Buyer School Research · Methods

Methodology: AI exposure and office-opportunity models

How the two industry measures are built, what each one covers, and what neither can show.

Buyer School modeled estimate. Every score, rank, share, coverage figure and cohort count on this site is calculated by Buyer School from the sources below. Employment totals are BLS statistics.

Both measures start from the BLS May 2025 national staffing file for private four-digit industries and attach a published academic exposure score to each occupation. The exposure ranking averages those occupational scores across an industry's matched employment and ranks 137 industries. The office-opportunity index is a separate construct for 83 mostly in-person industries. The two are not on one scale and should not be merged.

Two constructs, kept apart

Sources and dates

Four different dates apply and they should not be confused.

First published

October 5, 2026

Edition

2.0 (exposure ranking), 1.0 (office opportunity)

Analysis prepared

October 4, 2026

Staffing reference period

May 2025 (BLS OEWS national industry staffing)

Exposure measure vintage

Felten, Raj and Seamans 2021 and 2023; Eloundou and coauthors 2023

The staffing data are from May 2025 and the academic exposure measures are older. Neither is a measurement of current AI capability.

BLS supplies official staffing statistics. The exposure researchers supply original academic measures; these are not government AI-risk indices. Buyer School produced the model. BLS and Census supplied underlying data and did not review or endorse the findings.

1. [S1] U.S. Bureau of Labor Statistics, Occupational Employment and Wage Statistics, May 2025 national industry staffing tables. bls.gov/oes/tables.htm; downloaded file: oesm25in4.zip. Private ownership, detailed occupations, strict four-digit NAICS only.
2. [S2] Felten, E., Raj, M., and Seamans, R. (2021). Occupational, industry, and geographic exposure to artificial intelligence: A novel dataset and its potential uses. Strategic Management Journal 42(12), 2195-2217. doi.org/10.1002/smj.3286. Author data: github.com/AIOE-Data/AIOE. The research is a primary academic source, though not a government measure.
3. [S3] Felten, E., Raj, M., and Seamans, R. (2023). How will Language Modelers like ChatGPT Affect Occupations and Industries? arxiv.org/abs/2303.01157. Language and image exposure workbooks supplied by the authors. The main model uses their language-modeling occupational scores, not their already-aggregated industry rankings.
4. [S4] Eloundou, T., Manning, S., Mishkin, P., and Rock, D. (2023). GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models. arxiv.org/abs/2303.10130. Author replication data: github.com/openai/GPTs-are-GPTs. Sensitivity uses human_rating_beta from data/occ_level.csv, restricted to .00 occupational records; no averaging of O*NET specialties without employment weights.
5. [S5] BLS, 2010-to-2018 SOC crosswalk. bls.gov/soc/2018/crosswalks.htm. Download: soc_2010_to_2018_crosswalk.xlsx. Ambiguous multiple-parent mappings excluded.
6. [S6] Grundy, A., Breaux, C., and Khatiwoda, D., U.S. Census Bureau (May 26, 2026), Large Firms With at Least 20 Employees Biggest AI Users. census.gov/library/stories/2026/05/ai-use-businesses.html. Historical context through May 3, 2026; not a latest October adoption estimate. Question change documentation: AI Question Wording Updates (PDF).
7. [S7] O*NET Resource Center, database and content model. onetcenter.org/database.html. Provides occupational information underlying the academic models. This study does not rerun the original model on the current O*NET release.

How the exposure ranking is calculated

1. Industry universe. Start with the BLS May 2025 national industry staffing file [S1]. The raw file contains 247 industry groups. Keep private-ownership, genuine four-digit NAICS categories and detailed occupation rows only. After removing composites, broader rollups and government designations, 211 industries remain.
2. Match occupations to exposure scores. Join occupational employment to the language-modeling exposure scores of Felten, Raj and Seamans [S2] [S3] through the BLS 2010-to-2018 SOC crosswalk [S5]. A one-parent SOC split inherits the older exposure value. A merged occupation with several older parents is omitted because its employment weights are unavailable.
3. Convert to percentiles. Express each occupation's score as its percentile within the original exposure dataset, on a 0 to 100 scale.
4. Average by employment. Exposure points = sum(matched occupation employment × occupational exposure percentile) ÷ sum(matched occupation employment). Every matched occupation is included and the result is normalized to covered employment.
5. Alternate task model. Repeat with the human_rating_beta occupational data of Eloundou and coauthors [S4], restricted to .00 records, converted to percentiles and weighted by matched industry employment.
6. Apply the coverage gate. Rank an industry only if at least 85% of its employment is matched in both measures. 137 industries qualify and the same 137 are used for the main and alternate ranks.
7. Rank. Higher scores sort first, with minimum ranks for exact ties. The top 25 are selected before displayed points are rounded to one decimal place.

