AI doesn’t fire the junior, it stops hiring them

In February, Jack Dorsey signed a letter to Block’s shareholders that is worth reading twice. It announced that the company would shrink by almost half, from more than 10,000 people to just under 6,000, because a considerably smaller team, using the artificial intelligence tools the company itself builds, can do more and do it better. The same letter presented 2025 as a strong year. In May, Cloudflare announced that it would let go of some 1,100 people in order to adopt an operating model based on AI agents, in a quarter in which its revenue had grown by 34%. For decades, companies laid people off when they sold less. Now they lay people off while selling more, and they call it productivity.

I recognise the temptation to add up these announcements and conclude that the wave of white-collar job destruction has already arrived. But that would be populist and irresponsible. Not even the most unsettling study I know of, which I will come to shortly, finds a generalised fall in employment caused by AI. What is beginning to show is quieter and, in my view, more dangerous. For now, AI is closing the front door on a great many young people. And in Europe, where tax systems live largely off taxing payrolls, a pay slip that never comes into existence is a fiscal problem even before it is an employment problem.

Corporate announcements are best read with a certain suspicion. Block and Cloudflare are real cases, but they are not a statistic. Atlassian cut close to 10% of its workforce and explained it with a mixture of restructuring, efficiency and the need to free up money for AI and for sales to large clients. In May, the AP news agency observed that AI is rarely the only reason companies give when they lay people off. What can I say? “We are migrating to an agent-based model” sounds much better to analysts than “we hired too many people three years ago”. The first sentence sells, while the second admits a mistake. AI is not emptying companies overnight. The change that is hardest to see happens earlier: entry-level positions that are no longer offered.

Nor do the most widely circulated figures help. The ILO estimates that one in four people employed worldwide works in an occupation with some “degree of exposure” (a gentle way of saying at risk of redundancy) to generative AI. The IMF puts the share of affected jobs in advanced economies as high as 60%. For Spain, for example, the OECD puts “exposed” employment at 27.4%, with Madrid at 36.6% and Castilla-La Mancha at 19.6%, a gap that says more about where banks and consultancies have their headquarters than about the technology. Read in a hurry, these figures look like unemployment forecasts. And I am unable to say whether they really are or not.

For the OECD, being “exposed” means that at least a fifth of the tasks in a job could be done in half the time with the help of AI. A lawyer who used to spend an afternoon preparing the first draft of a contract and now takes an hour is highly “exposed”, and their firm may well bill more and hire another lawyer. Provided, of course, that it has more clients. That is why the IMF includes in its 60% the jobs in which AI complements the worker, which it estimates at around half. Seen this way, “exposure” does not translate directly into unemployment. It tells us where to look, and little more.

Productivity can rise without wages improving or more jobs being created. It depends on who keeps the savings: the company, the worker or the customer.

The figure that really worries me comes from Stanford. A team at the Digital Economy Lab has worked with ADP payroll data, covering millions of American workers, up to June 2026. They find no general fall in employment. But employment among 22- to 25-year-olds in occupations “exposed” to AI is 19% below where it would be had it kept pace with that of people the same age in less exposed occupations. And that gap does not come mainly from lay-offs. It comes from hires that were never made.

The authors present this as early, descriptive evidence, with no claim to causality, and the OECD, in its July employment report, is more cautious still, judging that the role of language models in young people’s labour-market difficulties appears, for now, to be limited. I will not hide the disagreement. But the United States is ahead in adoption, and I would not be surprised if Europe were seeing the early version of what is already showing up in American payrolls.

The explanation is almost mechanical. What generative AI does reasonably well is summarise, write a first draft, review documents, classify information and write simple code. Exactly what the newcomer used to do. Anyone who has spent time in a law firm, an audit firm or a bank’s research department knows that those tedious tasks were a paid apprenticeship. By reviewing a hundred contracts, one learns to recognise the clause that matters in the hundred and first. The partner of 2040 is the junior who today reviews documents at eleven at night. If that junior is not hired, this year’s savings are paid for with a lack of judgement fifteen years from now.

And here lies the statistical trap. Nobody counts a young person who is not hired as having been dismissed. They do another master’s degree, accept a worse-paid job outside their field or stay a few more years in their parents’ home, something we in Spain already do with considerable ease. The unemployment rate will be slow to pick it up. Entry-level salaries, internships that are never advertised and the training companies stop paying for will register it first.

In Europe, where a large part of tax revenue depends on payrolls, the evidence will also arrive through the tax office.

