The AI Layoff Trap: How Companies Can Fire Their Own Customers
Acronyms used: AI means Artificial Intelligence, software that performs tasks normally requiring human judgment or cognition. UBI means Universal Basic Income, a recurring cash payment given to people without requiring employment. CEO means Chief Executive Officer, the top executive responsible for running a company. SaaS means Software as a Service, software delivered online by subscription. GDP means Gross Domestic Product, the total value of goods and services produced in an economy.
The simplest way to understand the AI layoff problem is this: a company fires workers to save money, then discovers, rather late in the evening, that workers were also customers.
That is the small bomb inside The AI Layoff Trap, the paper by Brett Hemenway Falk and Gerry Tsoukalas. It is not saying the machines are evil. It is not saying every CEO wakes up, oils his horns, and plans the ruin of clerks, coders, designers, analysts, call-center workers, and all the other office creatures who live by keyboard, caffeine, and hope. The paper’s sharper claim is worse because it does not require villains. It requires only normal firms doing normal firm things: cutting costs, adopting AI, pleasing investors, reducing headcount, and calling the whole business “strategic transformation,” as if the English language itself had been laid off and replaced by a damp brochure.
The trap begins with a fact so obvious that business people often manage not to see it. Wages are costs to one firm, but income to the economy. Inside the company spreadsheet, the worker appears as expense. Outside the spreadsheet, the same worker is rent paid, food bought, fees paid, phones repaired, medicines purchased, school uniforms replaced, shoes bought, small pleasures taken, and debt kept from becoming a tiger at the door.
So when one company automates a job, the company keeps the saving. But the loss of that worker’s spending is scattered everywhere. The fired worker buys less from many businesses, not only the business that fired him. The pain leaks into the market like water from an upstairs flat. Everyone downstairs gets wet, but nobody agrees who should fix the pipe.
That leak is called a demand externality. The name sounds like something that should be buried under committee minutes, but the idea is plain. A firm does not bear the full cost of the demand it destroys. It gets the full benefit of reducing payroll, but only a small share of the future sales loss caused by weaker consumer spending. The private benefit is immediate and countable. The social damage is delayed, spread out, and easy to pretend is someone else’s weather.
This is why the paper is interesting. It does not merely repeat the familiar slogan that AI will take jobs. That has been shouted so often now it has become background noise, like a ceiling fan with a bent blade. The paper asks what happens when many firms automate at the same time, faster than the economy can create new high-paying work. The answer is not just unemployment. The answer is a feedback loop.
Workers lose jobs. Their spending falls. Firms face weaker demand. Firms cut more costs. AI makes more cost-cutting possible. More jobs go. More demand disappears. The machine becomes better at producing, while society becomes worse at buying. At the extreme, you get a magnificent bakery making endless bread for people who no longer have money for breakfast.
That is the gloomy joke. Productivity can rise while the market underneath it thins out. A company can become lean, efficient, automated, cloud-powered, AI-native, consultant-approved, and financially doomed because its customers have quietly been removed from the story.
The model in the paper is built around tasks. Firms decide how much work to automate. Some tasks are easy to automate, others harder. This matters. AI does not arrive like a magic servant and do “jobs.” It attacks tasks. Draft this email. Classify this ticket. Summarize this call. Generate this code. Fill this form. Answer this customer. Prepare this report. One by one the tasks are shaved away, and then the job that contained them begins to look like a banana peel after a monkey conference.
At first, automation looks wonderfully sensible. Why keep paying people to do work a system can do cheaper, faster, and without complaining that the office chair is shaped like a punishment device? If one firm automates and its competitors do not, it gains an advantage. Lower cost. Better margin. Maybe lower prices. Maybe higher profit. Maybe both, if the finance department is in a festive mood.
But when everyone automates, the advantage shrinks. The cost reductions spread across the industry. The market-share gain disappears because everyone is chasing the same gain. What remains is the damage to income and spending. Each firm did the rational thing alone. Together they may have built a trapdoor under their own market.
This is the prisoner’s dilemma of automation. If all firms could somehow agree to slow down, they might all be better off. But no single firm wants to be the noble fool standing in the rain holding an umbrella for civilization while its competitors run indoors with the money. The first firm to restrain itself risks losing to the firm that automates aggressively. So the individually smart move remains automation, even when the collective result is stupid.
This is where the paper becomes less like science fiction and more like a neighborhood quarrel. Everyone knows that dumping garbage into the common pond is bad. But if cleaning up costs one person money and the dirty water is shared by all, each person has an incentive to look innocent and carry on. Soon the pond smells like history.
UBI enters the discussion as a cushion, not a cure. It can keep people from falling through the floor, and that is no small thing. People who sneer at income support usually have very firm opinions from very soft chairs. But the paper’s point is that UBI does not automatically change the firm’s automation decision. The company still asks: can I save money by replacing this task? If yes, it proceeds. UBI may preserve some spending, but it does not by itself stop the private incentive to over-automate.
Upskilling has the same problem when spoken of as magic. Training is useful when real jobs exist at the other end. Without those jobs, it becomes motivational packaging for economic displacement. You can teach a man cloud architecture, prompt engineering, data analytics, and the correct tone for LinkedIn optimism, but if the labor market is producing fewer good seats than trained applicants, then the training program is not a bridge. It is a waiting room with certificates.
This is not an argument against learning. Learning is oxygen. But oxygen does not solve a broken staircase.
The paper also discusses an automation tax, the Pigouvian idea that a firm should pay for the social cost of the demand it destroys when it automates tasks. In theory this is neat. In practice it is a goat wearing spectacles. How do you define an automated task? Is a chatbot replacing five support agents automation? Yes. Is a coder using AI to do twice as much work automation? Maybe. Is a reporting tool that lets one manager replace three analysts automation? Very likely, but good luck getting the company to call it that. It will be called “workflow modernization,” “AI enablement,” or “enterprise productivity uplift,” because modern business has discovered that if you give a thing a soft enough name, people may not notice the teeth.
