Census AI adoption data goes viral — but the methodology is badly flawed
Sep 8, 2025
Key Points
- US Census Bureau data showing declining AI adoption among large firms is spreading widely despite a five-fold gap with private estimates, suggesting the methodology is fundamentally broken.
- Census's two-week usage window captures episodic tool use rather than sustained adoption, and bundles disparate technologies without distinguishing experimental use from strategic deployment.
- Ramp's 43% adoption figure based on paid subscriptions appears more durable than Census's 8% estimate, highlighting how mismatched methodology can turn government statistics into noise.
Summary
The US Census Bureau surveys 1.2 million firms biweekly about AI tool usage, asking whether they deployed machine learning, natural language processing, virtual agents, or voice recognition in the past two weeks. Recent data shows AI adoption declining among companies with more than 250 employees and has spread widely online.
The problem is a massive gap between two estimates of AI adoption. Ramp says 43% of US businesses have paid subscriptions to AI models, platforms, and tools. Census data puts the figure at 8%. A five-fold difference suggests one measurement is fundamentally broken or both are measuring different things entirely.
Census methodology appears weaker. A two-week window is too narrow to capture genuine adoption patterns. Many firms use AI episodically or for specific projects, not continuously. The survey also bundles disparate technologies without distinguishing between experimental use, vendor-mandated integration, and strategic deployment. Small sample sizes within industry segments could amplify noise. Point-in-time snapshots may not reflect sustained behavior.
Ramp's 43% figure rests on paid subscriptions, a clearer proxy for actual commitment. Whether that number is too high depends on how Ramp defines "paid subscription" and whether it includes free trials or bundled tools users don't actively choose. But it measures something more durable than Census's two-week usage window.
The viral spread of Census data despite these flaws shows how government statistics can mislead when methodology mismatches the phenomenon. A declining adoption rate among large firms, if real, would be newsworthy. But the gap between Census and Ramp suggests the survey captures noise rather than signal.