Alex Willen runs 12 e-commerce brands with no employees and expects between $2.5 million and $3 million in revenue this year. Kristin Ginn left Microsoft in March, started an AI-adoption consultancy, and seven months later has roughly $250,000 received or contracted through year-end with no staff and no regular contractors.
Both are the kind of case that gets used to argue the cost of running a company has fallen permanently. The argument is easier to make than to test, because the largest dataset in circulation counts applications, not companies that survived.
Start with the arithmetic a solo operator actually faces. A product photography shoot runs $2,000 to $5,000. A designer charges $300 to $1,000 for a set of listing images. Willen uses Anthropic’s Claude for bookkeeping, customer service, advertising and inventory, and OpenAI’s image tools for work that once went to photographers. Ginn spends under $60 a month on paid AI tools.
Remove those line items and the model changes shape. Willen says that without AI he would need one more full-time person plus contractors for photography, design, bookkeeping and ad management. He also rebuilt the Shopify store of a hunting-gear brand he acquired, Lone Trail, which had been running at roughly 1% of sales; in a recent 30-day stretch it accounted for about 30% of that brand’s revenue, which he puts at six figures a year.
The evidence that this is happening broadly comes from Liya Palagashvili at George Mason University’s Mercatus Center. In industries heavily exposed to AI, business applications showing no early hiring signals rose 26.8% from the first quarter of 2024 to the first quarter of 2026, from about 222,000 to 281,000 per quarter. Applications in the same industries that looked likely to become employers fell 6.4%. In construction and wholesale trade, her comparison group, non-employer applications slipped 0.4%.
A business application is a declaration of intent.
That is the limit of the series. The Census-based measure flags filings without indicators of near-term payroll or hiring plans; it does not follow those businesses to see whether they stayed at one person, hired later, or closed. Palagashvili describes her findings as preliminary and descriptive, and does not claim AI caused the rise.
Data from firms that do see payroll points the other way. Gusto tracked small businesses after they adopted AI and found headcount about 7% higher after a year than at comparable companies that knew about the tools and did not adopt them. Among firms with fewer than 10 employees, adopters added staff about 10% faster than non-adopters.
Both results can be true at once. A tool that lets one person start a company also lets an existing five-person firm take on work it would previously have declined, and the second effect shows up in payroll while the first shows up in filings. The FreshBooks survey sits between them: 79% of solopreneurs said they tried AI before hiring someone for a task they could not do themselves, and 86% when microbusiness owners are included.
What the solo model does not remove is dependence on distribution. Roughly 80% of Willen’s revenue still arrives through Amazon, with Meta and Shopify carrying direct sales. Ginn has no staff, and LinkedIn is among her largest sources of customers. The software bill is small; the platform bill is the business.
Erik Brynjolfsson at Stanford reads that as a feature of the shift rather than a bug. A longer tail of small producers raises the value of whatever aggregates them, he argues, because the platforms sell the tools and supply the customers. His point is not that the savings flow to big technology companies. Willen keeps some as profit and Ginn’s tool spend is trivial, so the transfer is thinner than the rhetoric suggests.
The more consequential claim in his work concerns firm size. In a September paper with Zoë Hitzig of the Anthropic Institute, Brynjolfsson argues that as AI codifies knowledge once held by individual workers and makes it available centrally, average firm size rises, industry concentration increases, and local managerial discretion declines. The authors describe early evidence as broadly consistent with those patterns.
Adam Branch has run the experiment twice. He built Incentive Technology Group to about 500 employees, where winning more work meant hiring more people. His current company, rhino.ai, has about 50 and does comparable modernisation work at three to four times the output per employee. At the earlier firm, labour was roughly 85% of operating spending and technology 5%. At rhino.ai he puts labour near 40% and technology, including cloud, software and models, near 25%.
The pattern Brynjolfsson draws from cases like that is a hollowing out of the middle: very small teams doing more at one end, very large organisations combining AI with scale at the other, and the management layers built for internal coordination squeezed out.
AWS’s own survey of 3,413 founders across 20 countries, restricted to companies under five years old using AI intensively, reports average annual revenue growth of 156% against 65% for startups generally, with more than half above $400,000 of revenue per employee. Ruba Borno, who runs the company’s specialists and partners group, describes the change as scaling with technology instead of headcount.
Surveys of that kind select for the companies that worked. A solo operator with a failed storefront does not appear in a sample of founders, and the $400,000 per employee figure is an average drawn from survivors of companies that were still trading when the survey was taken.
What would settle the question is longitudinal. The 281,000 quarterly applications in AI-exposed industries can be matched against payroll records in three and five years, which would show how many became employers, how many stayed at one person and how many disappeared. Nobody has published that yet.
Until then the honest reading is that the cost of trying has fallen and the cost of failing has not.
Ginn’s $250,000 in contracted work and Willen’s $3 million are outcomes, not a base rate. The filings that analysts are counting describe an intention to find out, and the finding out has barely started.