Until not long ago I still believed that the adoption of AI by firms, with models ever more capable and general, would wreak serious havoc on the labour market: a very aggressive reconfiguration of the workforce and a fall in labour demand as productivity took off. It was the logical conclusion: if a firm can produce the same or more with fewer people, it will.
Day-to-day experience, however, is showing me the other side of the coin, and it has to do with a mechanism inherent to the capitalist system itself: competition.
Producing the same with less, or more with the same
Efficiency has two possible uses. It can be used to produce the same with fewer resources, or to produce more with the same ones. And the choice between the two is not free.
When a firm brings AI into its internal processes, its productive capacity grows markedly: for the same price, or a lower one, it can offer far more product, or a far more complete one. At that point, its competitors are left without the opposite option. They can cut headcount and keep offering the same as always, now cheaper to produce, and the quarterly accounts will be grateful. But it is a short-termist profitability: whoever offers the same thing — or a notably inferior product — against someone offering considerably more for the same price is out of the market before long. Efficiency translates into a wider offering, not into cuts.
Something very human comes into play here: competitiveness. When there is someone on the other side, you do not play for a draw; you play to take it all. And it is not just instinct: it also happens to be the smart move in the medium and long run. Cutting improves today’s bottom line; expanding is the only thing that keeps you from falling behind tomorrow. That is why the key is not to shed the less productive resources but to use all of them, more productive or less, combining them as efficiently as possible and even adding to them. In a race for the maximum, every resource adds output, and giving one up is giving away ground. It is a game of maxima, not of optima.
Abundance without mass unemployment
The result is a scenario of abundance in which the minimum that is expected keeps rising as the technological frontier shifts. And to keep pace, firms need to keep expanding their productive resources, labour among them.
I think this explains why studies are beginning to appear that find a relationship between AI adoption and job creation or, at the very least, observe no significant impact on job destruction. The ECB, drawing on a survey of some 5,000 euro area firms, finds that those making intensive use of AI are about 4% more likely to hire, not fire. And the detail that interests me most: that effect is driven by firms using AI to scale up output and R&D, whereas those using it to cut labour costs — barely 15% — do shed jobs, but not enough to offset the balance. Across the Atlantic, an NBER working paper based on nearly 750 executives reaches a similar conclusion: productivity gains come mainly through revenue, innovation and demand channels, and there is little evidence of near-term aggregate employment declines, though there is an internal reshuffle with fewer routine clerical roles and more skilled technical profiles.
As in any technological revolution, there will be a reconfiguration of the workforce: occupations that disappear and others that are born. But it does not look as though we are heading for mass unemployment, at least not in this pre-AGI or pre-ASI phase.
The exception: where there is no competition, there are cuts
Everything above rests on one premise: that there is someone on the other side willing to offer more for the same price. When that premise fails, the reasoning flips.
In a monopoly, in a cartel or in the public administration itself there is no price competition. Nobody is going to push you out of the market for offering the same as always, because the customer has nowhere else to go. Without that pressure, nothing forces you to produce more with the same resources or to improve the product. The incentive runs the other way: keep the efficiency gain as margin, passing it on neither to the price nor to the product, and the most comfortable way to do that is to produce the same with less. Here AI does not translate into a wider offering but into cost reduction, and the easiest cost to cut is labour.
It is in these settings that net job destruction can indeed appear. Public employment, in particular, is a case to watch. It competes with nobody, carries a permanent pressure to contain spending, and concentrates a large share of the routine administrative tasks that AI already does well.
The other variable: the elasticity of demand
There is a second condition, besides competition, for efficiency to turn into more output rather than fewer jobs: the market must be able to absorb the extra production. In other words, demand must be elastic.
If demand is elastic, every product improvement or price cut drives a more than proportional increase in sales, and it pays the firm to produce more with the same resources, whether or not it has competitors breathing down its neck. If demand is inelastic — because the market is saturated or because the need is fixed — producing more only sinks prices. In that case not even competition saves jobs: the only way to hold margins is to cut costs, and we are back in the previous scenario.
None of this is new. Bessen documents it with two centuries of data from US textiles, steel and automobiles: while demand was elastic, automation created jobs in those industries for over a century. Once markets saturated and demand turned inelastic, the very same productivity gains began to destroy them. Public administration is again the extreme example: demand for permits, case files or tax returns is what it is, and no efficiency gain is going to generate more of it. At most, an improvement in processing and response times, whose upside is limited.
Crossing the two variables, the map is fairly clear. With competition and elastic demand, AI turns into more product and employment holds. Without competition but with elastic demand, nobody forces the firm to expand, but it may pay to do so. With competition and saturated demand, a price war and cost cutting. And without competition and with inelastic demand, job destruction is almost certain.
The bottleneck is the user
Let us return to the first quadrant, competition plus elastic demand. There, efficiency does turn into abundance: more product, more features, more complete. But that elastic demand has a limit, and it is not price. Demand is not exercised by an abstract market; it is exercised by people, and people have two resources that do not scale with technology: time and attention.
