On May 26, Sam Altman said he had been "pretty wrong" about AI quickly erasing entry-level jobs. Dario Amodei simultaneously walked back his earlier warning that the technology could wipe out half of all office jobs in the United States. One weekend was enough for one of the hottest forecasts in the AI debate to cool noticeably.

The background is no secret. OpenAI and Anthropic are preparing capital raises at valuations close to $1 trillion each. Ahead of such financing, it is rational for a CEO to soften the tone toward politicians, unions and pension funds. Amodei now argues instead that if 90 percent of a role is automated, the remaining 10 percent expands — a productivity thesis that echoes McKinsey and Goldman Sachs, but runs straight against Anthropic's own warnings from last year.

Two days later, on May 28, eight Swedish tech founders published an op-ed in Dagens Industri. They called for a cross-bloc, long-term plan for entrepreneurship and AI. They want talent visas, unrestricted employee stock options and broad political consensus. They list Lovable, Legora, Sana and Tandem as heirs to Ericsson, Spotify and Klarna. Tech already accounts for 8 percent of Sweden's GDP, they write. Tech companies will create new jobs, build the export industries of the future, drive productivity and finance the welfare state.

These are two letters in the same week, from the same economic logic. Both say essentially the same thing: trust the process, AI creates more than it destroys, the market solves distribution.

They may be right. It is worth asking whether they can be.

The first contradiction: job creation is symbolic

The simplest way to check a claim about job creation is to count the jobs.

The Swedish tech companies in the founders' letter employ few people. Spotify globally had, according to its 2025 annual report, a little over 7,000 employees on average. Klarna has gone from just over 5,500 employees in 2022 to about 3,400 at the end of 2024, and was reported during 2025 to be below 3,000. Lovable, Legora, Sana, Tandem, Lassie and Kognic are probably in the range of 30–200 employees each. Together, they represent a Swedish workforce on the order of ten thousand people, only a fraction of whom work in Sweden.

Volvo Cars alone employs just over 20,000 people in Sweden after the latest round of layoff notices (down from about 23,000 before the cuts). Forestry and the forest industry together employ more than 80,000. Retail employs about 250,000.

That does not mean tech companies are unimportant — they can be strategically important without being large employers. But it does mean that the claim "creates new jobs" is a rhetorical promise, not a quantifiable forecast. If Sweden manages over the coming decade to produce ten Lovables, five Sanas and three Legoras, that creates at best 20,000–50,000 highly qualified jobs, on an optimistic count.

How does that relate to the white-collar jobs that the same AI products can automate or sharply reduce the need for? That question is not asked in the op-ed. Nor is it asked by Altman and Amodei.

In the United States, outplacement firm Challenger, Gray & Christmas counted 27,645 AI-cited job cuts in the first quarter of 2026 — about 13 percent of the quarter's layoffs. After April, the figure had risen to 49,135, equal to about 16 percent of the year's job cuts through April. Layoffs in the technology sector were simultaneously 40 percent higher in the first quarter than during the same period a year earlier.

Stanford Digital Economy Lab, in turn, has shown that young workers, aged 22–25, in the most AI-exposed occupations have seen a 16 percent relative decline in employment since broad genAI adoption, while more experienced workers in the same occupations have fared significantly better. This is not a final verdict on the whole labour market. But it is an early signal that the entry point into some knowledge professions is already narrowing.

The economic research in the field does not support the tech founders' line either. Daron Acemoglu, the MIT economist awarded the 2024 Nobel Prize in economics, writes with Simon Johnson — former chief economist at the IMF — in Finance & Development (December 2023) that "the drive toward automation is perilous". There is no guarantee that AI on its current path will create more jobs than it destroys. The productivity bandwagon invoked by tech investors — the idea that productivity gains automatically become broader prosperity — requires, in their analysis, specific conditions that are not met when automation technology dominates the direction of development.

The first contradiction in the founders' model is therefore empirical. Not ideological: empirical. When the companies that are supposed to "create jobs" are examined numerically, and when the economic literature in the field is examined, they appear instead to redistribute value — upward, toward capital, and away from broad employment.

The second contradiction: agents do not need interfaces

This is the underestimated point — the one rarely discussed in the Swedish tech debate.

