The AI Layoff Wave Is Running Into Reality

Artificial intelligence was sold to boardrooms as a clean break with the old economy: fire the routine workers, install the software, and let the machines do the rest. Three years into the frenzy, the cleaner version of the story has failed. Companies that cut staff in the name of AI are finding that they have bought a costly illusion, then hired the same people back at higher pay.

The Great Correction Begins

The first phase of the AI boom had the feel of a corporate stampede. Chief executives announced that the age of human bottlenecks was over. Customer service, data entry, basic coding, order taking, and a long list of back-office tasks were supposed to be absorbed by machines that never tired, never complained, and never asked for a raise. Investors liked the sound of it. Boards liked the cost savings. Managers liked the promise of control.

What they have discovered is less glamorous. The systems work well when the task is narrow, repetitive, and easy to score. They stumble when the work depends on judgment, context, memory, persuasion, or common sense. That failure has produced a corporate boomerang: roles eliminated in haste are being recreated after the damage becomes visible.

The numbers in the source material are hard to dismiss. In a 2026 Robert Half survey of 2,000 US hiring managers, 32% of organisations that removed roles because of AI later rehired for the same jobs. Forrester’s 2026 Future of Work Report said 55% of employers that replaced workers with AI now regret the decision. It also found that 73% of organisations that made AI-driven cuts did not end up ahead financially. The lesson is not subtle. Automation did not remove the need for labour; it redistributed the work, then exposed how much of the job had been invisible to executives in the first place.

That should have been obvious. The people most eager to replace workers with software tended to describe employment as though it were a stack of interchangeable chores. It is not. A real job contains routine work, yes, but it also contains exceptions, negotiations, and the little repairs that keep systems from breaking. Those are the parts that do not show up neatly in spreadsheets, and they are precisely the parts AI struggles to imitate.

The Companies That Fired Too Soon

Klarna became one of the loudest examples of the new orthodoxy. The company’s chief executive announced that AI had replaced the work of 700 customer service representatives. The statement did what such announcements are meant to do: it reassured investors, impressed the press, and made competitors nervous.

Then the practical problem arrived. AI can handle the simple cases. It can reset passwords, track orders, and answer questions that resemble one another. It does not do well when the caller is angry, confused, or dealing with a problem that does not fit the script. According to the source material, the company soon began rehiring, because the machine could manage the easy 60% but not the high-value remainder that determined whether a customer stayed or left. That is the trap. Companies confuse volume with value and then discover that the most valuable part of the job was the part they could not automate.

McDonald’s provided a more comic but equally instructive example. The chain tested AI order-taking at 100 drive-throughs in the United States. The pitch was classic corporate technology theatre: faster service, lower costs, fewer human errors. Instead the system became an object lesson in what happens when a flashy tool meets the messiness of ordinary life. Video clips circulated of the bot adding hundreds of dollars’ worth of chicken nuggets to a simple order and mishandling basic requests. The company eventually shut down the programme and returned to people.

Ford’s case was less viral and more serious. The company cut 5,300 salaried positions while rolling out 900 AI-powered cameras to inspect manufacturing lines for defects. The assumption was that a machine could detect quality problems with the same reliability as experienced staff, and that data alone could replace practical knowledge. It did not work that way. Charles Poon, Ford’s vice-president of vehicle hardware engineering, later admitted that the company had assumed the system would deliver quality simply because it introduced AI and adjusted the rules. That is not engineering. It is wishful thinking dressed up as strategy.

Ford’s reversal was not small. The company rehired 350 veteran engineers, brought back weekly design reviews, and added more inspections and technical tests at the Kentucky truck plant for the 2026 Expedition. The result mattered. By 2026, Ford had moved from 15th place in J.D. Power’s Initial Quality Study in 2023 to first place, with 41 fewer problems per 100 vehicles. It beat Toyota and Honda. The win was not that Ford became more automated. The win was that it rediscovered the value of human scrutiny after technology had promised to make it unnecessary.

IBM found a different version of the same problem. The company let AI handle 94% of routine HR requests, but the remaining 6% included ethical dilemmas, policy judgment, and genuinely awkward human situations. Those are not edge cases; they are part of the job. IBM’s chief human resources officer, Nick LaMarr, identified the deeper danger: if companies stop hiring entry-level workers because software now does the routine work, they destroy the pipeline that produces senior people later. In three to five years, the company has no experienced employees to promote when judgment is needed. That is not efficiency. It is self-sabotage.

