---
id: "lib-2026-09-00-ai-exposure-and-adoption-among-u-s-firms-ces-wp"
type: "library"
title: "AI Exposure and Adoption Among U.S. Firms (CES WP 26-61)"
lang: "EN"
org: "US Census Bureau, BLS (CES)"
source_type: "Paper"
date: "2026-09-01"
week: "2026-W44"
url: "https://nosov.ai/library/lib-2026-09-00-ai-exposure-and-adoption-among-u-s-firms-ces-wp/"
original_url: "https://www2.census.gov/library/working-papers/2026/adrm/ces/CES-WP-26-61.pdf"
tier: "A"
tier_name: "Strong evidence"
tier_reason: "US Census Bureau and BLS economists with John Haltiwanger link confidential survey microdata to test 14 published AI-exposure measures against reported firm adoption; transparent method and robustness checks, limited to single-establishment firms."
summary_basis: "full-text"
original_minutes: "75"
tags: ["business transformation", "ai adoption", "ai exposure", "firm data"]
---

# AI Exposure and Adoption Among U.S. Firms (CES WP 26-61)

Tier A (Strong evidence): US Census Bureau and BLS economists with John Haltiwanger link confidential survey microdata to test 14 published AI-exposure measures against reported firm adoption; transparent method and robustness checks, limited to single-establishment firms.

Summary based on the full text.

## In short

A Census Bureau Center for Economic Studies working paper (CES 26-61, September 2026, also BLS Working Paper 591) by Matthew Dey, Emin Dinlersoz, Lucia Foster, Cheryl Grim, John Haltiwanger and colleagues. Many studies use occupation-based "AI exposure" scores as a stand-in for AI use. This paper checks that link directly, for the first time at firm level with official data. Exposure predicts adoption, but weakly: most of the difference between firms that use AI and those that do not comes from something other than their occupational mix.

## Key numbers

- 14: occupation-based AI exposure measures from the literature tested against reported firm AI use
- 4-11 pts: higher adoption probability per one-standard-deviation rise in exposure, falling to 1-8 pts with year and industry fixed effects
- 1-6%: share of cross-firm variation in adoption explained by exposure alone
- about 85%: firm-level variation in AI adoption left unexplained even with industry and firm controls
- 8-22 pts: amount by which actual adoption in Information and Professional Services exceeds the exposure-predicted rate

## What it reports

- All 14 exposure measures are positively and significantly associated with adoption in simple regressions.
- One standard deviation more exposure means 4-11 percentage points higher probability of using AI; with year and three-digit industry fixed effects, 1-8 points. For the two headline measures (Eloundou et al. GPT-4 beta and Massenkoff and McCrory observed exposure), the coefficient roughly halves with industry controls.
- Exposure alone explains about 1-6% of cross-firm variation; within detailed industries, 0.3-0.7%. About 85% of variation remains unexplained with all controls.
- Information has the highest adoption and the strongest exposure link; Education Services adopts heavily with a weak link; Accommodation and Food Services shows essentially none.
- Finance and Insurance adopts at the rate its exposure predicts; Information and Professional Services adopt 8-22 points above it.
- Employment-weighted adoption is 9% under the older survey question (AI in producing goods or services) and 27% under the revised one (AI in any business function).

## How they got it

Firm exposure is built from occupation-level exposure scores weighted by each firm's occupational employment in the BLS Occupational Employment and Wage Statistics survey, then linked to AI use reported in the Census Bureau's Business Trends and Outlook Survey and to the Longitudinal Business Database. Samples: about 57,000 firm-year observations on the old question and about 15,500 on the revised one, with propensity-score weights. Linear probability models, heteroskedasticity-robust errors; robustness checks on match quality, firm weights and question version.

## Limits

- Single-establishment firms only (average about 13 staff), which excludes many large firms that adopt AI more.
- Associations, not causal estimates; adoption is a yes/no survey answer.
- Exposure is measured from 2018-2020 occupational mix and adoption from 2023-2025.
- Estimates for the largest size class are imprecise.

## Why it matters for companies

Exposure indices are widely used to predict which roles and firms AI will change. This paper shows they say little about which firms actually adopt: organisational capability, costs and sector matter far more than task mix. For planning, an exposure score marks where AI could apply, not where value or disruption will appear. Firms in sectors whose adoption runs ahead of exposure (information, professional services) set the pace that competitors with similar task profiles are being measured against.

Original: https://www2.census.gov/library/working-papers/2026/adrm/ces/CES-WP-26-61.pdf