The share of American business spending directed toward freelance platforms has been divided by five in four years. Not slowed: divided by five.

This is the main lesson from data published by Ramp Economics Lab, the research division of fintech Ramp which manages corporate credit cards for thousands of American companies. Their advantage over any opinion survey: they see actual transactions, not intentions. And what the transactions show is unambiguous.

In four years, companies that outsourced work to platforms like Upwork have, for more than half of them, completely stopped doing so. Those that continue are spending much less. And the share that went to freelancers is now being absorbed, for the most part, by AI tools.

The Essentials

  • The share of business spending toward freelance platforms fell from 0.66% (Q4 2021) to 0.14% (Q3 2025), a 79% decline in relative volume, according to Ramp Economics Lab
  • More than half of the companies using these platforms completely stopped relying on freelance platforms over the same period
  • For companies most exposed to AI (those that spent the most on freelancers), the substitution ratio is 3 cents spent on AI for every dollar saved on freelancers — a replacement cost far below the outsourced work it replaces
  • The phenomenon first affects frictionless contractual tasks: writing, translation, code, design; support functions in large enterprises constitute the next frontier

Freelancers are the canary in the coal mine of the labor market. Without unions, without severance packages, without contractual notice periods, they absorb market shocks before salaried employees. What their decline portends for the rest of the labor economy deserves careful scrutiny.

In Four Years, an Unprecedented Collapse in Demand for Outsourced Work

The starting point for Ramp’s data is the fourth quarter of 2021. This was the post-pandemic peak for freelance platforms: companies had discovered remote work, they were turning to independent contractors for one-off projects, Upwork and its competitors were posting record valuations. The share of business spending going to these platforms reached 0.66% of total measured professional spending.

By late 2025, that share had fallen to 0.14%. The trajectory is nearly linear in its descent: no sudden collapse, no rebound. A constant erosion, quarter after quarter, as generative AI tools entered into professional use.

This figure has a useful characteristic: it measures a relative share, not an absolute volume. The American economy continued to grow over this period. If the share going to freelancers fell 79% in relative terms, the decline in actual real volume is massive.

In parallel, the share of spending toward AI suppliers rose from zero to approximately 3% over the same period. This is not a calendar coincidence. It is direct substitution, measurable in the same companies’ accounting lines.

The correlation is not perfect — other factors are at play, including the 2022-2023 economic slowdown and rising interest rates. But the substitution ratio calculated by Ramp is striking: for companies most exposed to AI, each dollar saved on freelancers corresponds to only 3 cents of AI spending ($0.03). In other words, AI costs a fraction of the human work it replaces, for comparable tasks. This is not marginal productivity gain. This is an order of magnitude change.

Frictionless Tasks Disappear First

Why freelancers before salaried employees? The answer comes down to one word: friction.

Firing an employee is expensive. It requires notice periods, severance, sometimes union negotiations. Modifying an internal position takes months. Abandoning a freelance contractor, by contrast, requires only a decision not to renew the project. Platforms like Upwork were built precisely on this argument: flexibility. That same flexibility is what makes them vulnerable first.

The tasks affected first are those defined by clearly specified deliverables: an article written, a logo created, a block of code produced, a document translated. Precisely the tasks for which generative AI tools today achieve acceptable performance, sometimes even superior on certain dimensions like speed and cost.

It’s not that freelancers worked poorly. It’s that their business model was based on an information asymmetry that AI has erased. A small company without a marketing department needed a freelance writer because it lacked in-house expertise. It can now use an AI tool to produce a first draft, which an existing employee reviews in twenty minutes. The transaction with the freelancer disappears.

This is the same mechanism described by economist David Autor in his work on labor market polarization: technology first eliminates clearly defined routine tasks, whether manual or cognitive. What remains — complex judgment, client relationships, high-value-added creativity — resists longer. On Upwork, the jobs that survive are those requiring specialized expertise and context difficult to convey to a machine: specialized development, strategic consulting, qualitative research.

More Than Half of Companies Cut Ties Entirely

The most surprising figure in Ramp’s data is not the overall decline. It’s how that decline is distributed.

More than half of the companies that used freelance platforms have completely stopped doing so. Not reduced their usage: stopped entirely. The other half has maintained some use, often at lower levels.

This says something important about the nature of substitution. It’s not a gradual rebalancing where each company reduces its freelance spending a bit and increases AI spending a bit. For the majority, it’s a binary rupture: on one side, companies that have entirely shifted to AI for the tasks in question, on the other, companies that continue outsourcing because their needs exceed what AI can do alone.

This bimodality suggests that substitution is not linear. It occurs in thresholds: below a certain complexity level, AI suffices and the resort to freelancers disappears entirely. Above that threshold, the freelancer remains useful and substitution is partial. The complexity threshold above which AI “suffices” rises with each new generation of models.