The language measure links AI applications to O*NET abilities [S7]; the task-based alternative uses a different construct. Both rest on occupational foundations, so their agreement is not fully independent validation.

Coverage rule and unranked industries

The 85% threshold is an author-set reporting rule, not a validated confidence threshold. It prevents publication of highly incomplete comparisons but cannot remove selection bias. 74 industries fail it and are listed as unranked with their coverage figures on the data page. Software publishing, computer systems design and travel arrangement are among them. Missing or unranked is not zero exposure: omitted staffing is never assigned a score of zero.

Occupational mergers, newer classifications, unavailable scores and suppressed detailed staffing can all reduce coverage.

Sensitivity checks

The data also carry ranks using the image-generation index and the best and worst rank across the four stated specifications. Image exposure is a different modality, not a newer version of the language model.

Missing-data bounds

The headline score averages observed matched employees. If omitted occupations differ from matched ones, the result can be biased in either direction. Two endpoints make that inspectable: lower endpoint = exposure points × main coverage fraction; upper endpoint = lower endpoint + 100 × uncovered employment fraction, capped at 100. They assign omitted employment the lowest or highest possible percentile. They are conservative sensitivity envelopes, not statistical confidence intervals.

How the office-opportunity index is calculated

1. Office roles. Narrow definition: SOC 43, office and administrative support. Broad definition adds SOC 11 and 13, management and business and financial occupations. Including management is an intentionally generous sensitivity assumption.
2. In-person roles. Narrow proxy: SOC 31, 35, 37, 39, 45, 47, 49, 51 and 53. Broad proxy adds SOC 29, health practitioners, and SOC 33, protective services. Group membership is not proof that every task is physical or protected from AI.
3. Score. Sum occupation employment shares multiplied by their language-exposure percentile across the designated office roles. Multiply that office exposure contribution by the employment share in the designated in-person roles. Express as index points.
4. Cohort. At least 85% matching coverage in both exposure measures, at least half of employment in the broad in-person proxy, and exclusion from the editorial digital-service candidate set. 83 industries qualify.
5. Rank range. Three specifications are reported: narrow office and narrow in-person roles; broad office and broad in-person roles; and the broad roles using the Eloundou alternative. The cohort stays fixed. The range is model sensitivity, not a confidence interval.

Any Census business AI-use figures quoted in the office-opportunity report [S6] are dated broad-sector survey context. They are not four-digit industry adoption rates and are not a component of either score.

Limitations

• Exposure is not a forecast. Neither measure estimates job loss, revenue loss, business failure, savings or any causal effect of AI.
• The exposure measures are older than the staffing data. This is a 2026 publication using May 2025 staffing and earlier academic measures, not an evaluation of current models.
• National staffing covers firms of every size and excludes self-employed workers. It does not isolate small businesses for sale. Owner work, contractors and outsourced support are missing.
• Four-digit categories combine different businesses. No separate HVAC, plumbing or electrical ranking can be drawn from NAICS 2382.
• Some ranked categories are institutional or nonprofit rather than ordinary acquisition targets.
• No range reported here accounts for BLS sampling error, subjective role definitions, staffing change, source age or the evolution of AI capability.
• The work was prepared with AI assistance and programmatic calculation from public sources and has not had independent practitioner or academic review.

Version history and corrections

Report addresses stay stable across editions. Each edition's files are kept unchanged in their own download folder. Corrections will be listed in this table with their date. No corrections have been made.

These counts describe this analysis. It does not cover a 152-industry series and does not rank every U.S. industry.