When AI Eats Its Own Customers

AI efficiency, demand deficit; cheaper labor, poorer customers; profits up, buyers down. The quest toward an optimal automation equilibrium isn’t trivial; it’s full of obstacles presented as coordination failures. AI that optimizes firms, but starves demand.

When Labor Becomes a Vanishing Customer Base

When every company cuts at once, who is left to buy? That’s the Layoff Trap, a theoretical economics paper (not an empirical estimate) of what is already happening in the real economy. AI can improve firm-level economics and still push the economy toward an inefficient outcome. Because each firm captures the labor-cost savings from automation while bearing only a slice of the spending loss that layoffs create, layoffs can make both workers and firm owners worse off.

The New Coordination Failure

Can competitive markets over-automate even when firms understand the danger?

Workers are also consumers. Their income loss matters for demand. And competing firms don’t fully internalize that loss, because any one firm only feels a fraction of the demand it destroys when it automates.

When markets are fragmented enough, worker spending matters enough, and replacement income is weak enough, firms automate beyond the efficient level. This doesn’t mean that there is an automation ceiling above which not only would we see diminishing returns, but harmful circumstances. In fact, if displaced workers end up in better-paid roles, so that reemployment more than replaces lost income , the economy can under-automate. The answer depends on whether AI mostly destroys income faster than it recreates it, or does the opposite.

The Layoff Trap

In the Layoff Trap paper, its authors argue that over-automation can be a pure efficiency problem, not just a fairness problem. They claim firms can automate so aggressively that they erode the customer base they still need, and they can do it rationally. That is a useful addition to the older task-based literature, such as Automation and New Tasks and The Wrong Kind Of AI?, which focus on displacement, reinstatement, and labor demand, and to Inefficient Automation, which finds excessive automation through slow reallocation and borrowing constraints.

When AI becomes more productive than humans, private automation rises further, but the cooperative benchmark does not move in the same way. The reason is a Red Queen dynamic: each firm sees a market-share gain from automating faster than rivals, but when all firms do it, the gains cancel out. That means “better AI” can widen the strategic distortion rather than solve it. It is one of the Layoff Trap paper’s best insights.

Retraining, UBI, capital-income taxes, worker equity, and bargaining can help in some cases, but they do not directly change the marginal incentive to automate. However, a per-task automation tax could do that.

Sectors & Open Economy

The Layoff Trap comes with an important caveat: workers laid off in the sector spend less on the sector, and that feeds back into sector revenue. In a real economy, however, spending leaks across sectors, into imports, into housing, into healthcare, etc. Some AI-heavy firms also sell globally, which weakens the link between local layoffs and local demand.

A huge amount of the real world depends on one parameter: the share of displaced income that gets replaced. That parameter stands in for reemployment speed, retraining quality, household savings, access to credit, unemployment insurance, wage insurance, fiscal policy, and plain timing. But timing is the whole macro question. Indebted Demand shows why debt-burdened households matter for demand, and The Productivity J-Curve shows why the gains from general-purpose technologies often arrive with a lag. The layoff trap is most plausible when layoffs are fast and income recycling is slow.

Neat On Paper, Messy In Reality

What exactly is an automation tax base in 2026: model calls, software seats, API spend, headcount reductions, revenue per employee, or something else? Firms can relabel substitution as augmentation, shift work offshore, or fold automation into general software budgets.

At the same time, if displaced workers move into better-paying jobs or the transition raises labor income, the distortion flips from over-automation to under-automation, and the right correction becomes a subsidy, not a tax.

Before anyone can say “tax automation,” they need a credible sector-by-sector view of whether AI is mostly replacing income, recreating it, or doing both with a long lag in between.

Evidence is still mixed. These economic papers seek to identify a structural vulnerability rather than diagnose an active crisis, suggesting that we keep an eye on profit erosion along mass layoffs.

How The Data Pushes Back

Start with exposure and adoption. GPTs Are GPTs estimated that around 80% of the U.S. workforce could have at least 10% of tasks affected by large language models, and about 19% could see at least 50% of tasks affected. The Rapid Adoption Of Generative AI found that by late 2024 nearly 40% of U.S. adults ages 18 to 64 were using generative AI, with 23% of employed respondents using it for work in the prior week, and 9% using it every work day. The ECB has since argued that diffusion in Europe is also happening fast, with employee AI adoption in the euro area rising from 26% in 2024 to 40% in 2025.