The SaaS industry — several hundred billion dollars globally — is built on a specific assumption: that humans need interfaces to interact with computers. Dashboards to monitor. Workflows to coordinate work. Approval chains to govern. Reports to communicate. Forms to register. The whole economy around enterprise software assumes that the end user is a human being who clicks, writes and approves.

Remove the human from the loop. Replace that person with an agent that has direct access through APIs and new agent protocols such as MCP (Model Context Protocol — a standard introduced by Anthropic in late 2024, allowing AI agents to discover and use external tools without prebuilt integration), and that has semantic understanding of the goal and the ability to make decisions autonomously.

What is needed then? Not a dashboard — the agent monitors the data directly and raises an alert on deviation. Not workflow — the agent coordinates itself with other agents. Not reports — the agent communicates semantically in whatever format is needed at the moment. Not forms — the agent gathers what is needed wherever it exists.

The result is that a large part of the existing enterprise software market is transitional technology. Built for a short era in which computers were fast but the interface was still human. When the interface disappears, the justification for many of the companies disappears too.

That does not mean all software dies. What is likely to survive is the infrastructure layer (compute, security, data, observability), tools for building and managing agents, physical-world interfaces (sensors, control, robotics), and what is regulated into human-in-the-loop settings (healthcare, law, safety).

The rest — the broad office software economy and the millions of administrative positions that make up its user base — shrinks or transforms. Several of Sweden's strongest tech companies sit on the wrong side of that line. They are built for a world with many knowledge workers who need help. In an agentic world, they either become platforms for other agents (which is an entirely different business) or become obsolete.

Part of what disappears is not just dashboards and forms. It is the whole integration economy — Zapier, n8n, custom connectors between business systems, integration consultants — that exists because humans and their vendors could not make systems talk to each other without translation. When agents can discover and combine capabilities directly, an intermediate layer that today turns over billions globally disappears.

This is not a political claim. It is an architectural observation. The question the founders need to answer is: which of your companies has a business model that survives when agents have taken over the interaction layer?

The third contradiction: who buys?

The question has been asked since Henry Ford. The story that Ford raised wages so that his workers could afford to buy his cars is historically simplified, but the point remains: capitalism requires consumers, consumers require income, and income still mostly comes from work.

KPMG announced in May 2026 a global alliance with Anthropic and began implementing Claude for its more than 276,000 employees worldwide — initially with a focus on tax and legal services. Standard Chartered, SEB and several major banks have announced heavy AI savings. KPMG, EY, Deloitte and PwC employ thousands in the Nordics alone in audit, tax and advisory — precisely the time-consuming work now being targeted by agents. Klarna's Sebastian Siemiatkowski said openly already in 2024 that AI lets the company do more with fewer people.

What happens to purchasing power when the broad knowledge-worker class shrinks?

The standard counter is: "capital owners consume too." That is true. But they consume differently. The economy that follows does not look like today's, but like something clearly split in two:

This is not hypothetical. It is what has happened in the United States over the past 30 years, only more slowly. AI compresses the same development from decades into years.

What do Klarna's founders do when consumers no longer have income to finance purchases with? What does Spotify do when music's purchasing-power base — knowledge workers in the service sector — shrinks? What does Sana do when the customers are AI agents rather than HR managers?

And what does Sweden do when tax revenue from work falls, because the "safety system that enabled risk-taking" — which the founders rightly praise — was historically financed by broad employment and taxation of labour?

That question is not answered in the founders' letter. Nor is it present in Altman's newly optimistic tone. It is a structural question on which the entire economic model rests, and both camps elegantly pass it by.

The serious counterargument: Jevons paradox

The strongest counterargument to the reasoning so far is Jevons paradox: when a resource becomes more efficient, total consumption tends to rise rather than fall. When the cost of cognitive labour falls toward zero, demand can explode. We begin applying AI to problems that were previously too expensive to solve — hyper-personalized diagnostics, continuous legal review, real-time analysis of every municipal budget, individualized education for every student. On that reasoning, AI creates an explosion of demand that absorbs the freed-up labour into new roles for which we do not yet have a vocabulary. This is the same logic Amodei now leans on when he says that if 90 percent of a task is automated, the remaining 10 percent expands.

The argument deserves to be taken seriously. It has historical grounding — every technology wave since industrialization has followed the pattern.

