---
id: "lib-2026-05-27-artificial-intelligence-productivity-and-the"
type: "library"
title: "Artificial Intelligence, Productivity, and the Workforce: Evidence from Corporate Executives (May 2026 revision)"
lang: "EN"
org: "Federal Reserve Bank of Richmond / Duke"
source_type: "Paper"
date: "2026-05-27"
week: "2026-W22"
url: "https://nosov.ai/library/lib-2026-05-27-artificial-intelligence-productivity-and-the/"
original_url: "https://www.richmondfed.org/-/media/RichmondFedOrg/research/national_economy/cfo_survey/academic_publications/AI_survey.pdf"
tier: "A"
tier_name: "Strong evidence"
tier_reason: "Economists at the Atlanta and Richmond Federal Reserve Banks and Duke's John Graham survey 734 financial executives through The CFO Survey, with weighting, open-ended coding and a productivity decomposition; strong, transparent evidence, limited by being self-reported attributions at an early stage."
summary_basis: "full-text"
original_minutes: "60"
tags: ["executive survey", "productivity", "workforce"]
frames: ["solow-paradox"]
---

# Artificial Intelligence, Productivity, and the Workforce: Evidence from Corporate Executives (May 2026 revision)

Tier A (Strong evidence): Economists at the Atlanta and Richmond Federal Reserve Banks and Duke's John Graham survey 734 financial executives through The CFO Survey, with weighting, open-ended coding and a productivity decomposition; strong, transparent evidence, limited by being self-reported attributions at an early stage.

Summary based on the full text.

## In short

The May 2026 revision of a paper by Salomé Baslandze, John Graham and colleagues at the Federal Reserve Banks of Atlanta and Richmond and Duke University (also issued as NBER w34984). It asks CFOs directly how much AI has changed and will change their employment, workforce mix and productivity. Adoption is broad but shallow and mostly subscription-based; net employment effects are small, negative at large firms and slightly positive at small ones; productivity gains come mainly from new products and serving customers better, not from cost-cutting.

## Key numbers

- 734: financial executives surveyed near year-end 2025 (The CFO Survey plus FEI, Nasdaq and Duke panels)
- less than 0.4%: expected AI-driven decline in US employment in 2026, implying 501,836 fewer jobs (95% CI 262,279 to 741,394)
- -0.70%: expected 2026 AI-attributed change in total employment at large firms (500+ staff)
- about 0.6 percent: AI-attributed labour productivity growth (revenue per employee) already realised in 2025
- more than 2 percentage points: expected fall by 2028 in the share of staff doing routine clerical work, mostly at large firms

## What it reports

- Spending: median expected 2026 AI spend $12,500 at small firms and $300,000 at large firms; about 30% of large firms plan to spend over $1 million. Operational spending (subscriptions, services, training) is 65% at small firms, 54% at large.
- Barriers for non-investors: immature technology, untrained workforce, privacy, uncertain capabilities.
- Roles: 50% of firms say AI will replace no roles, 6% are unsure, 44% describe some replacement.
- Employment: large firms expect -0.7% in 2026 due to AI; aggregate implied reduction 501,836 (under 0.4% of the workforce), concentrated in high-skill services (-271,290) and finance (-111,243).
- Composition: routine clerical share expected to fall more than 2 points by 2028; skilled technical roles rise, mainly at small firms, suggesting reallocation across firms more than within them.
- A Negative Exposure Index from open-ended answers shows office and administrative support most often replaced; professional, technical and sales roles mostly enhanced.
- Productivity: implied 2026 gains of 1.62% (small) and 2.41% (large); above 2% in high-skill services and finance. Little comes from capital deepening, because AI is mostly rented.
- Perceived productivity gains exceed those implied by revenue and employment data, a "productivity paradox" the authors attribute to delayed revenue realisation.

## How they got it

Two surveys fielded near year-end 2025: The CFO Survey (Duke, Richmond Fed, Atlanta Fed) across 48 states and all major nonfarm private industries, plus FEI, Nasdaq and Duke alumni panels. Categorical answers are converted to midpoints; results use representativeness weights (Census firm counts by sector and size) and importance weights (2024 employment). Open-ended answers are mapped to BLS SOC groups. Revenue-per-worker growth is decomposed into capital deepening and residual TFP. De-identified data are on a public dashboard.

## Limits

- AI attribution is the executive's own judgement, not measured causally.
- Executives are mostly finance leaders, who may see cost effects more clearly than workflow changes.
- Early-stage snapshot; agentic tools deployed in 2026 are largely outside the window.
- US only.

## Why it matters for companies

This is one of the most credible answers to "what are other companies actually doing with AI and headcount": most expect small net effects, and larger firms expect cuts concentrated in clerical and administrative work. The finding that productivity gains come from product and customer channels rather than labour savings is a warning against business cases built only on headcount reduction. The gap between perceived and measured productivity also argues for tracking revenue and quality outcomes, not executive impressions.

Original: https://www.richmondfed.org/-/media/RichmondFedOrg/research/national_economy/cfo_survey/academic_publications/AI_survey.pdf