The First Costs of AI Substitution May Be Paid by Workers We Count Poorly

Discussions about artificial intelligence and employment tend to look for visible events: layoffs, hiring freezes, falling employment, or occupations forecast to shrink. These indicators matter, but they do not capture every way a business can change its relationship with work. A company can alter what it buys, who performs a task, and how much external help it needs without dismissing a permanent employee.

Ryan Stevens’s working paper, Payrolls to Prompts, offers a way to examine this less visible part of the transition. Using firm-level expenditure data, the study compares spending on selected online labour marketplaces with spending on identified AI model providers. It finds that firms with greater prior reliance on online labour marketplaces adopted AI providers more intensively after ChatGPT’s public release and reduced marketplace spending relative to less-exposed firms.

The result is worth examining, but it should be interpreted carefully. Spending data can reveal changes in how firms acquire inputs. It does not directly establish how many people lost work, whether their earnings fell, or whether the same tasks were completed with equivalent quality. The broader question is what can be learned when a change in the purchase of work becomes visible before its consequences for workers are fully understood.

When a business changes what it buys

A firm’s demand for work is expressed through many decisions. It may hire an employee, commission a freelancer, purchase a service from an agency, use software, or develop a task internally. These arrangements can overlap, and a task can move between them as costs, capabilities, and requirements change.

AI introduces another option into that set of decisions. A business may use a model to produce a first draft, analyse a document, generate a design variation, or assist with a coding task. Depending on the quality threshold and the surrounding workflow, it may then purchase fewer external services, commission a different kind of work, or ask an existing team to do more.

These possibilities are not equivalent. A lower contractor bill could reflect a task completed with AI assistance, a task brought in-house, a change of supplier, a lower volume of work, or work that is no longer required. Expenditure alone cannot reliably separate these explanations.

The value of the spending signal is therefore not that it settles the question of displacement. It identifies a change in firms’ purchasing behaviour that deserves closer investigation. That is a meaningful finding, provided the signal is not mistaken for the full outcome.

What the paper measures, and what it does not

The study tracks expenditure on selected online labour marketplaces and AI model providers within the data available to the researcher. Between the fourth quarter of 2021 and the third quarter of 2025, the share of total spending allocated to the marketplaces covered by the analysis fell from 0.66% to 0.14%. The share allocated to the identified AI providers rose from zero to 2.85% over the same period.

The paper also estimates differences between firms according to their prior reliance on online labour marketplaces. Among the most exposed firms, the estimated spending changes are consistent with substitution between marketplace expenditure and AI-provider expenditure. The author discusses a relationship of approximately three cents of additional AI-provider spending for each dollar of reduced marketplace spending in this group.

That ratio should not be read as the price of replacing a dollar of human work with AI. It is an estimate based on expenditure measures, not a direct comparison of equivalent tasks or total production costs. The analysis does not fully capture internal engineering, infrastructure, integration, supervision, or maintenance costs. It also does not measure output quality or the amount of work completed.

The study’s design compares firms with different levels of prior marketplace exposure after ChatGPT’s release. This is informative, but it does not remove every interpretive concern. The paper notes that the most exposed firms had significantly higher marketplace spending in an earlier pre-release period, raising a potential pre-trend issue. Its sample and spending definitions also limit how far the results can be generalised.

The appropriate conclusion is specific: in the firms and expenditure categories examined, the data show a shift in spending consistent with some substitution between purchased online labour and AI-provider services. They do not establish a general rate of worker replacement or prove that the wider economy is losing jobs as a result.

The transaction is not the worker

When a company reduces a contract, its accounting records a change in expenditure. The worker’s experience depends on a different set of conditions. A freelancer may have several clients or depend heavily on one recurring account. A reduction in assignments may be manageable for one person and financially destabilising for another.

The spending data do not reveal these differences. They do not show whether a contractor found replacement work, changed specialisation, used AI to become more productive, accepted lower rates, or left the market. Those outcomes cannot be inferred from the firm’s payment record alone.

This distinction matters because the unit visible to the business and the unit experiencing the change are different. A firm sees a task, a supplier, a budget, and a delivery requirement. A worker experiences income, time, uncertainty, relationships with clients, and the need to adjust. The same purchasing decision can have different consequences depending on the worker’s circumstances.

It would be premature to conclude from this study that freelancers are universally the first or most affected group. The paper offers a reason to examine contracted work closely, not a ranking of who will bear AI’s costs. Its contribution is to make one part of the changing relationship between firms and external labour observable.

The missing connection between spending and outcomes

A fuller account of AI’s labour-market effects would connect firm spending to what happens to the work itself. Researchers would need to know whether assignments disappear, change in scope, move to other suppliers, or become more productive through AI assistance. They would then need to connect those changes to worker-level outcomes, including earnings, hours, assignment volume, client concentration, and movement into other work.

Productivity requires its own measurement. A lower cost for one input does not automatically mean that a firm produces more value with fewer resources. The business may incur costs elsewhere, spend time checking outputs, accept a different level of quality, or reorganise the task in ways that change the total process. A convincing productivity claim requires evidence about outputs and relevant costs, not just a change in one expenditure category.

Aggregate employment is another distinct question. Evidence that some firms reduce marketplace spending cannot, by itself, establish whether employment across an industry or economy will fall. Other firms may expand, new tasks may emerge, and the division of work between people and systems may change. Those possibilities should not be used to dismiss losses in particular places, just as individual examples of substitution should not be used to infer economy-wide displacement.

The analytical task is to connect the layers rather than collapse them. Spending data can identify where to look. Task-level evidence can help explain what changed. Worker-level evidence can show who experienced the adjustment. Productivity and employment measures can then help establish the broader economic consequences.

Why the first signal matters

The importance of the study lies in the question it makes possible. If firms change how they purchase work, researchers should not wait for every consequence to appear in headline employment figures before examining the transition. At the same time, they should not treat an early signal as a final verdict on the future of work.

This requires a disciplined distinction between observation and explanation. A spending decline is observable. AI substitution is one possible explanation that the paper investigates through its research design. A particular worker’s loss of income, a firm’s productivity gain, and an economy-wide employment effect are further claims that require additional evidence.

That discipline is not a reason to understate the issue. It is how the issue can be understood without turning a limited but useful finding into a sweeping conclusion. The firms’ purchasing decisions are part of the story; the subsequent organisation of work and the outcomes for people are the rest of it.