AI startups are exporting labor costs while calling it efficiency.
The Wall Street Journal reports a real change in how some early-stage companies are organized. Founders are asking whether a new hire is necessary before authorizing the hire. Some have answered by dismissing employees, ending contractor relationships, and replacing agency work with software. The engineering output may improve. The accounting is incomplete.
Butternut AI had nine employees at its peak and now has four. Co-founder Simar Singh told the Journal that five engineers were let go after the founders concluded that AI agents could perform their work around the clock. The company continues to serve about 10,000 paying customers. That establishes a business fact. It does not establish what happened to the five engineers.
The Journal does not report whether they found new work, accepted lower pay, left engineering, or remained unemployed. Those are not incidental details. They are the labor-market result of the decision. A firm can reduce its head count without reducing the work required to build and maintain its product. The work moves to software, to the remaining employees who supervise it, or to workers elsewhere who compete for fewer openings.
Lindy presents the same transaction in a different department. The company laid off marketing employees during a product pivot and did not refill their positions. Its marketing team remained at two people, supplemented by AI tools. Five outside creative agencies were also dropped as AI-generated video replaced their work. Lindy’s user base grew to roughly half a million, and the company continued hiring coders.
The growth figures show that Lindy can produce more with fewer paid contributors. They do not show whether the laid-off marketers found comparable jobs, whether the agencies lost a client or a material share of revenue, or whether the displaced work will support any small business in the contractor economy. An agency is not a line on an organizational chart. It carries employees, freelancers, equipment, leases, and other clients. Ending five contracts can transmit a revenue shock well beyond the startup’s payroll.
That is the missing public stake. The founder sees a labor cost. The worker sees an occupation becoming less secure. The agency sees a customer relationship that may not be replaceable. The public sees a labor market in which the gains from a new technology may accrue to the firm and its investors while the adjustment costs are assigned to people with less bargaining power.
LinkedIn data cited by the Journal show that U.S. startups founded in 2022 with at least five employees in their first year averaged 20.7 workers by year three, compared with 32.5 for similarly sized startups founded in 2018. That is a decline of roughly one-third in average head count. The data describe fewer jobs inside younger startups. They do not tell us whether the displaced workers moved to other firms, started businesses of their own, accepted lower compensation, or disappeared from the measure.
The distinction matters because head count is not the same as productivity. Productivity means more output from a given quantity of labor and capital. It does not answer who receives the resulting income, who bears the risk, or whether the displaced worker can obtain an equivalent position. A smaller company may be a productive company. It may also be a company that has shifted the cost of transition onto workers and contractors.
Y Combinator points to Emergent, with $15 million in projected yearly revenue and 15 people, and Retell AI, with $60 million and about 40 people. AgentCollect has reduced its staff from 50 to about 30, automating billing coordination, call-quality evaluation, and product-management work. These examples show that a lean company can generate substantial revenue. They do not show that the gains are broadly shared, or that the jobs removed were low-value jobs that the workers could readily replace.
The revenue numbers also describe a concentrated form of value capture. A small group of founders, investors, and remaining employees may control a larger stream of revenue with fewer claims against it. That can be a legitimate return on useful innovation. It can also become a mechanism for transferring income from a wide group of workers and suppliers to a narrow group of owners. The source material does not provide the ownership, compensation, or distributional data needed to decide which description dominates. It does provide enough to reject the claim that the matter is simply entrepreneurial clarity.
The investor-side counterargument is narrower and more credible. First Round Capital’s Liz Wessel says companies that are scaling rapidly still compete for engineers. Cowboy Ventures founder Aileen Lee says enterprise customers often want to buy from a company with human staff. Venture capitalists also expect startups seeking a Series A investment to reach roughly $4 million to $5 million in first-year revenue, a target that is difficult for a company with no employees beyond its founders.
Those constraints show that the lean model is not a universal law. It works better for self-serve software and products whose customers accept automated support than for businesses selling expensive services where human accountability is part of the product. The business model determines how far the head count can fall.
The source contains no worker-side interview, no contractor-revenue data, no wage series, and no reemployment measure. That absence is itself the limit of the argument. The Journal gives us founder testimony, investor testimony, company revenue, and employee counts. It does not give us the outturn for the people whose jobs and contracts ended. A policy-literate account cannot call displacement an efficiency gain until it records both sides of the ledger.
The résumés that now advertise AI fluency belong to that ledger. Workers are being told to demonstrate that they can use the tools that may replace the tasks they previously performed. That is a rational response to changed demand. It is not evidence that the transition is fair, or that every worker can become the supervisor of an AI system without losing income, status, or a profession.
The useful question is not whether a four-person startup can serve 10,000 customers. It can. The question is who owns the resulting capacity, who bears the cost when labor is displaced, and whether workers have bargaining power, portable benefits, retraining, or a credible route into the next job. Any serious productivity account must pair revenue and head count with those measures. Otherwise the score records the founder’s gain and leaves the worker’s loss outside the table.
The four-person startup is not the new forty-person startup for everyone. It is a test of who is permitted to count as a cost.