American labor productivity (non-agricultural sector) rose from +1.6% in 2023 to +2.3% in 2024, according to the Bureau of Labor Statistics. This acceleration coincides with the massive diffusion of AI in companies, even though the causal link remains unproven at the aggregate level according to available NBER studies. The gains exist. They are measurable. But surveys conducted among American business leaders reveal that they are first concentrated in margins, with a lag of two to four quarters before wages follow. The question is not whether progress is happening. It is to understand who sets the conditions for sharing it, and whether these conditions are changeable.

The Essentials

  • American labor productivity increased from +1.6% in 2023 to +2.3% in 2024 (Bureau of Labor Statistics, non-agricultural sector)
  • NBER surveys of business leaders confirm that gains are concentrated first in profits, with a lag of two to four quarters before wages follow
  • The most affected sectors are skilled services and finance, where AI assists or replaces repetitive analytical tasks
  • American median real wages advanced modestly in 2024, far behind the productivity curve
  • The transmission of gains to workers depends less on an automatic mechanism than on policy choices regarding corporate taxation, wage bargaining power, and training investments

The comparison that best illuminates the current American situation is that of the 1990s. Between 1996 and 2004, the United States experienced a similar acceleration in productivity linked to computerization. Wages had followed then, but with a lag of several years, and unevenly. Less-skilled workers had benefited little from the technological boom. The distribution of gains had concentrated toward the top of the qualification ladder. Today, economists at the Brookings Institution who track the same cycle warn against a repeat: the mechanisms that block transmission exist, they are known, and they do not correct themselves.

Productivity Accelerates, Data Confirms It

The acceleration measured in 2024 is not a statistical artifact. The Bureau of Labor Statistics publishes quarterly measurements of productivity by sector. The increase to 2.3% corresponds to a visible trend in most cognitively intensive sectors: business services, finance, insurance, legal counsel. These are precisely the sectors where adoption of generative AI tools has been fastest since 2022.

NBER surveys provide a qualitative dimension to these figures. Among the leaders surveyed, a significant share reports having observed productivity gains in their teams using AI. But an equally significant share indicates that these gains were first absorbed by improving operating margins, before considering wage increases. This sequencing reflects understandable business logic: you first improve the bottom line before redistributing. The problem is that this lag, averaging two to four quarters according to statements, can stretch if no external pressure is applied.

It should be noted in this regard that an NBER study of some 6,000 leaders reveals that 9 out of 10 observe no impact of AI on their company’s productivity over the past three years, which underscores how contested the causal link between AI adoption and aggregate productivity gains remains. Furthermore, according to BLS data for 2024, real hourly compensation advanced by approximately 2.0%, or about 87% of productivity progress (2.3%), a gap much smaller than crude correlation might suggest. The surplus exists, but the mechanisms of its distribution merit close scrutiny.

The Lag Is Not a Natural Law

Classical economics teaches that total factor productivity eventually benefits all factors of production, labor included. In a perfectly competitive labor market, companies that achieve productivity gains must raise wages to attract and retain talent. The 2024-2026 American reality complicates this mechanism at several points.

The first friction point is market concentration. Thomas Philippon has documented, in his work on the “great reversal” of American competition, how American companies have lost their competitive character since the 1990s. Entire sectors are dominated by a few players who have no need to outbid on wages to recruit. The competitive pressure that would normally have transmitted productivity gains to wages has weakened.

The second point is the structure of bargaining power. The American unionization rate hovers around 10%, one of the lowest among advanced economies. Workers who achieve productivity gains thanks to AI rarely have the institutional tools to formally claim a share. In Northern European economies, where collective bargaining remains strong, comparative OECD studies show that the transmission of productivity gains to wages is structurally faster.

The third friction is fiscal. Share buybacks and dividends, which have been tax-advantaged in the United States since the 2017 reform, offer companies a profitable alternative to wage increases for using their surplus. In 2024, share buybacks by S&P 500 companies reached record levels, according to S&P Global data. This allocation choice is not irrational from a shareholder perspective. But it confirms that the sharing of productivity gains results from a set of rules of the game, not an automatic mechanism.

Substitution or Complementarity: The Question Becomes Central Again

Axelle Arquié, economist specializing in AI and the labor market, posed this question in her recent work: does AI substitute for workers, or does it complement them? The answer is not the same depending on sectors, tasks, and qualification levels. 2024-2026 American data allows us to begin answering through figures.

In skilled services and finance, the two sectors that NBER surveys identify as most transformed, AI seems primarily complementary. It accelerates analytical tasks, reduces document production time, improves forecast precision. Jobs do not disappear massively in these sectors; they transform. Professionals using these tools see their output increase, but their numbers do not decrease proportionally. This is precisely the mechanism that generates the measured productivity gains.

But this complementarity has a downside. It concentrates gains on already-qualified workers, already well-paid, already in dynamic sectors. For workers whose tasks are more routine, in logistics, customer service, data entry, the pressure toward substitution is stronger. The risk is not homogeneous mass unemployment, but a bifurcation: significant gains for those who master the tools, stagnation or downward mobility for those who lack access. This is a question that autonomous agents pose even more acutely, as we analyzed in our article on the new single-operator entrepreneurial model.

What Concretely Blocks Transmission

The Brookings Institution analysis identifies three specific mechanisms that slow the transmission of productivity gains to wages in the current American context.

The first is a training deficit. Productivity gains linked to AI require new skills. Companies that invest in training their employees improve both their productivity and retention. But this investment remains unequal. Large technology and finance companies train massively; SMEs, which employ the majority of American workers, invest much less in upskilling. This differential widens the gap between workers who capture AI gains and those who suffer its effects without benefiting from them.

