Automation may be profitable for firms but costly for the wider economy

Artificial intelligence (AI) may deliver higher productivity for firms while quietly weakening the consumer demand that keeps the wider economy moving, according to a new economic model that challenges the standard pro-growth framing of automation. The paper finds that when AI displaces income from low-wealth workers who spend most of what they earn, firms can over-automate even as output per worker rises and high-skilled employees benefit.

The study, titled "The Demand Externality of Automation," was published as an economics working paper by Erhan Bayraktar. It uses a static benchmark and a stationary heterogeneous-agent general equilibrium model to show how automation can produce either productivity-led capital growth or demand-base erosion, depending on productivity gains, worker exposure, capital ownership and fiscal policy.

Automation's hidden cost lies in who loses income

The study primarily claims that automation should not be analyzed only as a production technology. In conventional firm-level reasoning, automation is attractive because it raises output per unit of input and reduces the need for paid human labor. A company that automates can lower costs, improve productivity and increase profit. That logic is real, Bayraktar argues, but incomplete.

The missing factor is demand. Workers are also consumers. When automation reduces labor income for households that spend most of their earnings, the economy can lose part of its consumption base. A firm deciding whether to automate does not internalize that broader decline in household demand. The paper calls this missing marginal effect the demand externality of automation.

This is especially important when exposed workers are low-wealth households with high marginal propensities to consume. Such households have limited savings and tend to spend a large share of any additional income. If automation cuts their wages or hours, the loss quickly shows up in consumption. By contrast, if automation gains accrue mainly to wealthy households, shareholders or foreign owners, more of the income may be saved, retained or moved outside the domestic consumption base.

Bayraktar's model shifts the focus away from a simple question of whether automation destroys jobs. The issue is not only how many workers are displaced. The more important question is whose income is displaced, who receives the gains and how those income changes affect consumption, savings, wages, investment and capital accumulation.

The paper builds a two-sided framework. On one side, automation can generate genuine economic gains. It can raise a productivity shifter, complement high-skilled workers and increase high-skilled labor income. These channels explain why firms adopt AI and why some workers and asset owners may benefit. On the other side, automation can reduce paid low-skill tasks, shrink the collective wage bill of exposed workers and weaken demand if those workers are central to consumption.

This dual structure allows the paper to avoid a one-sided conclusion. It does not claim that AI is inherently harmful or automatically beneficial. Instead, it identifies the conditions under which each outcome emerges. Automation becomes productivity-led when productivity gains are large, high-skilled labor is strongly complemented, capital does not become obsolete, and ownership of automation rents is broad. It becomes demand-eroding when low-wealth workers are heavily exposed, ownership is concentrated and fiscal support is limited.

The study uses a heterogeneous-agent model because a representative household would miss the central mechanism. If all households were treated as identical, the model would not show why income losses among low-wealth workers matter more for demand than income gains among wealthy asset owners. The framework includes households that differ by skill and wealth, face incomplete insurance and save through capital or equity claims. Wages and returns are determined by market clearing, while the wealth distribution evolves through a formal macroeconomic system.

That design matters because automation affects several forces at once. It changes productivity, labor income, capital returns, asset ownership and the stationary wealth distribution. A low-skilled worker with wealth may be partly protected by asset income. A high-skilled worker with little wealth may still depend heavily on labor income. The model's core point is that economic exposure depends on both skill and wealth.

The study also highlights the role of ownership. If automation rents flow broadly to domestic households, the income gains can support consumption and savings. If they are concentrated among high-wealth households or foreign owners, they do less to offset the loss of labor income among exposed workers. The paper therefore treats ownership not as a side issue, but as a central determinant of whether AI lifts the whole economy or mainly redistributes income upward.

AI can fuel growth, but only under the right conditions

The author reports two main regimes to show how the same model can produce different AI outcomes. The first is demand-base erosion. In this scenario, firms choose substantial automation because it raises productivity and high-skilled labor income. However, the broader economy suffers because automation sharply reduces low-skilled labor income, weakens consumption and lowers capital accumulation.

In the demand-base erosion baseline, decentralized firms choose an automation index of about 0.526. This is a large adoption scenario. High-skilled labor income rises substantially, showing that automation is not simply destructive in the model. Firms and high-skilled workers have clear private reasons to support adoption. Yet low-skilled labor income falls sharply, aggregate consumption drops and the capital stock declines.

The government target in that baseline is much lower automation, reaching zero in the reported exercise. This does not mean the study treats all automation as socially useless. Rather, it means that under those parameters, the demand loss from exposed households overwhelms the private productivity gains. The private firm's first-order condition does not include the lost consumption demand of low-wealth workers, so private automation becomes excessive.

The second regime is productivity-led capital growth. In this scenario, automation produces stronger productivity gains, complements high-skilled work more powerfully, exposes low-skilled workers less severely and causes less obsolescence of existing capital. Under those conditions, automation raises output, capital and aggregate consumption.

In the productivity-led case, decentralized automation increases the capital stock, raises output and lifts total consumption compared with the no-automation allocation. This is the optimistic AI scenario. The model shows that it is possible, but it requires favorable conditions: strong productivity growth, weaker displacement of the consumption base, limited destruction of older capital and enough ownership pass-through for households to benefit from automation rents.