Organizations that best capture AI’s gains are precisely those that have successfully redeployed their teams toward high-value-added tasks that machines don’t cover, as noted in an analysis on the relationship between organizational preparedness and actual AI gains.

Support Functions in Large Enterprises: The Next Frontier

Until now, the debate on AI and employment mainly concerned abstract categories: “jobs with routine tasks,” “low-skilled white-collar workers,” “knowledge workers.” Ramp’s data offers something more useful: a concrete case, measured on actual transactions, with a precise substitution ratio.

And this case suggests that the next terrain is the support functions of large enterprises.

The logic is the same as for freelancers, with a time lag related to contractual friction. An internal communications department producing press releases, presentations, and internal newsletters does exactly what freelancers did: well-defined, repetitive deliverables, from standardized briefs. The difference is that these employees have contracts. Replacing an internal position is costly socially, politically, legally.

But the calculation changes when the position becomes available for other reasons: natural departure, non-renewal of a fixed-term contract, reorganization. In these cases, the question “do we replace this position?” now arises with an additional option: AI. And the ratio of a few cents per dollar replaced makes the economic answer obvious for the tasks in question.

Several large American companies have already reported reductions in their marketing, communications, and legal support staffs. Duolingo explicitly announced in early 2025 that it would not replace departing content contractors, replacing them with AI tools instead. Klarna temporarily replaced the equivalent of 700 customer support agents with AI in early 2024, before having to rehire human agents beginning in May 2025 after service quality degraded. These decisions remained anecdotal as long as they weren’t measured in aggregate data. Ramp’s data shows this is a systemic phenomenon, not a series of isolated announcements.

This is precisely why several American states are seeking to regulate these transitions: without regulation of response times and reclassification obligations, the speed of substitution risks exceeding workers’ capacity to adapt.

What the Ratio Reveals About Sharing the Gains

The substitution ratio deserves closer examination. For every 100 dollars not spent on freelancers, companies most exposed to AI spend approximately 3 dollars on AI tools (a ratio of $0.03). The remaining 97 dollars go somewhere.

Three possible destinations. First option: they remain in the company’s pocket, in the form of improved margins. Second option: they are reinvested elsewhere in the business, generating new internal employment. Third option: they are passed on to customers as price reductions, with demand effects creating activity elsewhere in the economy.

Ramp’s data does not allow us to distinguish between these three scenarios. But the issue is central to understanding whether the observed substitution is net for employment or merely a displacement.

Economist Daron Acemoglu, in his recent work on automation and the direction of technological progress, distinguishes between two types of innovation: those that increase human productivity by complementing workers’ abilities, and those that substitute for them without creating enough new jobs. His diagnosis on generative AI is that the vast majority of current investments aim at substitution, not augmentation. If the remaining 97 dollars go to margins rather than new jobs, this scenario is confirmed.

The question is not rhetorical. If productivity gains are captured by a limited number of actors — companies and their shareholders, AI suppliers — without redistribution, substitution on Upwork is not a transformation of the labor market but an extraction of value from workers to capital. This is not inevitable. Policies for sharing productivity gains exist: redistribution through taxation, reclassification obligations, collective bargaining over automation conditions. But they require explicit political choices.

The End of the “AI Creates as Much as It Destroys” Argument

The debate on AI and employment has long remained abstract. On one side, optimists cited historical precedents: the power loom, the automobile, information technology. Each technological wave had ultimately created more jobs than it destroyed. On the other, pessimists argued that the speed and generality of generative AI broke with these precedents.

Ramp’s data does not settle the long-term debate. But it imposes an empirical constraint that the optimist argument must now confront: in a measurable sector, over four years, substitution is rapid, massive, and inexpensive. The historical reasoning “it has always created jobs in the end” says nothing about the duration of transition or who bears it.

For the 59 million American independent workers counted by the Freelancers Union, the end could be a long time coming. Resort to freelance platforms has constituted for many an informal safety net: supplementary income, activity during transitions between salaried jobs, a first step toward entrepreneurship. If this market shrinks durably, this net disappears without being replaced by another.

The trajectory on Upwork is clear. What remains open is the speed at which the same phenomenon will reach companies’ internal functions, and whether labor market institutions will be able to adapt their responses to a substitution measured in quarters rather than decades.

Sources

  1. Ramp Economics Lab — AI Labor Market Impact on Freelancers
  2. arXiv — Payrolls to Prompts (full academic paper)
  3. Fortune — Duolingo CEO replaces contractors with AI
  4. Entrepreneur — Klarna reverses AI customer service
  5. American Economic Review — Autor & Dorn (2013)
  6. Acemoglu, D. & Johnson, S. — Power and Progress, PublicAffairs, 2023 (no link)
  7. Freelancers Union — annual report on American independent work (no link)
  8. Autor, D. — work on labor market polarization, MIT (no link)