On August 27, 2026, the Bureau of Labor Statistics published its first occupational AI exposure measure. The BLS categories and the Opportunity Data AI Exposure Index score the same occupations from different signals: one from what AI can do and how it is used, one from what the work itself consists of. This comparison measures where the two agree and names the mechanism behind where they part.
The BLS AI Exposure Categories accompany the 2025-35 employment projections: each of 831 National Employment Matrix occupations is assigned one of four levels, Low, Moderate, High, or Very high. The categories synthesize five external sources: three research measures of theoretical exposure, in which human raters or language models assess AI capability against O*NET tasks and abilities, and two measures of observed usage, built from Anthropic Claude conversation data and Microsoft Copilot interactions mapped to O*NET tasks and work activities.
BLS states two limits that matter for reading this page: an exposure category "is not a forecast of employment growth or decline," and the categories do not distinguish AI that replaces work from AI that assists it.
Capability-based measures are snapshots: they can only know the AI of their collection window. Inside the BLS categories, the three theoretical inputs rate AI as it stood between the 2010s benchmark era and early 2023, one of them using GPT-3.5 as the rater, and the two usage windows close in January 2025. BLS states this limitation itself. None of the five inputs observed the frontier that was evident by spring 2026, Claude Opus 4.8 and beyond: agentic systems that write and run software, operate computers, and carry multi-step work end to end. The scale of that break is documented by independent measurement: Stanford's AI Index 2026 (April 2026) reports agent success on real-world tasks rising from 20 percent to 77 percent in a year on Terminal-Bench, and METR, whose task-length measure had doubled every seven months for years and then every four, found the strongest 2026 agents at the edge of what its evaluation could reliably measure.
Capability or data window of each input to the BLS categories, against the frontier the Opportunity Data AI Exposure Index was designed on. Sources: the BLS methodology note and each study's stated data period.
The Opportunity Data AI Exposure Index is built on the opposite assumption. It does not measure AI capability, so it does not age with it. By April 2026, frontier systems made one design choice defensible: treat current capability as sufficient for digitally mediated tasks, and measure what binds when capability does not. That residual is structural: hands-on work, physical presence, face-to-face care, responsibility for others' safety, and in practice regulation. The assumption has stated boundaries, and the AI Exposure Index monitors both: the physical anchor holds in the absence of capable AI-powered robotics, and the human-contact shield holds while institutions and preference keep people in the loop. A capability snapshot must be recollected every time models move; a structural measure moves only when the work itself does.
| Opportunity Data AI Exposure Index | BLS categories | |
|---|---|---|
| Philosophy | Assume the AI, measure the work: treat Opus 4.7-class capability and beyond as given, and score the O*NET work characteristics that still separate exposed from protected occupations when capability is not the constraint | Measure the AI: observe what models before 2025 could do and which tasks people worked on with them, then rank occupations by that snapshot |
| Built from | Nine O*NET work characteristic variables reported by incumbent workers: what the work physically and socially consists of | Five external research and usage measures of what AI models can do and what people use them for, mapped onto O*NET tasks, abilities, and work activities |
| Form | Continuous 0-1 score with three named dimensions | Four ordinal levels |
| Coverage | 911 occupations, 1,806 academic programs, state portfolios | 831 occupations, national |
| What they share | Both stand on O*NET's description of what each occupation does. The exposure signals are independent: no AI capability or usage measure enters the Opportunity Data AI Exposure Index, the AI Exposure Index is not an input to BLS, and none of the AI Exposure Index's nine variables serves as an exposure signal in any BLS source. Agreement between the two is therefore informative rather than mechanical. | |
The digital core of the economy ranks the same on both measures. Of the 761 occupations both score, the rank correlation is 0.76, 43 percent fall in the same tier, and 86 percent fall within one tier. The occupations both measures place at the top are the quantitative and software fields: actuaries (index score 0.85), market research analysts and marketing specialists (0.84), software developers (0.83), economists (0.82), data scientists (0.81), and database architects (0.80) are all rated Very high by BLS and Highly exposed by the AI Exposure Index.
