Firm task exposure only weakly predicts reported AI use

Dey, Dinlersoz, Foster, Grim, Haltiwanger, Meyer, Wulff Pabilonia, Pande, Stinson, Wang and Wolf, “AI Exposure and Adoption among U.S. Firms,” CES 26-61, September 2026; full working paper. Cross-links: identification brief, separate graduate study. A collaborating BLS/Census team maps 14 occupation-based measures of potential exposure to the Business Trends and Outlook Survey’s firm-reported AI use; the latter is a binary self-report of use in prior two weeks, with broader wording starting November 2025. They analyze matched single-unit firms: large multi-unit firms, whose establishment occupation mix might not describe the whole company, are excluded.

A one-standard-deviation higher exposure predicts a 4–11 percentage-point greater chance of reported AI use in unadjusted models, 1–8 points after year and detailed industry controls. Exposure alone explains only about 1–6% of cross-firm adoption variation; within-industry explanatory power is lower. It is related to, but nowhere near synonymous with, actual use. At subsector rather than firm level the fit is much higher (a cited earlier study finds 47% variation explained); aggregation can hide firm heterogeneity. The results characterize selected single-unit businesses and do not connect any firm’s AI use to its job applicants or graduates’ wages, prove adoption is causing changes, or show that the graduating-major exposure relationship is as weak numerically as the firm-level relationship. AI may alter expectations or graduate skill signals even before employers adopt it.

Useful skepticism about attributing graduates’ worse first jobs to employer AI substitution; not skepticism about the observed outcomes. A stronger study joins timed use, vacancy/offers, applicant major, remote policy, payroll, prices/output and subsequent earnings, comparing exposed and less-exposed candidates in adopting and nonadopting firms. Both papers are working papers whose interpretations are authors’ own, not the agency’s endorsed view.