The risk of saturating the user is becoming ever more real. Applications that suddenly start doing everything, piling on more and more features, end up overwhelming someone who, in reality, only wanted to solve a handful of very specific needs. An excessively complex or overloaded system drives them towards simpler solutions that address what they are looking for directly. We do not have infinite time to learn the whole solution, nor infinite attention to deal with matters we never intended to take on when we bought the product.
On top of that comes predation among the applications themselves: as they multiply, many solutions proliferate for the same pain point and it becomes impossible to keep track of every alternative on the market.
When everything is abundant, the virtue is to produce scarcity
That is why, today, the main problem is no longer the technical or epistemological bottleneck — a gap that AI closes to a greater or lesser extent — but something far harder to find: common sense. The ability to see clearly what is needed and what is not, to put oneself in the user’s shoes rather than the product’s, and to simplify rather than diversify. It is a soft skill intrinsic to each person’s mental setup, and only experience trains it.
With AI it is very easy to build software, generate knowledge, add features, cover everything. None of that used to be easy: it demanded time and people, and that constraint forced a trade-off that has now disappeared. You no longer have to choose because you can do it all, and precisely for that reason choosing is once again what makes the difference. When everything is abundant, the virtue is to produce scarcity, which is, in the end, the basis of value.
Hybrid profiles, not AI specialists
Which profiles fit this scenario? People who know where they want to go. Hybrid profiles, with a minimum of technical knowledge and a grounding in the business, able to build fast but without over-engineering, knowing what the user needs and delivering it in a way that is simple and intuitive. This, however, takes experience, which once again raises the entry barrier for juniors, already displaced because the tasks they used to take on to gain that experience are now done by AI.
On the other hand, I think profiles specialised in applying AI will disappear. As it becomes more intelligent, it becomes easier to work with, and there will no longer be a need for a specialist who knows how to squeeze the most out of it. It will become a basic technology: a requirement, not a merit. As already happened with office software or with the ability to browse the internet, in almost any skilled occupation we will take for granted that a person knows how to use it. It will become a commodity, and the value will remain in recognising where to apply it: the use cases.
What is going to be scarce, and what is going to be worth something, is judgement.
Further reading
What I describe here is a reflection from the trenches, not a study. But the topic already has a sizeable literature, and the evidence coming in points in the same direction: no mass unemployment for now, a notable reshuffle of tasks and profiles, and a serious warning for newcomers. Some references if you want to dig deeper:
- AI and jobs. A review of theory, estimates, and evidence (del Rio-Chanona et al., 2025). A survey synthesising theory, exposure measures and empirical evidence. A good starting point: productivity gains of 20-60% in controlled experiments and 15-30% in the field, with signs of declining demand for novice workers.
- Still Waters, Rapid Currents (Humlum and Vestergaard, NBER, rev. 2026). Danish administrative records: two years after ChatGPT, null effects on earnings and hours (ruling out impacts above 2%), but task reorganisation and new AI oversight and integration roles.
- Canaries in the Coal Mine? (Brynjolfsson, Chandar and Chen, Stanford, rev. August 2026). ADP payroll data in the US: no widespread displacement, but employment of workers aged 22-25 in exposed occupations is 19% below that of their less-exposed peers, through reduced hiring rather than layoffs. Where AI complements rather than substitutes, employment is flat or rising.
- Generative AI as Seniority-Biased Technological Change (Hosseini and Lichtinger, 2025). Using data on 62 million résumés and 285,000 firms, they find that junior employment falls in firms adopting generative AI while senior employment keeps growing.
- AI and the US labour market: effects on employment growth (ECB, Economic Bulletin 4/2026). Between 2019 and 2025, occupations at high risk of substitution grew around 15 percentage points less than low-risk ones, accelerating after ChatGPT.
- Automation and Jobs: When Technology Boosts Employment (Bessen, Economic Policy, 2019). Two centuries of data from US textiles, steel and automobiles show that automation creates jobs while demand is elastic and destroys them once the market saturates. The reference framework on the role of demand.
- The Simple Macroeconomics of AI (Acemoglu, Economic Policy, 2025). The sceptical counterpoint: a task-based model estimating modest aggregate productivity gains (under 0.66% TFP over ten years) and a widening gap between capital and labour income.
- GPTs are GPTs (Eloundou et al., Science, 2024). The classic exposure reference: around 80% of US workers have at least 10% of their tasks exposed to LLMs, with higher exposure in high-wage jobs.
- Generative AI at Work (Brynjolfsson, Li and Raymond, QJE, 2025). The most cited field experiment: an AI assistant in customer support raised productivity by 14% on average, with the largest gains among the least experienced workers.
This is a selection biased towards the most cited and the most recent; dozens more appear on Google Scholar every month.