Even in the productivity-led case, however, the paper finds an important warning. Aggregate consumption can rise while low-skill, low-wealth households still lose. Bayraktar's results show that poor exposed households may see consumption fall even when the overall economy becomes richer. That distinction matters for policymakers because growth alone may not protect vulnerable groups.

The study separates aggregate abundance from distributional protection. A society can become richer in total while specific groups lose income, bargaining power or consumption capacity. If policymakers look only at output or productivity, they may miss the people who are harmed. If they look only at displacement, they may miss the productivity benefits. The model argues for tracking both.

The paper also addresses why capital and returns can move in ways that appear counterintuitive. In the demand-base erosion case, capital falls as automation rises, but the investment return does not necessarily rise. That is because automation also reduces production labor and can increase obsolescence of legacy capital. In the productivity-led case, capital rises, but the return on productive capital can fall because capital becomes more abundant. These outcomes show why automation must be studied in general equilibrium rather than through firm-level cost savings alone.

The model's short-run diagnostics are directly relevant to the current AI debate. A productivity-led path should show broad consumption growth, productive investment and employment reallocation into new high-value tasks. A demand-base erosion path should show rising high-skill premiums and stronger owner income alongside stress among exposed low-wealth workers. The paper suggests that the next phase of AI's economic impact may be split rather than uniform.

This is consistent with the study's use of empirical guideposts. The paper draws on evidence showing that high-AI-exposure jobs are often higher-paid and that equity ownership is concentrated. That combination supports both sides of the debate: AI may complement high-skilled workers and raise productivity, but the gains may flow disproportionately to people who already own capital or occupy higher-wage jobs.

The study shows that these facts are not separate. They interact. If the workers most helped by AI are high-skilled and the owners of automation rents are wealthy, the gains may not replace the demand lost when lower-wealth workers lose income. The resulting economy can experience private adoption, higher productivity in some tasks and weaker household demand at the same time.

Tax policy must target the automation margin and protect demand

The paper's policy analysis focuses on the difference between taxing automation as a slogan and designing a fiscal mechanism that actually changes firm behavior and household welfare. Bayraktar argues that the policy target is not the word tax. The target is the missing derivative in the firm's automation decision.

A firm automates based on the private marginal benefit of automation compared with its cost. A corrective policy must alter that margin. A literal automation tax can do this by making automation more expensive at the margin. A retained-labor subsidy can also do it if it rewards firms for keeping paid human task input. Different instruments can produce the same marginal effect at the target level of automation, but they may have very different fiscal and distributional consequences.

The study stresses that fiscal closure is not optional. If a government taxes automation, it must decide what happens to the revenue. It can rebate the money to households, lose some of it through administrative or political frictions, spend it elsewhere or transfer it to other claimants. Without specifying that closure, the incidence of an automation tax is unclear.

One important result is that a tax that fully deters automation may raise no long-run revenue because it eliminates the tax base. If automation falls to zero, the government collects no automation tax. This creates a tension between welfare and revenue. A welfare-focused government may prefer a high tax that prevents excessive automation in a demand-base erosion regime. A revenue-focused government may prefer a lower tax that leaves some automation in place and collects revenue from the remaining base.

The paper shows that rebate design matters. A lump-sum rebate gives the same amount to every household. A labor-income-proportional or total-income-proportional rebate can favor higher-income or higher-wealth households. A progressive rebate targeted toward low-wealth and low-earnings households gives more support to the households with the highest marginal propensities to consume.

This is critical to the demand externality. If the problem is the loss of spending by low-wealth households, then rebates targeted to those households are more effective in supporting consumption. The study finds that progressive rebates can deliver much larger direct consumption support to low-skill, bottom-wealth households than equal rebates, because those households both receive more and spend a larger share of the transfer.

The framework does not present taxation as a universal answer. In the productivity-led regime, a tax that eliminates automation can reduce consumption and capital by sacrificing productivity gains. In that case, the better policy may be a positive-revenue tax paired with targeted transfers, allowing society to keep some productivity benefits while protecting exposed households. The appropriate policy depends on which AI regime the economy is actually in.

The paper also examines domestic ownership. If households own a larger share of automation rents, asset returns can partly offset labor income losses. But ownership protection is limited when capital ownership is concentrated. Wealthy households may be insulated, including wealthy low-skilled households, while low-wealth exposed households remain vulnerable. Broad ownership matters more than aggregate ownership alone.

Policymakers should not rely only on productivity forecasts, job exposure counts or headline adoption rates. They need to measure which workers lose labor income, which households have high spending sensitivity, who owns the capital and automation rents, and whether fiscal systems transfer gains toward the consumption base.

The paper also suggests that simple representative-agent models may understate AI's risks. If average income rises, those models can make automation look socially beneficial. But if the gains accrue to lower-spending households and the losses hit high-spending households, aggregate demand can weaken even when productivity improves.

The study explains why individually rational decisions can add up to a weaker economy. Each firm may automate to reduce costs, but if many firms reduce income for high-MPC workers at the same time, the resulting demand loss can hurt the broader market. This is the same logic behind the paper's title: automation has a demand externality because firms do not fully account for how their labor-saving decisions affect the consumer base.