Korinek himself addresses Jevons paradox in the VoxDev podcast and reaches the same conclusion as the one that follows here. Even if the new demand absorbs part of the workforce, the transition remains deeply disruptive, because workers are forced into the ever narrower roles where human input is still required. He warns of a sharp devaluation of cognitive labour — exactly the kind of work that today carries Nordic middle-class consumption.

There are two specific reasons to be cautious about transferring the historical Jevons pattern directly to the AI wave.

The first is architectural. In previous technology waves, the new demand expanded within the domain of humans — more cars required more car mechanics, more computers required more IT consultants, more websites required more web developers. The expansion Amodei describes — "the remaining 10 percent grows" — assumes that the 10 percent is reserved for humans. There is no structural reason it would be. If agents can do 90 percent of a task today, there is no architectural barrier to their doing the expanding remainder tomorrow. The technology that opens demand is the same technology that can absorb it.

The second is practical policy. Even if Jevons holds — even if new jobs appear in sectors we cannot yet name — it does not solve the consistency problem. Policy still needs to know where the new jobs emerge, when they emerge, which skills they require, how the transition is financed and how the tax base is preserved in the meantime. A future productivity boost is not a labour-market plan. It is an assumption.

The point here is not to forecast the exact number of eliminated jobs. It is to show that the founders' model lacks a measurement and distribution mechanism for the scenario they themselves make technically possible.

A consistency analysis, not a political objection

It would be wrong to read this as an attack on the tech founders. The Swedish signatories have a perfectly valid point about talent visas and employee stock options. Sweden competes for talent globally. Employee stock options are genuinely worse designed here than in London, New York and Berlin. Cross-bloc long-term policy for business conditions is a reasonable political request.

Their strongest argument is not even the one they write — but the one they implicitly assume. The AI transition will happen regardless of what Sweden does. In that case, the choice is not between "no disruption" and "disruption plus Swedish companies," but between "disruption plus we capture some of the value" and "disruption plus we capture none at all." From that perspective, talent visas and employee stock options are reasonable measures.

But that is a different argument from "tech creates more jobs than it destroys." That sentence lacks boundary conditions and allows the senders to avoid the follow-up question: how do you combine "we build AI companies that are meant to do more with fewer people" with "we need broad employment to finance the welfare state"?

It is the same contradiction Altman and Amodei are now trying to soften ahead of their IPOs. Same week. Not because they are coordinating — but because the economic logic in every direction pulls toward calm messages for politicians and investors.

That is where the serious political discussion needs to land. Not the counterargument "you are wrong" — but the additional question "and what is the plan for what you do not mention?"

What the government actually needs

A cross-bloc, long-term plan for entrepreneurship and AI is reasonable. Such a plan without parallel planning for the broad labour-market shift that AI companies themselves trigger is not long-term, however — it is wishful thinking.

What needs to be included, beyond talent visas and employee stock options:

That the question is no longer extreme is visible in who is asking it. Anton Korinek — economist at the University of Virginia, senior fellow at Brookings and PIIE, and a member of Anthropic's own Economic Advisory Council — stated (Vox, Explain It to Me, December 2025) that when the AI revolution really breaks through there is "no guarantee that we can earn a decent living based on labor", and that we will then need a new system for income distribution. When one of Anthropic's own economic advisers says that — at a time when the company is seeking capital at valuations close to $1 trillion — it is no longer a marginal warning from the outside.

This is not a left-wing programme. It is risk management. And it is exactly what the tech founders' letter — and Altman's and Amodei's new optimism — lacks.

Finally

In the week when eight Swedish tech founders asked for long-termism, two of the world's leading AI CEOs walked back their warnings. Not because they changed their minds — but because capital-market logic rewards a calmer message ahead of an IPO. That is not a problem with their character. It is a problem with the information those of us planning Sweden's economy receive from them.

If Sweden builds its AI policy on forecasts that are turned as needed, it builds on unstable ground. The long-termism the founders ask for requires us to take their own forecasts seriously — including the forecasts they are currently walking back.

This is not a yes or a no to their agenda. It is a demand that it pass the consistency condition: your proposed measures and your descriptions of the future must fit together. When they do not, we cannot plan a country on them.

Rolf Skogling runs ai-skiftet.se — a Swedish voice on how AI is changing society, work and leadership.