The second mechanism is the absence of sharing clauses. In some European collective contracts, automatic productivity gain-sharing mechanisms exist: a portion of performance improvement translates directly into wage increases or reduced working time. In the United States, such arrangements are rare and often reserved for sectors still strongly unionized, like parts of the automotive industry. For the vast majority of workers, individual negotiation remains the norm, and it structurally favors the employer when the labor market loosens slightly.

The third mechanism is the timing of capital investments. When a company deploys an AI tool, it incurs an immediate cost: licenses, infrastructure, initial training, reorganization. Productivity increases, but it first serves to amortize this investment before generating a redistributable surplus. This cycle is not manipulation; it is the normal logic of return on investment. But it explains the two-to-four-quarter lag observed in NBER surveys, and it signals that the redistribution window is approaching, but requires being anticipated by public policy to avoid being once again absorbed by the shareholder.

The Levers Exist, Some Have Already Worked

American economic history offers useful precedents. After World War II, the rise in productivity linked to mechanization of industry had been accompanied by sustained wage increases. This transmission was not spontaneous: it resulted from powerful unions, progressive fiscal policy, and a labor market tightened by full employment. Conditions have changed, but the levers remain identifiable.

On taxation, several economists, including Dani Rodrik in his work on “good jobs,” argue for redirecting incentives toward companies that invest in their employees rather than shareholders. Tax credits conditional on wage increases or training investment are tools already experimented with at the state level. Massachusetts and Washington tested such mechanisms with encouraging results according to early evaluations.

On training, the issue is access. Federal government-funded retraining programs remain undersized relative to the scale of the transformation underway. Brookings estimates that public investments in vocational training should be tripled to keep pace with AI deployment in companies. The Biden Administration had initiated programs in this direction via the Chips and Science Act; their continuity under the current administration remains uncertain.

On competition, enforcement of antitrust law by the FTC under Lina Khan, then by her successors, has begun raising questions about concentration in sectors most affected by AI. More active competition in digital and financial services would mechanically exert upward pressure on wages. This lever is slow, but structural. The governance of AI agents themselves poses similar questions about concentration of economic power, as analyzed in this article on governance challenges of autonomous agents.

The Precedent of the 1990s and What It Changes

Carl Benedikt Frey, whose work on technology and employment is authoritative, has shown that major technological waves do not resemble each other in their distributive effects. Nineteenth-century mechanization first destroyed real wages for several decades before they rebounded. Computerization in the 1980s-1990s was less destructive, but widened qualification inequalities. Generative AI from 2022-2026 has this particular feature: it affects qualified cognitive tasks, not only manual or routine ones.

This shift upward on the qualification ladder changes the necessary economic policy. When technology affects unskilled workers, classic responses are vocational training and social safety nets. When it affects intermediate professions and managers, the response must be more sophisticated: retraining toward high-value relational and creative tasks, work-time sharing, new forms of ownership over digital tools. These responses exist in American public debate, particularly among economists like Daniel Susskind or Diane Coyle who think about measuring and sharing value in the digital age. They have not yet found legislative translation at the federal level.

The American dynamic illustrates a broader phenomenon. Productivity gains linked to AI are real, measurable, and probably lasting. But their distribution is not automatic. It results from power relations, fiscal rules, training investments, and bargaining structures. The surplus exists. The question now is whether American institutions are capable of devising, within the two-to-four-quarter timeframe companies give themselves, the mechanisms that will allow the greatest number to benefit before the gains are entirely absorbed elsewhere.

Sources

  1. Brookings Institution, “AI Growth Acceleration versus Distributional Fairness”, https://www.brookings.edu/articles/ai-growth-acceleration-versus-distributional-fairness/
  2. Bureau of Labor Statistics, Productivity and Costs, https://www.bls.gov/productivity/
  3. Federal Reserve Bank of Atlanta, Wage Growth Tracker, https://www.atlantafed.org/chcs/wage-growth-tracker
  4. NBER, survey of 750 American leaders on AI gains, March 2026, National Bureau of Economic Research (URL not available at publication date)
  5. Thomas Philippon, The Great Reversal: How America Gave Up on Free Markets, Harvard University Press, 2019
  6. Dani Rodrik, work on “good jobs” and industrial policy, https://drodrik.scholar.harvard.edu/
  7. S&P Global, data on S&P 500 share buybacks, 2024, https://www.spglobal.com/
  8. OECD, reports on collective bargaining and transmission of productivity gains, https://www.oecd.org/employment/
  9. BLS Productivity 2024, official annual figure, https://www.bls.gov/opub/ted/2025/productivity-up-2-3-percent-in-2024.htm
  10. NBER Working Paper 34984, survey of 750 CFOs, March 2026, https://www.nber.org/papers/w34984
  11. NBER Working Paper 34836, survey of 6,000 leaders, AI and productivity, https://www.nber.org/papers/w34836
  12. Atlanta Fed Wage Growth Tracker, https://www.atlantafed.org/research-and-data/data/wage-growth-tracker
  13. BLS Union Membership 2024, https://www.bls.gov/opub/ted/2025/union-membership-rates-highest-in-hawaii-and-new-york-lowest-in-north-carolina-in-2024.htm
  14. Philippon, The Great Reversal, Harvard University Press, https://www.hup.harvard.edu/books/9780674260320
  15. Brookings Institution, AI and the labor market 2026, https://www.brookings.edu/articles/research-on-ai-and-the-labor-market-is-still-in-the-first-inning/
  16. Chicago Fed — IT and productivity, 1990s — https://www.chicagofed.org/publications/chicago-fed-letter/2003/september-193