The current narrative on the employment impact of AI is pretty dark. It goes like this: As AI gets more capable, companies need fewer people. Whole categories of jobs (though mostly white-collar) disappear. We’re already seeing mass layoffs attributed to (as yet mostly unrealized) AI productivity gains.
This is for sure causing some economic disruption and suffering, but the end game here is likely to be incredibly positive for those same categories of employment. The reason is that AI isn’t just replacing workers so much as it’s changing the economics of work itself. And that is very likely to produce a different outcome than people expect.
The Klarna and Amazon Reality Checks
Let’s start with two companies that went all in on AI and had to walk it back.
In February of 2024, Klarna announced that two-thirds of all customer service chats were being handled by agents. Later, in December of 2024, Klarna made the bold claim that it had stopped hiring humans in favor of AI and had reduced headcount 22%, mainly through attrition. Then in May 2025, Klarna reversed itself and began hiring humans for customer service again, saying essentially that cost had proven to be the wrong metric to prioritize, and quality of service had suffered. The pattern here is that humans will lose out to AI in the short term, but over the longer term, when cooler heads prevail, the higher productivity future will be humans working alongside AI.
Amazon is another example. Earlier this year, the company pushed internal teams to use AI coding tools more aggressively. It didn’t go well. Several outages forced it to mandate that senior engineers review AI-generated code before it is released to production. It’s not exactly a case of AI replacing engineers. It did change the nature of the engineering work, increasing the load on senior engineers. It’s not entirely clear what is going to happen to the junior engineering ranks. Since October 2025, the company has cut about 30,000 jobs, nearly 10% of its workforce, with both AI productivity improvements and AI infrastructure spending cited as the driving reasons.
There’s a similar pattern in systems that are supposed to be fully autonomous. For example, Waymo is often regarded as the gold standard for autonomy. But even Waymo still relies on human intervention to handle edge cases. And despite removing the driver, its cost structure hasn’t clearly beaten human-driven alternatives like Uber and Lyft. That’s not a failure of the technology, but rather an example of technology reaching its limit. AI excels at handling common-sense situations, but humans are still required to manage everything else.
When Efficiency Creates More Work
There’s another important dynamic here that often shows up in technology cycles. When something becomes cheaper, we don’t use less of it, we use more. This is called the Jevons Paradox. The short version is that when technology makes a resource cheaper to use, overall usage rises despite the cost savings on a per unit or per use basis; the per unit savings get overwhelmed by how much new demand the lower cost creates.
A version of this dynamic has played out in radiology over the last decade. In 2016, Geoffrey Hinton famously declared that we should stop training new radiologists because AI would soon be better than humans at reading images. More recently, he reversed himself, saying his prior comments were overly broad. In fact, radiology jobs are now projected to grow faster than other roles as a result of embracing AI.
However, today’s main angst about AI labor displacement lies in software engineering.
For the last five or so decades, the difficulty (and ultimately cost) of building and maintaining software has created a situation in which, for nearly all organizations, buying software built by a professional software development organization was more cost-effective in the long run than trying to build your own custom software even though it came with functionality trade-offs that might not be optimal for your business.
If building and maintaining software becomes easy enough, that trade-off will flip and companies will begin building their own custom software that’s tailored to their specific bespoke needs. This will result in dramatically more software being built — just not necessarily by professional software organizations.
The public markets recently seemed to discover this possibility and drew down public SaaS company multiples dramatically as a result. It remains to be seen if things will play out this dramatically. But if they do, it won’t necessarily mean the disappearance of software engineers. All that code needs to be maintained, monitored, secured and fixed when it breaks. Which means demand for technical work doesn’t disappear, it just becomes expressed in a different form. In other words, it might mean software engineers will mostly work inside enterprises instead of for software companies.
So Why Hasn’t Productivity Spiked?
By all measures, AI hasn’t meaningfully improved overall workforce productivity yet. But that’s not because AI isn’t useful. Simply plugging AI into existing workflows doesn’t automatically make them better. The earlier Amazon example shows that.
A historical precedent that might be at play is another famous paradox…the productivity paradox, sometimes called the Solow paradox. The short version of this is that it takes time for companies to properly figure out how to leverage technology changes. It took decades for regular (i.e., non-AI) IT technologies to show up in the productivity data at the macro economy level. We’re probably in the early stages of that same dynamic for AI.
The Job Destruction Narrative Is Too Simple
To be clear, it’s no secret that people are losing jobs in this economy. But attributing that exclusively to AI is a stretch. A lot of companies that are cutting headcount expanded aggressively during the pandemic. Headcount grew faster than revenue. What’s happening now often looks like a correction that reflects current economic reality. AI may be part of the reason, or the excuse for layoffs, but it’s not the sole root cause.
Even in cases where AI is clearly involved, like the Amazon cuts, the story is more complicated. The company isn’t just reducing labor. It’s also increasing capital expenditure on AI infrastructure. And furthermore, most of Amazon’s labor cost is comprised of warehouse and logistics workers — in other words, not white collar jobs — so the corporate workforce layoffs, while significant in terms of total numbers, are a drop in the bucket when compared to their overall workforce numbers.
Looking at this dynamic graphically can help. Comparing headcount to gross profit for Amazon shows a pattern of over-hiring during the pandemic, then flattening out headcount growth until gross profit catches up.
Amazon Gross Profit vs Headcount 2019 Through 2026E.
Applying the same analysis to Klara is also illustrative. The same pandemic over-hiring occurred, but instead of flattening out, the company reduced headcount through 2022-2024, ended up very slightly below where it started and has started growing headcount again. This implies an over-hiring correction, followed by reconfiguration of human vs. AI work and a return to more profitable, higher leverage scaling.
Klarna Gross Profit vs Headcount 2019 Through 2025.
Finally, Block (NYSE: XYZ) is also currently in the news as an example of AI job losses, having announced a layoff of 40% of its workforce in its Q4 2025 earnings call. CEO Jack Dorsey said, “intelligence tools have changed what it means to build and run a company.”
Block Gross Profit vs Headcount 2019 through 2026E
Looking at this visualization shows a familiar pattern: Dramatic pandemic over-hiring, followed by a drawdown back to — or maybe just below — linear growth vs. gross profit. What that’s followed by in 2026, projections of a steep drawdown in labor vs. gross profit, is a story in progress.
It remains to be seen how it ends.
The Bust Is In Process, But The Boom Is Likely To Follow
The general consensus is that AI is a headcount reduction tool. This is the most obvious effect, so it gets the most attention. But historically, the bigger impact of technology isn’t that it reduces work, but that it changes the economics of the work, often eventually resulting in more demand, not less. This seems to be what’s in process with AI. It isn’t just making workers more efficient. It’s changing what’s economically viable to build, operate, and scale, which is going to end up with dramatically more demand in the long run.
