The AI harvest is beginning

In March 2026, artificial intelligence became the leading reason for job cuts in the United States for the first time. Of the 60,620 cuts announced that month, one in four were attributed to AI, according to outplacement firm Challenger, Gray & Christmas. 

For the last few years, OpenAI, Google DeepMind and Anthropic have been spending hundreds of billions on AI infrastructure. For a while, AI felt like a fun gimmick. Then it could search the web. Then it started learning about you and retaining memory over time. Analysts called it a money pit. The world’s largest technology companies have committed roughly $700 billion in combined capital expenditure for 2026 alone, nearly doubling their 2025 outlays. The returns are arriving. And so are the layoffs. 

The AI harvest has begun.

It will be disinflationary, profitable for companies, and painful for workers whose roles can be codified. 

The productivity J-curve 

Erik Brynjolfsson, director of Stanford’s Digital Economy Lab, describes this pattern as the productivity J-curve. General-purpose technologies, from the steam engine to the computer, do not deliver immediate gains. They need years of massive, often unmeasured investment. This includes reorganising business processes, retraining workers, developing new business models. During this phase, measured productivity is suppressed as resources are poured into investments that have not yet paid off. 

As Brynjolfsson wrote in the Financial Times in February 2026, the updated US data suggests we are now transitioning out of this investment phase and into a harvest phase. Revised payroll figures show that total US job growth in 2025 was approximately 403,000 lower than initially reported.  

US productivity grew 2.5% in 2025, nearly double the 1.4% annual average of the past 15 years. The earlier investments in AI infrastructure and process redesign are beginning to show measurable output.  

The cuts have started 

The evidence is already in the labour market. The technology sector announced 52,050 job cuts in Q1 2026, up 40% year-on-year. Since Challenger began tracking AI as a cause in 2023, nearly 100,000 cuts have been attributed to it, with the pace accelerating: 3.5% of all cuts in 2023, 5% in 2025, 13% in Q1 2026.

A significant example came in February 2026, when Jack Dorsey cut over 4,000 workers from Block, roughly 40% of the company’s workforce, and directly attributed it to AI.  

He was not alone. Amazon cut 14,000 roles in October 2025 with nearly 47% of these being engineering roles. And Atlassian cutting 1,600 roles in March 2026.  

But headlines only capture the most visible cuts. Stanford’s Enterprise AI Playbook, which studied 51 successful deployments across 41 organisations, found that 45% led to direct headcount reduction. The researchers warn this may represent “a floor, not a ceiling.”

Agentic AI, where systems autonomously execute multi-step tasks, currently represents only 20% of deployments. But it delivered median productivity gains of 71% when used. Most companies haven’t deployed it yet. When they do, the floor rises.

The tools have changed 

I use AI tools four to five hours a day. I have done so since late 2022, upon the release of the original ChatGPT 3.5. In those early days, AI tools were useful but unreliable: prone to hallucinations, unable to remember previous conversations, confined to a single chat window.

That has changed materially. The major models now search the web, retain memory of who you are, and reference past conversations. The compound effect of that continuity on output quality is difficult to overstate: context from previous problem-solving feeds directly into stronger work today.

I have used AI to build and ship a full custom website for a professional services firm (theme, responsive design), a project that would have required hiring a developer 18 months earlier. I use it to draft analytical writing, process data, and build the charts in this article. Anthropic’s Cowork agent, launched in January 2026, autonomously executes multi-step tasks across files and applications. Tasks that used to take hours compress into minutes.

Brynjolfsson’s own research suggests most businesses still use AI for translation or summarisation, what he calls “glorified dictionary” use. But a small cohort of power users, and I count myself among them, are automating entire workstreams, compressing weeks of effort into hours. As more workers move from the first category to the second, the productivity gains will compound. 

This time might actually be different

Every previous technological revolution eventually created more employment than it destroyed. I do not believe that pattern will hold this time, at least not fast enough for the people being displaced now.

There is a fundamental distinction between what prior technologies replaced and what AI replaces. The steam engine, the loom, the tractor: these amplified our muscles. One worker with a power loom could produce what previously took dozens by hand. The work was still physical, but each person’s effort went vastly further. And in doing so, these technologies freed humans to move into cognitive work.

When computers and offshoring shrank manufacturing from the 1970s onward, displaced workers moved into service jobs that computers themselves had created: financial analysts, IT support, logistics coordinators. The personal computer alone enabled 15.8 million net new jobs in the US since 1980, according to McKinsey. There was somewhere to go.

AI replaces our intelligence. It automates the cognitive work that was supposed to be the safe harbour after the machines took the factories.

Anthropic’s own research, based on millions of real-world Claude conversations, illustrates the scale of what is coming. In computer and mathematical occupations, AI can theoretically handle 94% of tasks. And business, finance and management jobs are similarly exposed. Observed usage today is a fraction of that. But the gap is not a ceiling. It is a lag. The occupations with the lowest exposure are overwhelmingly physical: construction, agriculture, food service. The cognitive professions that were supposed to be the destination after the factory floor are now the most exposed.

This is already visible in the data. Research from Brynjolfsson, Chandar, and Chen found that early-career workers aged 22 to 25 in the most AI-exposed occupations have experienced a relative employment decline of 13 to 16% since late 2022, with young software developers hit hardest at nearly 20%.

Prior technologies destroyed routine tasks but left entry-level cognitive work intact as a training ground. A junior analyst or graduate accountant didn’t just perform calculations. They learned judgment by doing the grunt work. AI removes that training ground entirely. The first rung of the career ladder is being pulled away before a generation can climb it.

AI will create new categories of work. But unlike every prior revolution, it does not open a new domain of human capability that I can easily see. It compresses all cognitive domains simultaneously. And unlike offshoring, you cannot reverse it with tariffs or trade policy. The cheaper labour is software, and it is not going away.

What this means 

If productivity growth is sustainably higher, this is disinflationary. More output with fewer workers means lower unit costs, downward pressure on prices, and eventually on interest rates. The building phase of AI infrastructure and current geopolitical tensions are mildly inflationary in the near term, but as deployment spreads, the net effect on prices should be downward.

For investors, disinflationary productivity growth means higher margins and stronger earnings. For workers entering the labour market now, the message is harder: the entry-level roles most exposed are precisely those that used to be the safe first rung on the career ladder.

The harvest is beginning. The question is who gets to eat.