BLS 2025-35 Projections: AI Now Has an Official Exposure Score

For the first time, the United States government has assigned every occupation it tracks an official AI exposure score. On August 27, 2026, the Bureau of Labor Statistics released its 2025–35 Employment Projections — and tucked alongside the usual growth tables was a brand-new data product that sorts 831 detailed occupations into four AI exposure categories, built in part on usage logs from Claude and Microsoft Copilot. If you have ever wondered whether the government thinks AI can do your job, there is now an official answer. It comes wrapped in more warnings than any data product BLS has shipped in years — and those warnings are half the story.

A Decade of Much Slower Job Growth

Start with the headline. [Fact] BLS projects the US economy will add 5.9 million jobs between 2025 and 2035, taking total employment from 170.3 million to 176.2 million — growth of 3.5%. [Fact] The decade that just ended, 2015–25, recorded 10.9% growth.

Run the arithmetic on those two figures and the contrast sharpens. [Estimate] A 10.9% gain implies the previous decade added roughly 16.7 million jobs; the coming decade is projected to add about 35% of that — barely a third of the pace, in absolute terms.

Before anyone blames the robots: this slowdown is not primarily an AI story. The release attributes its growth patterns overwhelmingly to demographics — an aging population, chronic disease prevalence, slower labor force growth. Healthcare and social assistance alone is projected to account for [Fact] 37.0% of all new jobs through 2035, adding over 2.2 million positions. The fastest-growing detailed occupation is nurse practitioners at [Fact] +41.0% — work built on physical presence and clinical judgment, about as far from a chatbot as a job gets.

AI Shows Up on Both Sides of the Ledger

What makes this release different is where AI appears in the causal narrative. It is now on both sides.

On the growth side: [Fact] utilities is projected to be the fastest-growing major sector at +9.8%, and [Claim] BLS attributes nearly all of that to electric power generation and transmission — including AI's power demands and data center buildout. [Fact] The computing infrastructure, data processing, and web hosting industry is projected to grow 25.1% and add 120,400 jobs. [Fact] Data scientists (+34.6%) and computer and information research scientists (+21.8%) both rank among the ten fastest-growing occupations. [Fact] Professional, scientific, and technical services is projected to add 926,700 jobs — the second-largest gain of any sector — [Claim] driven by demand for AI-based systems, R&D, and consulting.

On the decline side: [Fact] office and administrative support occupations are projected to shrink 4.0% and shed 752,100 jobs — the largest decline of any major occupational group, and [Claim] BLS points to automation tools "including those powered by AI" as a key reason. [Fact] Sales occupations are projected to fall 1.4% and production occupations 0.4%. [Claim] Generative AI may also limit demand in arts, design, entertainment, sports, and media roles.

Same technology, both columns of the spreadsheet.

One honest complication: office and administrative support has been shrinking for decades — spreadsheets, e-filing, and enterprise software were eroding these jobs long before large language models existed. The −4.0% figure is not a clean measurement of AI's bite; it is the continuation of a 40-year automation trend that AI now accelerates. Treating it as a pure AI signal would overstate the case.

Where the New Exposure Categories Come From

The genuinely new artifact is the AI exposure categories product. BLS combined five external data sources into two dimensions for every occupation, then used a clustering algorithm to assign each one of four labels: [Fact] Low, Moderate, High, or Very high relative AI exposure.

The two dimensions matter more than the labels. The first is theoretical exposure — whether AI technology could assist or complete an occupation's work — built from three academic measures: Felten, Raj, and Seamans' ability-based survey scores, Eloundou et al.'s "GPTs are GPTs" task ratings, and Eisfeldt et al.'s GPT-based task scoring. The second is observed exposure — what AI is actually being used for — built from two corporate telemetry sources: [Fact] Anthropic's measure of Claude conversations and API traffic mapped to O*NET tasks, and Microsoft's measure of Copilot usage mapped to work activities.

Pause on that second dimension. A federal statistical agency is now building an official data product partly out of private AI vendors' usage logs — millions of Claude chats and Copilot sessions, aggregated and mapped to the government's own occupational taxonomy. To our knowledge, no official US statistical product has done this before. It is a quiet institutional milestone: the telemetry of two commercial AI systems has become, in effect, source data for the national statistical system.

The seams show if you look closely. [Fact] Of 4,155 possible occupation-source combinations (831 occupations × 5 sources), 3,944 were observed and 211 had to be imputed. [Estimate] That works out to roughly 5% of the data grid filled in by statistical modeling, touching 75 occupations — about 9% of the total. And BLS itself flags a deeper limitation: [Fact] the three theoretical sources capture AI capabilities "no later than mid-2023." The official yardstick for what AI could theoretically do is already more than three years — and several model generations — old.

The Agency Spends More Words on What This Is Not

Here is the counterintuitive part. The BLS documentation devotes remarkable space to warning readers against the exact interpretation most headlines will apply. [Fact] The agency states explicitly: an exposure category is not a forecast of employment growth or decline, not a worker replacement estimate, not a wage forecast, not a probability of AI adoption, and does not distinguish automation from augmentation. Exposure is relative, not absolute — a "Very high" label means an occupation's tasks overlap more with AI capabilities than other occupations' tasks do, nothing more.

The projections themselves prove why the warnings matter. [Fact] Computer and mathematical occupations sit squarely in AI's capability zone, yet they are projected to be the fifth fastest-growing major group at +7.3%. High exposure coexisting with strong growth is not an edge case — it is the pattern for the entire software field, where [Claim] BLS expects AI proliferation to create demand. If you read "Very high exposure" as "job doomed," the government's own numbers contradict you on the same page.

From Adjustment Factor to Standalone Product

Last year's 2024–34 release was the first in which BLS explicitly incorporated AI impacts into its occupation-level projections — we analyzed that cycle in detail in our earlier coverage, including the projected 5.5% decline for customer service representatives. Twelve months later, AI has graduated from a methodology adjustment buried in footnotes to a named, standalone data product with its own publication page and downloadable table.

The direction of travel is clear even if the numbers stay conservative. [Fact] BLS's own interpretation guidance says adjustments for emerging technologies are "generally applied conservatively," and advises readers to focus on the direction and relative size of changes rather than precise values.

What This Means for Your Job

The practical upgrade for workers is a two-axis reading that was not possible before: growth direction × exposure category. An occupation can now be checked against both.

If you work in administrative support or customer service, the projection and the exposure data point the same direction — that alignment, not either signal alone, is the warning worth acting on. If you are in software development or data science, you are in the odd quadrant: maximum exposure, strong projected growth — which reads as augmentation, not replacement, at least for this decade. And if you are a nurse practitioner or anywhere in hands-on healthcare, you are standing where the government expects a third of all new American jobs to appear.

The honest takeaway from this release is not a number. It is that the US statistical system has decided AI exposure is now a permanent, measurable attribute of every job — while insisting, loudly, that exposure is not destiny.

Sources

• BLS, Employment Projections — 2025–2035 (news release, August 27, 2026)
• BLS, Artificial Intelligence (AI) exposure categories
• BLS, AI impacts on employment projections (factsheet)

This article was produced with AI-assisted analysis. All figures were verified against the BLS primary release text; derived calculations ([Estimate] tags) are our own arithmetic from those published figures.