Data from Where Work Is Moving and The Price of Judgment
Work to 2035: occupations, programs and places
Search 830 occupations and 1,432 academic programs. See how much of each job's work AI is likely to take on by 2035, who should stay in charge of the decisions, and where the change lands first.
Our own model, 2025–2035. Robots and other physical automation are not yet included, and we do not yet track agentic AI, which carries out multi-step work on its own. Jobs are full-time equivalents of labor demand, not layoffs.
2026 editionReleased November 2026 · Updated quarterly · Data as of November 2026 · What the next edition adds →
Loading the data…
| Occupation | Jobs, 2025 | Median wage | AI reduction by 2035 | Net change 2025–35 (range) | Who decides | AI value per hour |
|---|
How many jobs in each state sit in occupations where AI removes a large share of the work, how many workers may need to change occupation, and the wages tied to the work AI takes over. Select a state to see its most exposed occupations.
New: state and county briefs. A short executive summary for every state and every county: its industries, how much of their work AI may take on, a slow-recovery stress test, and how the counties compare. Print or save any brief as a PDF.
Select a measure, then a state
| State | Jobs, 2025 | Jobs in highly exposed occupations | AI reduction by 2035 | Net change | May need to change occupation | Wages tied to work AI removes | Most exposed occupation |
|---|
Highly exposed occupations: occupations where AI removes 20% or more of labor demand by 2035, before growth. May need to change occupation: jobs in occupations that shrink. Wages tied to work AI removes: annual wages for the work hours removed, at state median wages. State model: national occupation rates with the AI cost test re-run on each state's wages and job mix.
Which model. The state and institution views use our 50-state model, with growth from the BLS 2025–35 projections and AI adoption from the Census Business Trends and Outlook Survey. It re-runs the AI cost test on each state's wages and job mix. It does not yet include robots and other physical automation, so states and programs tied to physical work look less exposed than they will. Added up, the state results come close to the national figures at the top of this page but do not match them exactly, because each state re-runs the AI cost test on its own wages.
| Program family | Level | Completions, 2023–24 | Demand for graduates, 2025–35 | Tier | AI share of graduates' work hours |
|---|
For each college and university: how many of its programs, and how many of its graduates, fall in each tier, and how it compares with similar institutions in the same state.
Peer groups in
Share of graduates by tier, all institutions of each type
By award level
Share of graduates in Invest programs, by level and institution type
| Institution | Type | Programs Invest · Reinvent · Review | Graduates by tier | Graduates | Invest share vs peers | In high-AI programs |
|---|
Programs are 4-digit fields of study by award level, IPEDS 2023–24. Peers are institutions of the same type in the same state. High-AI programs: AI pays for half or more of graduates' work hours. Tiers describe demand for graduates, not program quality.
From headline value to realizable value
AI almost always does a task faster than a person. But someone must check the work, redo what fails, and absorb the cost of errors that slip through. Counting those costs, about a fifth of the headline value of AI-exposed work is realizable today.
Value of AI-exposed work hours, U.S. economy, billions of dollars. Base case: 90% of failures caught; an error that escapes costs 5 times the task's value times its consequence.
Why AI fluency is an economic measure
If reviewers check AI work 25% faster and let 25% fewer errors through, realizable value rises from $649B to $848B.
- Faster review only
- $753B
- Fewer escaped errors only
- $741B
- Both, 10% better
- $727B
- Both, 50% better
- $1,061B
Who decides: four governance types
Occupations sorted by how costly an error is and how complex the work is. The type sets how fast AI can take over and how much human review stays in place.