None of this denies that productivity exists. At a customer service company studied by NBER researchers, access to an AI assistant raised the number of queries resolved per hour by around 14%. Among novices the improvement was around 35%, while the most experienced staff barely noticed any difference. There is something ironic in that result. The tool that helps the novice most is the same one that allows the company not to hire them.

The European Commission asked Europeans in February and March, and 91% of those who use AI at work say they finish sooner, with an average reported saving of 7.4 hours a month. Almost a full working day. I suspect few have taken it off.

So the question is who keeps that time saved. A 14% gain in productivity does not mean 14% fewer staff or 14% higher pay. The company can serve customers faster with the same people, sell more and hire, leave vacancies unfilled or invest the savings in something else, and whichever of these responses predominates decides almost everything else. If productivity lowers prices, raises real wages and opens up new markets, households spend more and some of the displaced employment is reabsorbed, as happened in earlier technological waves. If the savings stay in the margin, it is a different story.

A euro that is no longer paid out in wages does not disappear. It can come back as a dividend, as investment or as a lower price. But it does not return to the same place or at the same pace. The clerk who loses their job, or the graduate who never gets one, spends almost everything they earn at the supermarket, on rent and at the corner bar. The shareholder who receives that euro as a dividend saves a much larger share of it. If income shifts towards those who consume a smaller proportion of each extra euro, output can grow faster than spending. A more productive economy and weaker households at the same time. It sounds like a paradox, and it is pure arithmetic.

The IMF has sketched two scenarios for AI. In the one it calls runaway diffusion, with rapid adoption concentrated in a few companies, consumption weakens as labour displacement increases, household balance sheets deteriorate and defaults rise. It is not a forecast, and the Fund itself warns that there is no baseline future that can be taken for granted. Meanwhile, investment in data centres, chips and electricity is propping up demand, so the blow, if it comes, is unlikely to show up in next year’s GDP. It is of little use to an investor to know which company “uses AI”, because soon they all will. What is useful is knowing who keeps the savings: the customer, the worker or the shareholder.

That the jobs most affected by the arrival of AI are those of the youngest is no coincidence. It is a signal.

This brings me to the part that, for Spain, strikes me as most concrete. According to the OECD, using 2023 data, social security contributions accounted for 34.7% of all Spanish tax revenue, compared with an OECD average of 25.5%, and those paid by employers alone amounted to 25.9%. Personal income tax (IRPF) contributed another 24.4%, and in Spain that tax rests mainly on earned income. Together, they make up almost six out of every ten euros of revenue, and most of it depends directly on there being payrolls. VAT, another 17.6%, is largely paid by those same employees when they consume.

A machine pays no contributions. Imagine a medium-sized insurer that replaces three administrative jobs with a subscription to an AI model. Three employer contributions, three employee contributions and three income tax withholdings disappear. In exchange, the tax authority collects whatever tax is due on the insurer’s additional profit, and the model provider’s margin is taxed wherever that provider declares its profits, which is rarely Valladolid. No simple rule of three applies, because contribution ceilings and rates alter the result, but a system that rests on payrolls is, by construction, the most sensitive to any change that reduces them.

Some will say that corporation tax will make up for it. It may raise more if profits grow and are declared here, but in Spain it contributes less than one in every ten euros, below the OECD average of 11.9%, and the profits from AI tend to be concentrated in multinationals whose tax residence has little to do with the place where the work was replaced. The IMF has warned about this. If AI reduces labour’s share of income and capital is taxed less, the tax base narrows precisely when more people need to be redeployed.

Are we there yet? We do not know. According to the INE, Spain’s national statistics institute, 21.1% of Spanish companies with ten or more employees were using some form of AI technology at the start of 2025, but that measures use and says nothing about substitution. Nor is there a reliable official count of lay-offs caused by AI, and even if there were, no register would show whether a model is today doing the work of the person who was never hired.

That is why I believe we are looking at the wrong indicator. We count lay-offs with names and surnames because they appear in the newspapers and because a chief executive signs them off. What ought to be monitored is considerably more boring: social security registrations by age and occupation, entry-level salaries, junior vacancies and, above all, how contribution revenue evolves relative to GDP. The day the economy grows and contributions fall behind, we will know that substitution has begun, even though nobody will have walked out of any office carrying a cardboard box.

And there is an irony I cannot get out of my head. The young people who today cannot get in because a machine does their apprentice work more cheaply are the very people the system will need, twenty years from now, to be paying contributions to fund the pensions of those who did get that first job.