AI Would Not Destroy Jobs by Mistake, but by Poorly Designed Incentives

Automation with artificial intelligence often presents itself as a technical decision: a company identifies a repetitive task, delegates it to a model or agent, reduces costs, and gains efficiency. But a new academic paper suggests a more uncomfortable interpretation for the tech sector: the problem isn’t that companies misuse AI, but that they use it exactly in ways that benefit them in the short term.

The study The AI Layoff Trap, authored by Brett Hemenway Falk from the University of Pennsylvania, and Gerry Tsoukalas from Boston University, models an economy transitioning toward AI-driven automation. Its main thesis is clear: when many companies replace workers with automated systems simultaneously, each captures the savings, but none fully accounts for the demand destruction caused by eliminating wages. This difference fuels a race to automate that can be rational for individual firms but harmful to the economy as a whole.

The failure isn’t in the model but in the market

Public discussions about AI and employment tend to focus on whether models are good enough to replace human tasks. The paper shifts the question. Even if AI works—improves productivity and cuts costs—the trap can occur if displaced workers don’t find new income quickly.

In this scenario, dismissed employees are not just a eliminated cost line. They are also consumers: buying products, paying for services, subscribing, traveling, financing mortgages, or consuming technology. When a single company lays off, demand impact is spread across the market. But when many do so simultaneously, the loss can become systemic.

This mechanism resembles an incentive bug. For an individual company, automation may be rational. For all firms doing it at once, it can be destructive if the economy doesn’t create new tasks and jobs at the same pace. The model shows that knowing this isn’t enough: companies still have incentives to automate because the cost of not doing so can mean losing margin to competitors.

Why agentic AI makes the problem more relevant

The thesis gains strength in the era of agentic AI. A traditional chatbot might assist with writing, summarizing, or answering. An agent can chain actions: reading repositories, executing tasks, coordinating tools, calling APIs, preparing documents, resolving tickets, reviewing contracts, or automating entire workflows. Replacement no longer involves just isolated tasks but entire sequences of work.

This scales the impact. When automation was slow, costly, and imperfect, the labor market had more time to absorb the shocks. With AI agents, deployment can be much faster. A SaaS provider can reduce support staff, a fintech can automate document review, a consulting firm can produce deliverables with fewer analysts, and a development team can deliver more with fewer junior profiles.

The paper doesn’t claim that all automation is bad. In fact, it acknowledges that historically technology has created new tasks and occupations. The difference lies in the pace. If job creation can’t keep up with displacement, productivity growth may come at the expense of weakening demand.

The corporate prisoner’s dilemma

The most powerful aspect of the model appears when automation becomes cheap and easy to integrate. It then resembles a prisoner’s dilemma. If all firms restrained automation, demand could be better preserved. But each has the incentive to automate first.

If one company chooses not to lay off and its competitors do, it risks cost disadvantages. If all lay off, no one gains a lasting structural advantage, but demand across the economy weakens. The outcome isn’t a clean transfer of workers to capital—in fact, the model suggests it could be a net loss where both workers and owners end up worse off than in a coordinated solution.

Competition—common in many markets for controlling prices and enhancing efficiency—can amplify this problem. The more companies compete, the smaller the demand share each internalizes. In a monopoly, the firm would bear the whole demand decline it causes. In a fragmented market, each firm might believe its contribution to the damage is small.

Why usual policies miss the mark

The study reviews several known responses: universal basic income, capital income taxes, profit-sharing, reskilling, corporate agreements, and private negotiations. It concludes that many can mitigate social damage but don’t address the root incentive to over-automate.

For example, basic income can raise income floors and sustain consumption but doesn’t alter the company’s calculation when deciding to replace a human task with AI. Profit taxes likewise may not affect that decision, as they target overall profit, not the specific margins of individual tasks.

The only instrument that closes the distortion in the model is a pigovian tax on automation: a levy per automated task equal to the share of demand the firm destroys and doesn’t assume. Its logic resembles an environmental tax—if private decisions generate external costs, public policy can require internalization.