Updates
A living estimate, updated quarterly
AI is changing faster than the data that measure it. We treat these results as a starting point. They are updated quarterly as new federal data are released, each update is dated, and the full model is re-run once a year, so the picture sharpens over time.
| Edition | Status | What it includes |
|---|---|---|
| 2026 edition | Current | Generative-era baseline: 830 occupations, 1,432 program families, 50 states and DC. Data set on OSF. |
| Agentic-era test | Waiting for data | Pre-registered test of entry-level employment, run as soon as the 2025 Census microdata are released. |
| 2027 edition | Planned | New federal wage, task and AI-use data, the agentic-era results, and revised scenarios. |
Change log
- November 20, 2026. State and county briefs: new sections on the work people do (jobs by occupation group), how much the work itself changes (High, Medium and Lower), and four kinds of change in High occupations, with graduates from local colleges in the same fields. College profiles: new Curriculum change section. Existing figures unchanged.
- November 18, 2026. State and county briefs added: a two-page executive summary for every state and county, with industries (BLS QCEW 2025 and OEWS May 2025 staffing patterns), a slow-recovery stress test and county comparisons.
- Version 2 · November 18, 2026. State and institution data and the headline figures re-run with the November 2026 state workbooks; college profiles added for 5,668 institutions, with local industries from BLS QCEW 2025, the program map and the AI line; data updated quarterly from here on.
- Version 1 · October 25, 2026. First release. Study, data and code registered on OSF.
Each edition keeps its own version of the data on OSF, so earlier results stay citable after an update.
Cite and download
Thalassinidis, A. E. (2026). Where work is moving [Data set]. OSF. osf.io/ae95y
Papers: Where Work Is Moving · The Price of Judgment
Model results are estimates under stated assumptions, not forecasts for any single employer, program or place.
Data sources and credits
How AI changes work
- O*NET 31.0 (U.S. Department of Labor): tasks, work context, skills
- Eloundou, Manning, Mishkin & Rock (2024), Science: which tasks AI can do
- Anthropic Economic Index, Claude.ai task data (January 2026): task success and time saved
- Census Business Trends and Outlook Survey (November 2025 to October 2026): speed of AI adoption
- Census Statistics of U.S. Businesses 2022: employment weights for adoption
Jobs, growth and where new jobs land
- BLS Employment Projections 2025–35: growth, separations, entry education, and BLS's own AI adjustments
- BLS National Employment Matrix 2025–35
- BLS input-output tables (real final demand, 2025 and 2035)
- BLS domestic employment requirements table (2025)
- BLS Occupational Employment and Wage Statistics (OEWS), May 2025: employment, wages, staffing patterns
- BLS Quarterly Census of Employment and Wages (QCEW) 2025: county employment by industry
- Weldon Cooper Center national population projections (vintage 2026): school-age drivers
Places, commuting and moving
- USDA Economic Research Service 2020 Commuting Zones
- Census LODES 8 (2024): where workers live
- Census county boundaries 2024: distances
- Census American Community Survey 2020–24, table C24010: county check only
Colleges and programs
- NCES IPEDS Completions 2023–24 and Institutional Characteristics
- NCES/BLS CIP 2020–SOC 2018 crosswalk
- Census Post-Secondary Employment Outcomes (PSEO): graduate outcomes, and the share of graduates working in the state
- DOL Eligible Training Provider list (TrainingProviderResults.gov)
- DOL list of apprenticeable occupations
Model and scripts
- Thalassinidis (2026), Where Work Is Moving, on OSF (osf.io/ae95y)
Each state's workbook lists its own state sources.
Licenses
- This site includes information from the O*NET 31.0 Database by the U.S. Department of Labor, Employment and Training Administration (USDOL/ETA), used under the CC BY 4.0 license. Angelo E. Thalassinidis has modified all or some of this information. USDOL/ETA has not approved, endorsed, or tested these modifications. O*NET® is a trademark of USDOL/ETA.
- Anthropic Economic Index data are used under the CC BY 4.0 license.
- Task exposure data from Eloundou et al. (2024), Science, 384(6702), at github.com/openai/GPTs-are-GPTs, Copyright (c) 2024 OpenAI, used under the MIT License.
- Data from BLS, the Census Bureau, NCES, USDA and the Department of Labor are federal works in the public domain in the United States.