A regional Fed survey now asks CFOs what they spend on AI

The setup

The CFO Survey is a quarterly panel run jointly by the Federal Reserve Banks of Richmond and Atlanta with Duke University's Fuqua School of Business, a continuation of a Duke research panel that predates it by nearly 25 years. Panel members are recruited mainly by telephone through the Atlanta Fed's outreach lab, using purchased business-contact lists sorted to represent U.S. industries by their contribution to GDP, according to the survey's own methodology documentation. Special questions rotate in at the end of the core questionnaire, and the fourth-quarter 2025 wave carried one on artificial intelligence spending and impact, retrieved as published 16 September 2026.

What the documents show

That Q4 2025 release states that over half of surveyed firms had made AI-related expenditures in the prior 12 months, with adoption reported at 77.7% among large firms against 48.3% among small firms, and that respondents expected to spend more on AI in the coming year than in the last one. On effects, the survey's own summary is notably restrained: most firms reported no change yet in labor productivity, decision-making speed or how tasks get allocated, with only small improvements reported by some, though a larger share expected improvement looking ahead.

The friction

The survey's public methodology document is candid about its own limits: panel members come from purchased contact lists rather than a full census, response is voluntary, results are weighted by industry, firm size or revenue to correct for uneven response rates, and members who miss three straight quarters get dropped from the panel. None of that invalidates the AI figures, but it means the reported adoption share describes financial executives who agreed to join and kept responding to a Fed-and-Duke panel, not a random draw of every U.S. company.

What changed in the work

What a module like this adds, compared with a single vendor's usage dashboard, is a business-conditions source explicitly designed for macroeconomic research and cross-checked against other Fed surveys, run by people with no product to sell. Reading a single special-question wave as an established trend goes beyond what the survey itself claims, since special questions rotate and are not asked every quarter in identical form.

• Was this AI question asked with the same wording in a prior quarter, or is a like-for-like comparison actually possible?
• Does the reported adoption figure come from firms weighted for size and industry, or a raw unweighted count?
• Does 'spending on AI' in this context mean a purchased subscription, an internal build, or either?

Because CFO Survey special questions rotate, the AI module may not reappear next quarter in the same form, and any comparison across waves should be checked against the specific question wording the survey team used each time, rather than assumed to be constant.