What History Says About Machines and Work

The panic around AI is only the latest chapter in a long, forgetful history. Each generation convinces itself that this time is different. Each generation turns out to be partly right and mostly wrong.

The printing press did not eliminate scribes so much as create publishing, journalism, and mass literacy. Photography did not kill painting; it expanded the market for visual work. The spreadsheet did not destroy accounting; it transformed it into a more analytical profession. The ATM was supposed to wipe out bank tellers. In fact, United States bank teller employment doubled over five decades after Barclays introduced the first automated teller machine in 1967. Once humans were freed from the simplest cash transactions, banks used them for service, advice, and complex interactions that machines could not manage.

This is the pattern economists have described for years. Automation usually does not erase work; it changes where the work is. Machines take over execution. Humans move into oversight, adaptation, and trust-building. The job becomes more abstract and often more valuable.

There is, however, a line that matters. Some roles disappear completely when the core task offers no room for judgment. Elevator operators were one such case. Once automatic systems arrived, there was no adjacent work to migrate into. The job died. The Census last recorded elevator operators in 1960.

Pilots tell the other story. Aircraft have had autopilot systems for generations, and modern planes can land themselves. Yet the number of pilots rose, not fell. Why? Because the task at the centre of the job is not moving the plane on a perfect day. It is deciding when to trust the automation, when to override it, and how to handle the abnormal event that no software trained on past data can fully anticipate. The same logic applies in law, medicine, engineering, manufacturing, and customer service. When the stakes are high and the environment is unpredictable, the human remains the anchor.

The source material also points to a less discussed economic force: non-consumption. Many jobs arise because technology makes something economically possible that used to be too expensive or too difficult to offer. A product, service, or process does not exist until the cost of producing it falls below what customers can pay. That is why new tools create new labour. A famous illustration is the old idea of a Maslow pyramid rendered as a physical object. The moulding cost alone would have been about $17,000, making the thing commercially ridiculous. New fabrication tools, including 3D printing and cheap design software, change that calculus. What once made no business sense becomes ordinary.

David Autor’s work is central here. The MIT economist has shown that a large share of current jobs did not exist half a century ago. That is not a warning about collapse. It is a reminder that labour markets are not static. They reorganise around what technology makes worth doing. AI may erase some tasks, but it also creates the possibility of work that is currently too tedious or too expensive to exist at scale. The danger is not that all work disappears. The danger is that executives mistake one kind of work for all work and fire people before the new arrangement has actually proved itself.

The New Math of “Efficiency”

One of the ugliest surprises of the AI era is that the bill keeps growing after the announcement posters come down. The original sales pitch was simple: software is cheaper than payroll. What has emerged instead is a system that can become more expensive than the labour it was meant to replace.

Modern AI is priced in consumption units. Every query costs something. Every image costs something. Every batch of code suggestions costs something. At small scale, the numbers seem harmless. At enterprise scale, they become punishing. A company that encourages thousands of employees to use AI all day, every day, can find that its “efficiency” programme now looks like a second payroll, only one with less predictable returns.

The source material describes a phenomenon called “token maxing,” in which workers use AI for tasks that do not need it, or run extra prompts to look productive. That is what happens when management turns software use into a metric. People optimise the metric, not the outcome. They generate more tokens because tokens are visible, while actual usefulness becomes secondary. The result is a theatrical economy: more apparent activity, more invoices, less value.

The cost problem is not abstract. Average token prices have risen sharply since the beginning of 2025. AT&T reportedly consumes nearly 8 billion tokens a day. Microsoft has cancelled internal licences for an AI coding assistant. Uber is said to have exhausted its annual AI budget in four months. Nvidia, the company most likely to benefit from the boom, has warned its own developers to be careful with usage. When the vendor of the hardware tells customers to slow down, the market signal is loud enough.

There is a further twist. The benefits of AI often show up where they are easiest to measure, not where they matter most. In coding, the tool may look cheap because it speeds up a narrow task. In customer service, the price can fall behind the human alternative on a spreadsheet while service quality quietly erodes. In data entry, the software may reach parity with labour, but parity is not victory when the machine still needs correction.

That is why some large companies have already begun to reverse course. The billing model is unforgiving. The more use there is, the more the cost rises. And unlike a worker, a model can be cheap at first and expensive later once the company has locked itself into the workflow. What looked like a one-time saving becomes a recurring charge.

