750 American executives anticipate productivity gains of 3.0% thanks to artificial intelligence, yet national statistics still capture no measurable effect. Three years after ChatGPT, the American economy is replaying the Solow paradox: “We see AI everywhere, except in productivity figures.”
This dissonance between perceptions and measurements is not unprecedented. Between 1987 and 2000, personal computing had created the same enigma before triggering the productivity boom of the 1990s. AI is following a similar trajectory, but with more pronounced sectoral concentration that delays its macroeconomic visibility.
The Essentials
- American executives anticipate 3.0% annual productivity gains from AI, concentrated in finance and skilled services
- No measurable effect yet appears in the aggregated statistics of the Bureau of Labor Statistics
- AI presents a 78% adoption rate weighted by employment, while executives forecast a 0.7% decline in employment
- The phenomenon reproduces the computing “Solow paradox” of the 1980s-1990s, resolved only after 15 years of diffusion
Executives See What Statistics Haven’t Yet Measured
A survey conducted by the Federal Reserve Bank of Atlanta among 750 American business executives reveals a spectacular gap between expectations and measurable reality. 67% of executives report observing productivity gains linked to AI in their organizations, with an average of 3.0% annual improvement. The finance sector (4.2% anticipated gains) and professional services (3.7%) are leading this transformation.
Yet aggregated data from the Bureau of Labor Statistics detects no significant acceleration of American labor productivity since 2023. Annual productivity growth remains stuck around 1.2%, at its pre-digital revolution level.
This contradiction is explained by the still-emerging nature of AI deployments. “Gains are concentrated in specific activities — code writing, data analysis, automated customer service — that represent limited fractions of total economic activity,” the Atlanta Fed study explains. A programmer who codes 40% faster thanks to Copilot only generates macroeconomic gains if this efficiency spreads beyond his workstation.
AI Transforms Work in a Context of Employment Adjustment
AI adoption now reaches 78% when weighted by employment, illustrating its massive diffusion throughout the American economy. However, this transformation comes with adjustments: executives anticipate a 0.7% reduction in employment in their organizations.
Detailed data shows that 82% of productivity gains come from improved existing performance, while only 18% result from complete task automation. A financial analyst equipped with AI can process 60% more cases, a developer produces code 35% faster, a customer service center resolves 25% more requests.
The energy required for data centers illustrates this technological expansion: more AI means more infrastructure, therefore more technical jobs to build and maintain it.
The Solow Paradox Repeats Itself 40 Years Later
“We see AI everywhere, except in productivity figures.” This rewording of the famous paradox stated by economist Robert Solow in 1987 about computing perfectly captures the current situation. Solow had observed that personal computing, omnipresent in American offices, was generating no measurable productivity gains at the national level.
It took until 1995-2000 for computing to trigger a spectacular acceleration of American productivity, rising from 1.4% annually in the 1980s to 2.9% in the second half of the 1990s. This resolution of the paradox had required three conditions: massive infrastructure investments, reorganization of work processes, and training millions of workers on new tools.
AI is following a similar but accelerated trajectory. AI investments reached 191 billion dollars in the United States in 2025, equivalent to an entire computing decade of the 1980s condensed into a single year. Microsoft, Google, and Amazon are redefining their internal processes around generative AI. Universities are massively integrating AI into their curricula.
The crucial difference: while computing took 15 years to resolve its productivity paradox, AI could do it in 5-7 years according to the Atlanta Fed’s projections.
Sectoral Concentration Delays the Spillover Effect
The macroeconomic invisibility of AI is largely explained by its concentration in high value-added sectors representing only a limited fraction of total employment. Financial services, computing, and professional services — which capture 70% of observed productivity gains — employ only 18% of the American workforce.
Conversely, employment-intensive sectors (retail, food service, transportation, healthcare) are adopting AI more slowly. A server, a delivery driver, or a nurse sees their daily routine little transformed by ChatGPT or Claude. These “physically anchored” professions represent 52% of American employment but capture only 12% of AI-related productivity gains.
This geography of adoption creates a statistical shadow effect. Even 5-6% gains in finance or computing get diluted into the national average dominated by less-affected sectors. The Bureau of Labor Statistics estimates that it would take 8-10% productivity gains in technology sectors to generate a detectable impact of 0.5 percentage points at the aggregate level.
Signals of Acceleration Multiply
Several indicators suggest the macroeconomic effect could emerge faster than expected. The number of companies reporting AI use in their processes rose from 23% in January 2024 to 47% in December 2025 according to the Atlanta Fed survey. This diffusion is accelerating beyond pure technology players.
Corporate spending on AI training surged 340% in 2025. General Motors is training 120,000 employees on generative AI. JPMorgan has deployed 60,000 Copilot licenses. Walmart is testing conversational AI in 2,800 stores. This scaling is progressively transforming entire swaths of the economy.
Most significantly, 34% of surveyed executives plan to expand AI use to operational functions (logistics, production, maintenance) by 2027. This transition from “white collar” to “blue collar” could trigger the missing spillover effect.
Early macroeconomic signals are already emerging. Productivity in the “Professional and Business Services” sector advanced 4.1% in 2025, its best level since 2001. Information and telecommunications show +5.7%. These sectoral performances are beginning to weigh heavily enough to influence national aggregates.
The American economy appears on the verge of resolving its AI productivity paradox. The question is no longer whether gains will appear in the statistics, but when this transformation will become visible across the entire economy.