Now the positive side. Generative AI At Work found meaningful productivity gains, with the biggest improvements going to less experienced and lower-skill workers. The same study showed treated workers with 2 months of tenure performing about as well as untreated workers with more than 6 months. That’s augmentation, not replacement.

Then the caution. Artificial Intelligence And The Labor Market finds strong substitution at the task level, but only modest overall employment effects on average because productivity gains at AI-adopting firms can offset losses in exposed jobs.

Canaries In The Coal Mine? finds a sharp relative employment decline for early-career workers in the most AI-exposed occupations. That makes the junior end of the market look fragile even if aggregate headcount is still holding up.

Large Language Models, Small Labor Market Effects finds precise null effects on earnings and recorded hours in Denmark 2 years after chatbot adoption, ruling out effects larger than 2% in that window. Statistics Canada finds that employment generally grew regardless of potential AI exposure from late 2022 to late 2025, though younger and less educated workers saw weaker growth and under-30 coding workers stagnated. The ECB likewise says the European evidence so far is mixed, with some employment-neutral findings and some large-firm reports of process optimization weighing on hiring.

Capital markets tell a different story though. Generative AI And Firm Values finds that firms with more labor exposure to generative AI saw stronger stock-market reactions after ChatGPT, and its “Artificial-Minus-Human” portfolio gained 5% in the 2 weeks after the launch. The ECB also notes that AI narratives have been a central driver of U.S. equity performance since early 2023. That doesn’t refute the layoff-trap mechanism. It does suggest that, so far, markets are pricing the first-round savings more heavily than any second-round demand damage.

There is one more layer. The IMF’s 2026 note on new jobs creation in the AI age says roughly 1 in 10 job postings in advanced economies now requires at least one new skill, often in AI or IT, with wage gains for those roles and rising polarization around them. That fits a transition story where AI is not simply destroying work, but shifting the skill mix faster than many workers, especially younger ones, can adjust.

Automation J’s Curve: Profits Now, Consumers Later

It all comes down to timing. The hard macro problem is not whether AI can do more tasks. It is whether the economy can recycle the resulting gains into purchasing power as fast as firms recycle tasks into software. If new tasks, cheaper prices, broader ownership, stronger transfers, and faster hiring elsewhere show up quickly, the layoff trap stays small. If substitution is immediate and recycling is delayed, the Layoff Trap becomes much more plausible.

That makes the current transition look less like a jobs apocalypse and more like a distributional J-curve. Productivity and valuations can move first. Hiring, wages, and career ladders can lag. Meanwhile, the early-career evidence tells us that the labor market can look “fine” in aggregate while quietly removing the bottom rungs that people use to enter software, design, customer support, analytics, and other white-collar tracks. If AI preserves senior output while thinning junior intake, the economy may not lose consumers all at once. It may lose training pipelines, bargaining power, and future middle-income demand in slow motion.

That’s where the danger from labor-saving AI is not only a wage-share story. It is also a market-structure story. Fragmented industries with many competing firms may be under the most pressure to automate fast, because each firm feels only a thin slice of the aggregate demand damage.

Sources and References

• AI And The Euro Area Economy
• Artificial Intelligence, Firm Growth, And Product Innovation
• Artificial Intelligence And The Labor Market
• Automation And New Tasks: How Technology Displaces And Reinstates Labor
• Bridging Skill Gaps For The Future: New Jobs Creation In The AI Age
• Canadian Employment Trends In The Era Of Generative Artificial Intelligence
• Canaries In The Coal Mine?
• Generative AI At Work
• Generative AI And Firm Values
• GPTs Are GPTs
• Indebted Demand
• Inefficient Automation
• Large Language Models, Small Labor Market Effects
• Robots, Trade, And Luddism: A Sufficient Statistic Approach To Optimal Technology Regulation
• Should Robots Be Taxed?
• The AI Layoff Trap
• The Productivity J-Curve
• The Productivity J-Curve and the Hidden Economics of AI Transformation, Simon Robinson, Medium, 20 February 2026
• The Rapid Adoption Of Generative AI
• The Simple Macroeconomics Of AI
• The Wrong Kind Of AI? Artificial Intelligence And The Future Of Labor Demand