Even the broad projections show strain. Gartner’s survey, as described in the source material, found that three-quarters of executives expected technology budgets to rise this year, with nearly half expecting double-digit increases. Enterprise AI spending is projected to hit $680 billion by 2027. That figure is not a sign of confidence so much as a sign of dependence. Once a firm reorganises around the technology, it must keep paying for it.

The deeper problem is that the economics are moving in the wrong direction for the suppliers. As models improve, users often need fewer queries and less computing power to get the same result. That should be good news for customers. It is terrible news for vendors whose revenue depends on high-volume consumption. A better product can mean a smaller bill. That is not a stable foundation for a trillion-dollar industry.

The Trillion-Dollar Bet Nobody Can Explain

The scale of investment is what makes the AI story more than a workplace dispute. It has become a market-wide financial gamble. The source material says five of the world’s largest technology companies have committed more than $1 trillion to AI infrastructure. That amount of money can support a long illusion.

The problem is that no one has yet shown a convincing path to returns that justify the spend. Microsoft puts money into OpenAI, which in turn spends heavily on Microsoft cloud services. The accounting produces revenue for one side and demand for the other, but it is the same money moving through the system. That is not the same thing as a durable business model. It is a loop.

The broader financing structure is even more fragile. OpenAI and Anthropic have committed to enormous computing purchases. Oracle is borrowing roughly $100 billion to build data centres for OpenAI, even as OpenAI burns cash and has no clear path to profitability. Much of the build-out is now being funded through debt and private credit rather than through operating cash flow. That matters because debt has a way of becoming a problem before the technology matures.

The Bank for International Settlements, often treated as a sober voice in a noisy market, has warned that disappointment in returns could trigger a sudden pullback in financing and turn the spending boom into a bust. The organisation’s concern is not academic. If investors decide that the returns do not match the build-out, the financing channels can seize up quickly. The BIS also warned that a major correction could have larger macroeconomic consequences than in the past because technology stocks are now woven deeply into retirement funds, index funds, and pension portfolios.

The comparison to past bubbles is uncomfortable. The dot-com crash destroyed roughly $5 trillion in market value. The financial crisis of 2008 erased more than $8 trillion in global wealth. AI is larger than either in capital intensity and more integrated into the financial system than either was at the peak of their manias. If the market turns, the damage will not be confined to Silicon Valley.

J.P. Morgan’s estimate in the source material is a useful reality check. The industry would need to generate about $650 billion in new annual revenue just to earn a 10% return on the infrastructure already being built. That is an enormous hurdle for a sector still subsidising users in the hope of capturing market share. The price cuts that attract customers also deepen the losses. Providers are effectively betting that scale will arrive before patience runs out.

There is no certainty that it will. The current system has several weaknesses at once: models are widely available, switching costs are low, open-source alternatives are improving, and no provider has yet secured the kind of monopoly power that would justify the valuations. The infrastructure itself may also age quickly. Data centres are expensive to build and chips go obsolete fast. The railway barons built assets that lasted for generations. AI firms are building ones that may be stale before they are fully switched on.

Michael Burry has already made his view plain through bearish positions against Nvidia, Tesla, Palantir, and Caterpillar, which benefits from the construction spree. His bet rests on a simple suspicion: that the money flowing into AI is being recycled through the same ecosystem, while depreciation assumptions are being stretched and earnings made to look better than they are. In plain English, the market may be mistaking financial engineering for industrial progress.

The Workers Paying the Price

All of this would be merely a market story if people were not losing work along the way. They are.

The first half of 2026 saw more than 150,000 jobs cut with AI named as a factor, according to the source material. Atlassian eliminated 1,600 positions. Cloudflare cut 1,100 even while reporting record revenue. Block reduced its headcount from 10,000 to fewer than 6,000. Cisco cut 4,000. Citigroup is targeting 20,000 reductions. The BBC said it would cut as many as 2,000 jobs, or 10% of its workforce.

These are not isolated mistakes. They are the visible edge of a wider change in how companies think about labour. AI gives management a story that sounds modern and inevitable. A layoff once had to be justified as a response to demand, profit, or restructuring. Now it can be sold as a step into the future. That matters because it lowers resistance. If the machine is coming, why argue?

The people most exposed are not the most senior. A Stanford preprint from November 2025 found that early-career workers aged 22 to 25 in AI-exposed occupations saw a 16% relative decline in employment. That is exactly where the damage should worry employers most. Entry-level work is not a waste product. It is how organisations produce their future managers, specialists, and supervisors. Remove the bottom rung and you eventually weaken the entire ladder.

