Fed research reveals 80% of firms report no AI productivity impact — but the headline misleads
Feb 25, 2026
Key Points
- A Fed-backed survey finds 80% of firms report no AI productivity impact, but the headline conflates "no impact" with "hasn't shaped hiring yet"—70% actively use AI while averaging only 1.5 hours per week of executive engagement.
- Executives forecast just 1.4% productivity gains over three years and expect AI to create more jobs than it destroys, contradicting Silicon Valley narratives that half of white-collar work will vanish.
- If policymakers cite low adoption numbers as evidence of an AI bubble, executives may cut AI investment, handing competitive advantage to faster-moving startups over incumbents.
Summary
A Federal Reserve-backed survey showing 80% of firms report no AI productivity impact is being misread and risks shaping policy based on flawed data.
The paper, authored by researchers from the Atlanta Fed, National Bureau of Economic Research, and central banks in the UK, Germany, and Australia, surveyed 6,000 verified business leaders across four countries. Unlike typical online surveys that pay random respondents $10 to claim CFO status, this study ID-verified participants and reality-checked their job titles.
The headline stat conflates "no impact" with "hasn't shaped hiring plans yet." The actual findings are messier. 70% of firms actively use AI, yet average executive use is only 1.5 hours per week. One quarter of executives report zero AI use. Text generation via LLMs is the most common use case at 41% of firms, meaning 59% aren't using LLMs for text generation or proofreading at all.
Measurement problems
Many executives don't realize they're using AI because it's embedded in SaaS tools they already use daily. Payment processing software with AI image generation, customer support agents indistinguishable from humans, and AI queries made on weekends about work-related problems don't register as work AI adoption in the executive mind. The survey's definition of "AI adoption" is broad enough that any machine learning, LLMs, visual content creation, or robotics counts, which inflates the adoption figure to 78% of U.S. firms.
Managers forecast only 1.4% productivity gains over three years. Economic researchers characterize this as sizable, but it underwhelms if you believe frontier AI narratives. More striking: 63% of firms expect no employment impact from AI. This contradicts Silicon Valley claims that 50% of white-collar work will disappear. Most managers believe AI will create more jobs than it destroys, even as some roles become obsolete.
The token consumption puzzle
The disconnect between reported low AI impact and high token consumption at labs like Stripe and Anthropic suggests executives either don't recognize the leverage they're getting or are downplaying adoption in surveys. One hour of prompting in 2026 generates far more output than the same time would in 2023. Waiting time, agent overnight work, and asynchronous processing don't register as "hours using AI" when executives self-report.
Policy risk
This paper will circulate within the Fed and likely be cited in policy discussions and New York Times coverage of an "AI bubble." If executives see low adoption numbers in authoritative research, they may deprioritize AI investment, creating a self-reinforcing slowdown. Meanwhile, faster-moving competitors will gain ground, which favors startups over incumbents.
Better measurement
Instead of asking CEOs how much AI they use, ask them how much they think competitors use. Neighbor polling worked better than direct polling in presidential elections by removing the stated-preference bias. Corporate executives have strong incentives to claim AI adoption. They have less incentive to underestimate rivals.