built a nocturnal civilisation that had never existed and that nobody had designed.

Applied intelligence is getting cheap. And the main mistake almost everyone is making right now — from the individual specialist to the ministry — is that we are fitting the new thing to the old. We ask: how do we do faster what we already do? By how much will document preparation time drop? How many people can we avoid hiring?

Those are questions about candles. Legitimate, but small.

The more interesting question sounds different: what work has become possible only now — and why does someone need it?

Which checks that nobody ever ordered. Which services that were unprofitable at any volume. Which decisions that used to be taken blind, because working them out cost more than getting them wrong. What help for the people a professional service never reached — and in my field, incidentally, that means most of the population.

I do not know how it will all turn out. Nobody does, and the healthy response is not to pretend otherwise. But you still have to choose a strategy, and mine is this: it matters more, strategically, to learn not how to do the old work faster, but how to notice what new work has now become possible.

Intelligence got cheap. What is left is to work out what we want.

That, it seems, is the only work that cannot be delegated.

Sources and further reading

• Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence — E. Brynjolfsson, B. Chandar, R. Chen, Stanford Digital Economy Lab
• No Widespread Displacement, but the AI Employment Gap for Young Workers Has Widened to 19% — data update, August 2026
• The Productivity J-Curve: How Intangibles Complement General Purpose Technologies — E. Brynjolfsson, D. Rock, C. Syverson, NBER / AEJ: Macroeconomics
• GDP-B: Accounting for the Value of New and Free Goods in the Digital Economy — E. Brynjolfsson, A. Collis et al., NBER
• Using massive online choice experiments to measure changes in well-being — PNAS, on estimating the consumer value of free digital services
• What AI Really Changes in Kazakh Labor Market — The Astana Times, estimates of the transformation of Kazakhstan's labour market
• AI in Kazakhstan Country Report — country report, January 2026