Some workers who were displaced were rehired as consultants overseeing the systems that replaced them. That can be humiliating and practical at the same time: the firm still needs the knowledge, but it wants it on a cheaper, more disposable basis. Others were simply gone. Expertise that took years to build was treated as expendable, only to be found missing when the software failed to cope.

The remaining employees are not having a good time either. Once companies start measuring AI use, they create pressure to use it all the time. Workers report being judged on token consumption rather than results. Some are prompted to run AI over everything, regardless of whether the task benefits. That converts a tool into a surveillance mechanism. It also creates resentment, because people know when they are being asked to pretend.

The most striking accounts in the source material are the ones that show how quickly management can lose its bearings. One lawyer described a boss who insisted that all ideas be discussed with AI before meetings, and who made staffing decisions based on what the model said. A sales strategist who brought real customer conversations to his founder was told the customers must be wrong because Claude and ChatGPT disagreed with them. The strategist left.

That is the deeper risk. AI can give people a surface of confidence without the discipline of friction. It answers quickly. It sounds fluent. It rarely pushes back. For a manager already inclined to confirm a belief, the machine becomes a mirror. The danger is not that it thinks too much. The danger is that it thinks too little, but speaks too well.

What AI Can Do, and What It Still Cannot

The honest case for AI is narrower than the marketing would have us believe. It is useful. In some places, it is very useful. It can accelerate coding. It can summarise large amounts of text. It can improve service speed in limited settings. It can reduce the time needed to find an issue or draft a response. The source material cites a 2026 field experiment at Alibaba showing gains in speed and some improvements in subjective satisfaction.

But speed is not the same as quality. A system that gets to the answer more quickly is not necessarily a system that gets to the right answer, or the whole answer, or the answer that preserves trust. A fast chatbot that solves the easy problem and leaves the user frustrated is not a substitute for a competent employee. It is a faster way to discover the limits of automation.

That distinction matters because business leaders often reward what they can see immediately. Reducing chat duration looks good. Cutting response times looks good. Shrinking headcount looks excellent. These are shallow victories if they come at the cost of repeat complaints, quality failures, missed context, and hidden damage that shows up later in returns, warranty claims, churn, or reputation.

The source material suggests that the real value of human workers lies in the 40% of a job that AI does not handle well, which can represent 90% of the economic value. That is an uncomfortable claim for executives who want to reduce labour to a line item, but it matches how organisations actually function. The trivial part of the job is rarely the valuable part. The valuable part is the part that keeps the business from embarrassing itself.

This is why some of the strongest cases for AI are not about replacement but augmentation. Let the software do the first pass. Let it surface possibilities. Let it remove dull work where the stakes are low. But keep the person who can judge, correct, and absorb the abnormal case. The moment a company decides that a machine’s fluency is a substitute for human responsibility, it crosses from efficiency into carelessness.

That line is now being tested across industries. In law, software can find relevant material, but it cannot determine the strategy. In manufacturing, cameras can detect patterns, but they do not understand the cost of a defect. In customer service, bots can answer simple questions, but they do not defuse anger or rebuild trust. In hiring, models can rank candidates, but they cannot substitute for judgment about character, fit, and potential.

The broader truth is that AI has exposed how much corporate work was already bad. It became popular in part because firms wanted a tool that would excuse them from hiring, training, and patience. That was always a tempting fantasy. It is now an expensive one.

The Myth Is Fading, the Bill Remains

The AI story is not collapsing because the technology is worthless. It is fading because the claims made for it were too large, too fast, and too careless with the facts. The companies that rushed to replace workers discovered that routine tasks are only a fraction of the enterprise. The companies that bought into the promise of cheap automation discovered that the service bill can rival the payroll it was supposed to replace. The investors who treated the boom as a straight line to profits are beginning to meet the older truth of markets: capacity is not the same as demand, and hype is not a business model.

There is still a place for AI, and there probably will be for a long time. But the boast that it would hollow out the labour market has already collided with reality. What remains is more complicated and less flattering to the people who sold the fantasy. Human judgment has not been removed from the system. It has become more valuable.

The real lesson of the last three years is not that machines failed to matter. It is that the people making decisions about them mistook demonstration for replacement, and novelty for discipline. That mistake has cost companies money, workers jobs, and investors patience. The bill is still being written.