AI Adoption Reaches 45% of U.S. Workers But Remains Shallow

Generative AI has reached 45% of the U.S. workforce as of May 2026, up from 33% in August 2024, according to new research from the Federal Reserve Bank of St. Louis. But the headline figure masks a more complex reality: while AI tools have spread across most occupations, they remain lightly used within nearly all of them.

The research, based on the Real-Time Population Survey of nearly 14,000 workers, introduces the first nationally representative measures of AI adoption at the detailed task level. The findings challenge assumptions about how deeply the technology has penetrated workplace routines.

Widespread but not deep

At least 20% of workers use AI in more than 80% of occupations tracked by the survey. More than 40% of distinct job tasks show adoption rates above 20%. By that measure, AI has achieved broad distribution across the economy.

Yet only 40% of occupations have adoption rates exceeding 50%, and just 16% surpass 70%. The task-level data is starker: fewer than 3% of job tasks have majority adoption, and none exceed 70%. Even in high-adoption occupations, workers use AI for different parts of their jobs rather than universally applying it to the same activities.

Where AI is—and isn't—used

Computer and information research scientists lead adoption at 87%, followed by information security analysts at 85% and network administrators at 82%. The most AI-assisted tasks are cognitive and information-intensive: reading technical documents, preparing research reports, and analyzing data trends.

Adoption remains lowest in hands-on or face-to-face work. Animal caretakers use AI at just 5%, receptionists at 8%, and licensed practical nurses at 10%. Tasks with zero reported AI use include driving trucks, presenting menus, assisting with medical procedures, and collecting biological specimens—work that requires physical presence or direct human interaction.

Exposure scores miss the mark

Existing research has attempted to predict AI adoption by ranking occupations and tasks by their theoretical "exposure" to the technology. The new data shows these predictions explain roughly half the variation in actual adoption—better than random, but far from definitive.

Medical secretaries and administrative assistants, for instance, were predicted to adopt AI at 61% but actually use it at just 17%, likely due to privacy regulations and error costs. Conversely, computer repairers, special education teachers, and even laundry workers adopt AI at roughly double predicted rates, finding uses in planning, communication, and information-gathering that exposure models didn't anticipate.

More tellingly, exposure scores poorly predict whether any individual worker adopts AI. Demographics like age, education, and gender explain little. Instead, the research points to learning from experience: workers who have used AI for at least six months adopt it for more tasks, and those who use it at work are more likely to use it outside work.

Why it matters

As AI reshapes labor markets, policymakers need accurate data to target workforce policies effectively. Understanding that adoption patterns depend as much on individual learning and experimentation as on job characteristics suggests that training and familiarity—not just technical capability—will determine how deeply AI integrates into work. The gap between what AI can theoretically do and what workers actually use it for reveals barriers beyond technology itself, from regulatory constraints to the costs of making mistakes.

The findings were first reported by the Federal Reserve Bank of St. Louis and the Washington Center for Equitable Growth, with data publicly available for researchers studying AI's labor market effects.