The Federal Reserve Board published a survey reconciliation paper on 3 April 2026, authored by staff economist Jeffrey S. Allen for the FEDS Notes series, comparing three separate federal instruments that should describe the same US economy and do not. The Business Trends and Outlook Survey (BTOS), run by the Census Bureau on a firm-weighted basis, put AI adoption at 18% for late 2025. The Research on Practices Survey (RPS), which asks individuals whether they personally use AI at work, returned 41%. The Survey of Business Uncertainty (SBU), which weights by how many workers are employed at AI-using firms, came in at 78%. Daily AI use across the US workforce sits at 12%; weekly use at 35.2%.
The three figures measure different things: the share of firms using AI, the share of workers personally using AI, and the share of the workforce employed at firms that have adopted it. The Fed paper's point is not that the surveys are faulty; it is that no federal agency has formally chosen between firm-weighted, individual-weighted and employment-weighted units, so the same quarter can be described as roughly one-in-five, two-in-five or four-in-five AI-adopted depending on which federal instrument is cited.
Forty-six days earlier, the bipartisan nine-senator coalition led by Josh Hawley and Mark Warner had written to the Department of Labor and the Bureau of Labor Statistics urging expanded AI workforce data collection . The Fed Board's reconciliation is effectively the federal answer: there is no single figure, and none of the three existing instruments will produce one until an agency chooses a canonical unit. The BLS itself has so far chosen none, skipping a separately scheduled GenAI workplace publication eleven days later.
In practice, private datasets set the headline number by default. Challenger, Gray & Christmas counts announced AI layoffs; Stanford's JOLTS-based analysis puts the real labour impact at a far higher multiple; Goldman Sachs's earlier monthly substitution model sat between the two. None of those are federal. None can be legislated against without a statutory benchmark that the reconciliation paper has just confirmed does not exist.
The 4.3x divergence is also the quantitative foundation for the Hawley-Warner demand: a letter asking for one number has been met by a paper documenting that the government currently produces three, none of them designated as authoritative. Whether the BLS is resourced to build a fourth, harmonised instrument, or whether the NY Fed's Survey of Consumer Expectations becomes the de facto federal measure by attrition, is likely the first concrete thing the next US Congress will have to decide on AI workforce policy.