| BLS category | Index: AI-protected | Moderate | AI-exposed | Highly exposed | Occupations |
|---|---|---|---|---|---|
| Low | 83% (167) | 17% (35) | 0 | 0 | 202 |
| Moderate | 60% (110) | 40% (72) | 0 | 0 | 182 |
| High | 25% (46) | 49% (92) | 22% (42) | 4% (7) | 187 |
| Very high | 1% (2) | 31% (58) | 44% (84) | 24% (46) | 190 |
Each row shows where the Opportunity Data AI Exposure Index places the occupations BLS assigns to that category; rows sum to 100 percent, shading follows the share, and counts are in parentheses. The BLS levels are relative (occupations compared with one another) while the AI Exposure Index tiers are fixed score thresholds, so the two grids are calibrated differently; the rank correlation is the comparison that does not depend on calibration.
Two features of this table carry the comparison. First, the empty upper right: none of the 384 occupations BLS rates Low or Moderate reaches either exposed tier on the Opportunity Data AI Exposure Index, so every serious disagreement runs in one direction, which is the signature of a stricter test rather than measurement noise. Second, the Very high row: BLS's top category splits 31, 44, and 24 percent across the AI Exposure Index's moderate, exposed, and highly exposed tiers, meaning the single BLS bucket contains structurally different populations, the digital core alongside high-usage, high-contact fields, and the structural measure separates them.
106 occupations sit two or more tiers apart, and every one of them diverges in the same direction: BLS rates the occupation High or Very high while the Opportunity Data AI Exposure Index rates it Moderate or AI-protected. The divergent set is not random. It concentrates in teaching at every level, nursing and medicine, counseling, clergy, and front-line sales and service work, and it employs 37.9 million people, about a quarter of the workforce compared here.
| Occupation | Employment, 2025 | BLS category | Opportunity Data AI Exposure Index tier | Index score |
|---|---|---|---|---|
| Retail salespersons | 3,998,100 | Very high | AI-protected | 0.39 |
| Registered nurses | 3,465,400 | High | AI-protected | 0.43 |
| Elementary school teachers | 1,419,300 | High | AI-protected | 0.41 |
| Educational and career counselors | 389,500 | Very high | Moderate | 0.57 |
| Nurse practitioners | 336,300 | High | AI-protected | 0.43 |
| Genetic counselors | 4,200 | Very high | Moderate | 0.59 |
| Flight attendants | 133,700 | High | AI-protected | 0.31 |
| Family medicine physicians | 114,300 | High | AI-protected | 0.38 |
Eight of the 106 divergent occupations, chosen for employment size and range. Employment is BLS 2025 National Employment Matrix.
The mechanism is visible in the constructions. Two of BLS's five sources measure observed usage, and teachers, clinicians, and counselors are heavy users of AI at work, so usage-based inputs rate these fields highly exposed. The Opportunity Data AI Exposure Index instead measures the structure of the work: its Human Contact dimension treats assisting and caring for others, working with the public, and face-to-face discussion as protective, and these are the highest-contact occupations in the data. The disagreement is therefore a measured version of the distinction BLS says its categories cannot make: exposure through use is not the same as exposure to replacement. Which reading is right for a given decision depends on the question; the divergent set is exactly where the question matters.
Neither measure forecasts employment. BLS's categories carry the strength the Opportunity Data AI Exposure Index lacks, grounding in observed use of current AI systems; the AI Exposure Index carries structure the categories lack, continuous scores, named dimensions, and the program and state layers that connect occupations to education pipelines. Where both measures rate an occupation highly exposed, the signal is strong from two independent directions. Where they diverge, the divergence itself is the finding: work that uses AI heavily but is anchored by human contact.
Opportunity Data AI Exposure Index scores are from the current data (O*NET 31.0, OEWS May 2025); construction is documented in the methodology. BLS categories, employment, and projections context are from the BLS AI Exposure Categories file (August 27, 2026). Index scores were aggregated to 6-digit SOC codes and joined to the 831 BLS occupations; 761 matched. The full concordance table, all 761 occupations with both measures, scores, employment, and wage context, is provided on request.
Suggested citation: Rowe, B. (2026). The Opportunity Data AI Exposure Index and the BLS AI Exposure Categories: a comparison. Opportunity Data. https://opportunitydata.org/ai-exposure-bls-comparison.html
© 2026 Opportunity Data.