The key practical challenge: measuring what constitutes automation

The proposal is conceptually elegant but difficult to implement. What counts as an automated task? Is an agent that assists an employee or one that replaces them? A tool that boosts productivity or one that eliminates an entire role? How do you measure lost demand if the worker finds another job after six months? What if a company automates in one country and sells in another?

These questions matter because AI-driven automation often doesn’t appear as a visible robot replacing a person. Instead, it manifests as software improvements, new features in SaaS, integrations with models, internal agents, or future hiring reductions. Tracking layoffs is easier than measuring roles never created.

Hence, the tax proposed in the paper should be seen more as a signaling tool rather than an immediate recipe. If companies don’t pay for the externality of demand they generate, they might automate beyond the optimal level. Specific policies could take different forms but must influence the decision margin, not just compensate for damages afterward.

Implications for the tech sector

For the tech industry, this view is uncomfortable because it challenges its own narrative. The sector has long claimed that automation creates new markets, jobs, and productivity. That might still hold long-term. But the paper warns that the transition could be costly if it happens too rapidly without effective reabsorption mechanisms.

This risk is especially high where AI replaces entry-level cognitive roles. If many junior positions in programming, support, analysis, marketing, operations, or customer service disappear, the market isn’t just losing employment—it’s also losing pathways for learning, internal promotions, and talent renewal.

The question for tech companies isn’t whether they should use AI; it’s how to prevent adoption from turning into a race to cut staff without a clear transition plan. Each individual automation might be justifiable, but the real issue arises when thousands implement similar cuts simultaneously.

A warning for investors and executives

The paper also offers a financial perspective. If a company announces more automation, the market might reward cost savings. But if many companies do so, overall demand can weaken. The risk for investors is confusing short-term margin improvements with sustainable growth.

In sectors like software, commerce, digital consumption, or mass services, demand depends on households with income. Automating internal processes may boost short-term EBITDA, but a deteriorating labor market can reduce subscriptions, purchases, renewals, and spending. AI doesn’t eliminate that relationship.

For managers, the practical takeaway is that automation should be evaluated using broader metrics: cost per task, quality, customer impact, reputational risk, internal reabsorption capacity, and talent effects. Not every staff reduction equates to sustainable efficiency.

The paper doesn’t predict the end of the economy

It’s important to clarify that The AI Layoff Trap doesn’t claim the economy will inevitably collapse because of AI. It shows that under certain assumptions, a competitive over-automation trap can emerge. If displaced workers find quality jobs, new tasks are created, wages re-balance without impoverishing most, and policies effectively recycle income, outcomes could differ.

But the warning deserves attention because it goes beyond generic fears of technology. It doesn’t just say AI destroys jobs; it highlights that demand destruction could threaten the support base of companies if the transition relies solely on private incentives.

This is the key consideration for the tech sector. Deploying AI agents isn’t just about APIs, benchmarks, cost per token, or productivity metrics. It’s also about the economic architecture: if technology allows producing more with fewer people, who will have sufficient income to buy what is produced?

AI can be a powerful productivity tool, but productivity without demand isn’t prosperity—it’s idle capacity. Falk and Tsoukalas’s paper reminds us that economies don’t break when a single company automates a task. They can break when all do so simultaneously, driven by rational logic, without mechanisms to sustain the consumer base.

Frequently Asked Questions

What does The AI Layoff Trap propose?

It models rational companies that automate more than optimal by capturing cost savings but only partially consider the demand they destroy by displacing workers.

Is it a prediction of economic collapse?

No. It’s a theoretical model with specific assumptions. Its value lies in illustrating a risk: incentives toward over-automation if the labor market can’t reabsorb displaced workers at the same pace.

Why is agentic AI especially relevant?

Because agents can automate entire workflows, not just individual tasks, potentially accelerating layoffs or reducing future hiring across sectors.

What solution does the model suggest?

The only solution within the model is a pigovian tax on automation: a levy per automated task, designed to internalize the demand destruction—that is, to ensure companies pay for the externality they generate.