Still, messy measurement does not mean the problem is fake. It means the problem has entered the zone where policy becomes difficult, industry becomes evasive, and consultants begin breeding in the walls.
The more useful lesson is not that we should tax every bot tomorrow morning. The useful lesson is that layoffs cannot be treated only as firm-level efficiency. When one company reduces headcount and keeps output stable, that may look like progress. When hundreds of companies do it together, especially in high-income sectors, it may become demand cannibalization. The same number means different things at different scales. A bucket of water is useful in the kitchen. The same bucket poured into your laptop is a family emergency.
This is also why so many AI discussions go wrong. They measure output but ignore purchasing power. They count productivity but not income replacement. They celebrate cost savings but do not ask who now has less money to spend. They look at the factory and forget the bazaar.
The video mentions tech layoffs, gaming, consulting, SaaS, and big firms moving money toward AI infrastructure. That detail matters because AI is not free. These systems need chips, data centers, electricity, cloud contracts, specialized engineers, model access, integration teams, security controls, and the usual managerial carnival around anything expensive. So a firm may cut human labor partly to fund machine labor. People imagine AI as a cheap invisible brain floating in the sky. In reality it is a power-hungry industrial apparatus in expensive clothes.
There is something darkly funny about this. The worker is told he is too costly. Then the company spends heavily on the infrastructure required to replace him. The machine does not demand a salary, but it does demand capital, energy, cooling, vendors, maintenance, and a priesthood of people who understand why the clever thing hallucinated a refund policy from 2018. It is not laborlessness. It is a different kind of cost, with better public relations.
The paper’s strongest point is that the danger can emerge without anyone making a mistake. That is what makes it structural. If a foolish company automates badly and fails, the market can punish it. But if many sensible companies automate well and weaken demand together, market punishment arrives as a fog. Each firm sees softer sales, margin pressure, and uncertainty. Each responds by cutting more. Nobody caused the storm alone, so nobody feels responsible for the rain.
This is how systems fail. Not usually with one dramatic villain pressing a red button. More often with many well-designed incentives pointing slightly in the wrong direction until the whole machine walks into a ditch.
The India angle is not decorative here. India has millions of people who have treated education, English, coding, support work, testing, analytics, back-office processing, and global services as a ladder out of family panic. Not paradise. Not luxury. A ladder. Sometimes shaky, sometimes humiliating, sometimes held together with exam coaching, parental sacrifice, and the stubborn belief that if you learn enough, type fast enough, speak politely enough, and swallow enough office nonsense, life may move one notch upward.
If AI compresses white-collar service work, that ladder gets crowded at the bottom and chopped at the top. The person in Bengaluru, Kolkata, Pune, Hyderabad, Manila, Warsaw, Nairobi, or Ohio is not living inside a research model. He is living inside rent, loan payments, aging parents, medical bills, school fees, and the quiet terror of becoming obsolete before becoming secure.
This is why “new jobs will come” is not enough. When? Where? At what wage? For whom? With what bargaining power? A new job paying half the old wage is not full recovery. A gig assignment is not a career. A training badge is not demand. A country cannot run forever on motivational webinars and delivery apps.
The paper does not prove that AI will cause collapse. It does something more modest and more valuable. It identifies a mechanism by which rational automation can become excessive. The question is empirical now. Are displaced workers finding comparable income quickly? Are AI-heavy firms expanding total demand or merely cutting payroll? Are productivity gains flowing into broad spending or concentrating as capital income? Are layoffs temporary correction after overhiring, or the early pattern of permanent task substitution? Are firms becoming more profitable, or merely more automated in a market that is quietly losing buyers?
Those are the questions worth asking. Not whether AI is good or bad, which is a kindergarten question wearing a doctoral gown. AI is a tool, an industry, a capital project, a labor substitute, a productivity engine, a coordination problem, and a political problem all at once. Asking whether it is good or bad is like asking whether electricity is polite.
The practical direction is boring, which usually means it is close to reality. Measure automation-linked displacement. Track income replacement, not just employment. Separate genuine worker augmentation from role elimination. Stop letting companies hide layoffs behind perfume words. Tie public training money to actual labor demand. Study whether AI adoption in one sector weakens demand in connected sectors. Consider taxation or insurance mechanisms where displacement is measurable and broad. Do not pretend this will be clean. Clean solutions belong to textbooks, pitch decks, and people who have never had to fix anything after deployment.
The central point remains plain. A wage is not just a cost. It is also someone else’s revenue. Cut enough wages across enough firms and the economy does not become a sleek machine. It becomes a restaurant where the kitchen is modern, the staff is gone, the menu is optimized, and the customers are outside looking through the glass because nobody has money for dinner.
That is the AI layoff trap. Not killer robots. Not lazy workers. Not evil executives twirling mustaches. Just a market clever enough to automate tasks and foolish enough to forget that the people doing those tasks were also holding up the market with their monthly spending.
A company can fire a worker.
An economy, if it is not careful, can fire its customers.
P.S. References: Brett Hemenway Falk and Gerry Tsoukalas, “The AI Layoff Trap,” arXiv, 2026: https://arxiv.org/abs/2603.20617; Daron Acemoglu and Pascual Restrepo, “Automation and New Tasks: How Technology Displaces and Reinstates Labor,” Journal of Economic Perspectives, 2019; Tyna Eloundou, Sam Manning, Pamela Mishkin, Daniel Rock, “GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models,” Science, 2024.