The Frontline Exposure Gap: Evidence on AI Adoption in Retail and Service Occupations from Task-Level Usage Data

Abstract

This paper measures realized generative-AI adoption in frontline service occupations using task-level and occupation-level data from the Anthropic Economic Index, U.S. employment data, and O*NET. Sales, office and administrative support, food preparation and serving, and personal care and service account for 31.74 percent of U.S. employment but only 11.13 percent of task-matched AI usage, a representation index of 0.351. Correcting occupational classification reduces the administrative-support index from 0.645 to 0.338, the gap remains large across the February and August 2025 releases, and observed frontline use is tilted toward automation-style interactions. New analyses separate the extensive and intensive margins across occupations. In a two-part model covering 732 occupations, a one-standard-deviation increase in computer use is associated with a 16.7 percentage-point higher probability of any observed exposure, while physical presence is associated with a 10.4 percentage-point lower probability. Among 333 occupations with positive exposure, physical presence is associated with 0.544 lower log exposure. Underrepresentation also persists under three definitions of frontline work: the current SOC groups, high physical presence, and high customer interaction yield representation indices of 0.351, 0.104, and 0.348. A potential-versus-observed benchmark identifies larger gaps in computer-intensive and customer-facing occupations, but it does not measure denied access because the two exposure measures use different task universes and methods. The evidence is descriptive and does not identify causal effects on employment, wages, productivity, or access. Data and code: https://github.com/raamnandhakumar-eng/polecoai