MIT researchers David Autor and Neil Thompson have published findings showing that automation recomposes jobs in ways that can deliberately compress wages for specific groups of workers, challenging the simple displacement narrative and raising a harder question: whether AI adoption is becoming a structural mechanism for employers to discipline wage growth where workers have the least leverage.
The framing matters. The dominant story told about automation and AI has been about replacement: technology does the job, worker loses it, economy adjusts. Autor and Thompson's research shows the mechanism is more precise and more consequential than that. When automation removes a job's most complex tasks, the remaining work becomes easier to enter, competition for it rises, and wages fall. Inventory clerks are the paper's most direct example: as computerisation targeted their most expert tasks, employment in the role more than doubled while real wages fell 13 percent. More workers, cheaper pay, productivity gains captured by the employer. That is not displacement. That is wage suppression through task erosion.
The inverse is equally instructive. When automation removes the simpler tasks from a job, the remaining work demands more expertise, fewer people qualify for it, and wages rise. Bookkeepers who survived spreadsheet automation became financial analysts. The role shrank in headcount but expanded in pay. The policy implication is uncomfortable: automation does not affect all workers equally, and the direction of the effect depends entirely on which part of the job gets automated, a decision made by the employer, not the worker. That asymmetry is where the power dynamic lives.
Parallel research from the San Francisco Federal Reserve, published in the American Economic Review in 2024, formalises the bargaining dimension directly. Sylvain Leduc and Zheng Liu argue that the threat of automation weakens workers' bargaining power in wage negotiations, creating endogenous real wage rigidity that persists even when labour markets are tight. The threat does not have to be executed to be effective. Employers who can credibly point to automation as an alternative to wage increases have a negotiating tool that depresses settlements below what productivity growth would otherwise justify. That mechanism is now more credible and cheaper to deploy than at any previous point in the technology cycle.
For regulated industries and lower-skill service sectors, the implications are already visible. MIT's November 2025 study, using the Iceberg Index developed with Oak Ridge National Laboratory, estimated that current AI tools can perform tasks tied to 11.7 percent of the US labour market at a cost competitive with human labour. The $1.2 trillion in annual wages that represents is not uniformly distributed. Finance, healthcare, and professional services are most exposed. These are sectors where workers traditionally captured a share of the rents from specialised expertise. AI that commoditises the routine portions of those roles does not just threaten headcount; it reduces the premium that expertise commands by making those tasks accessible to lower-credentialed employees or automated systems at lower cost.
For SF readers, the research raises a question that does not appear in most AI efficiency decks. When a startup sells a workflow automation product that handles the routine portions of a knowledge worker's role, the margin expansion story for the enterprise customer is straightforwardly a wage compression story for the worker whose specialised tasks just became commoditised. The product does not advertise that outcome, but the mechanism is the same. Healthcare AI that automates diagnostic image review reduces the bottleneck that justified radiologist compensation. Legal AI that automates contract review compresses the leverage junior associates had in salary negotiations. The efficiency gain is real. Who captures it is a distributional choice.
Founders building in the AI efficiency category should expect this dynamic to shape regulatory and labour responses in ways that affect their go-to-market. The European Union's AI Act already requires impact assessments for high-risk AI in employment contexts. US labour unions have begun negotiating technology clauses that limit automation without consultation. As the research evidence linking specific automation tools to wage suppression for identifiable worker groups accumulates, procurement decisions will face more scrutiny, particularly in the public sector and in industries with strong collective bargaining. The sales motion for AI efficiency tools is not changing yet, but the political environment around it is.
The research does not argue that automation is inherently bad for workers. Autor and Thompson are explicit that augmentation, automating the easy parts while leaving expert tasks to humans, can raise wages and create more valuable roles. The strategic implication for AI developers is that product design choices have distributional consequences that go beyond feature lists. Founders who understand which tasks they are automating, and what that does to the wage structure of the roles affected, are better positioned to navigate the regulatory and reputational environment that is forming around this question.
Also read: Anthropic's 80x growth projection tests whether safety sells enterprise AI at frontier scale • Scale AI's $500 million Pentagon contract reframes who gets to build America's national security AI stack • Meituan leads Moonshot AI to $20 billion valuation, betting big on Kimi as China's consumer AI gateway
