Methodology
How we calculate the composite AI automation risk score for each occupation.
Composite Risk Score Formula
Risk = 0.30ยทF + 0.20ยทT + 0.20ยทB + 0.15ยทL + 0.15ยทG
Each component is normalized to a 0โ100 scale before weighting.
Frey & Osborne Automation Probability
30%Based on the 2017 Oxford study that estimated automation probabilities for 702 occupations using machine learning classifiers trained on O*NET task data. We use the original probabilities mapped to SOC codes, normalized to a 0โ100 scale.
The Frey/Osborne model classifies occupations based on three engineering bottlenecks: perception and manipulation, creative intelligence, and social intelligence. Tasks requiring these capabilities are less automatable. We apply their original 702-occupation probabilities, extended to 925 occupations via SOC crosswalk and O*NET task similarity matching for occupations not in the original study.
๐ Frey, C.B. & Osborne, M.A. (2017). The Future of Employment. Technological Forecasting and Social Change.
OECD Task-Based Analysis
20%The OECD approach analyzes individual tasks within occupations rather than whole jobs. It counts the proportion of tasks that are routine (manual or cognitive) versus non-routine, using O*NET task importance and frequency ratings. Occupations with >60% routine tasks score higher.
Unlike the Frey/Osborne approach that classifies entire occupations, the OECD method recognizes that most jobs contain a mix of automatable and non-automatable tasks. We calculate the routine task intensity (RTI) index using O*NET's 41 work activity variables, weighted by task importance (scale 1-5) and frequency (scale 1-7). The RTI score is then normalized to 0โ100.
๐ Nedelkoska, L. & Quintini, G. (2018). Automation, skills use and training. OECD Social, Employment and Migration Working Papers No. 202.
BLS Employment Projections
20%10-year employment projections from the Bureau of Labor Statistics. Occupations projected to decline receive higher risk scores. We invert and normalize the projected growth rate: a -20% projection maps to ~90 risk; +20% maps to ~10 risk.
BLS projections incorporate macroeconomic modeling, industry output projections, and staffing pattern analysis. While BLS doesn't explicitly model AI displacement, their projections capture historical automation trends and industry shifts. We use the 2024โ2034 projections (released September 2025), applying a linear normalization: projected growth of -30% or worse = 100 risk; +30% or better = 0 risk.
๐ Bureau of Labor Statistics, Employment Projections program (2024โ2034).
Layoff & Restructuring Signal
15%Real-time signal derived from WARN Act filings, SEC 8-K restructuring disclosures, and news-reported layoffs mentioning AI or automation. We aggregate events by NAICS โ SOC crosswalk, weighted by recency (exponential decay, ฯ=90 days) and scale (affected headcount).
We collect WARN Act filings from all 50 states plus DC, SEC 8-K filings mentioning 'restructuring,' 'workforce reduction,' or 'automation,' and verified news reports from major outlets. Each event is coded by industry (NAICS), mapped to affected occupations via BLS staffing patterns, and weighted by: (1) recency โ exponential decay with 90-day half-life, (2) scale โ log-transformed headcount, and (3) AI attribution โ events explicitly citing AI/automation receive 2x weight.
๐ State WARN Act databases, SEC EDGAR filings, verified news reports.
GenAI Exposure Index
15%Measures how exposed an occupation's core tasks are to large language models and generative AI specifically. Based on Eloundou et al. (2023) methodology: each task is rated ฮฑ (no exposure), ฮฒ (LLM alone), or ฮณ (LLM + tools). The exposure score is the weighted proportion of ฮฒ+ฮณ tasks.
We extend the original Eloundou et al. framework with 2025 capability assessments. As LLMs have improved, some tasks previously rated ฮฑ (no exposure) have been reclassified to ฮฒ or ฮณ. Our updated ratings incorporate real-world deployment data from enterprise AI adoption surveys, GitHub Copilot usage statistics, and customer service automation benchmarks. The exposure score weights ฮฒ tasks at 0.5 and ฮณ tasks at 1.0.
๐ Eloundou, T. et al. (2023). GPTs are GPTs: An Early Look at the Labor Market Impact Potential of LLMs. arXiv:2303.10130.
Data Sources
| Source | Type | Update Frequency | Usage |
|---|---|---|---|
| Bureau of Labor Statistics (BLS) | Government | Annual (May release) | Occupational employment, wages, projections |
| O*NET OnLine | Government | Biannual (v29.1 current) | Task-level occupation data, skills, work activities |
| Federal Reserve (FRED) | Government | Monthly | Unemployment, JOLTS, labor force participation |
| WARN Act Filings | Government | Continuous (state-level) | Mass layoff notifications |
| SEC EDGAR | Government | Continuous | Corporate restructuring disclosures |
| Census Bureau (CBP) | Government | Annual | County-level employment by industry |
| Frey & Osborne (2017) | Academic | Static (seminal study) | Base automation probabilities |
| Eloundou et al. (2023) | Academic | Static + our 2025 update | LLM task exposure ratings |
| OECD (2018) | Academic | Static (methodology) | Task-based automation framework |
| layoffs.fyi | Community | Continuous | Tech layoff event tracking |
Score Interpretation
| Score Range | Label | Interpretation |
|---|---|---|
| 0โ20 | Low Risk | Core tasks require human judgment, creativity, or physical dexterity that current AI cannot replicate. |
| 21โ40 | Moderate | Some tasks are automatable, but the role adapts through augmentation rather than displacement. |
| 41โ60 | Elevated | Significant task automation is underway. Workers should actively build complementary skills. |
| 61โ80 | High Risk | Majority of core tasks face automation pressure. Career transition planning is advisable. |
| 81โ100 | Very High Risk | Most tasks can be performed by current or near-term AI. Significant workforce reduction likely within 5โ10 years. |
How We Compare to Other Indices
Several organizations measure automation and AI risk. Here's how our approach differs from the most commonly cited frameworks.
AI Exposure (Ours)
โ Strength: Most comprehensive โ combines academic models, government data, and real-time signals
โ ๏ธ Limitation: US-only; layoff signals can be noisy
Frey & Osborne (Oxford)
โ Strength: Seminal study; widely cited; granular occupation-level estimates
โ ๏ธ Limitation: Pre-LLM; treats occupations as binary (automatable or not); no task-level nuance
OECD Automation Risk
โ Strength: International coverage; recognizes within-occupation variation
โ ๏ธ Limitation: Lower granularity; uses broader occupation categories; less frequent updates
World Economic Forum (WEF)
โ Strength: Incorporates employer intent and investment plans; forward-looking
โ ๏ธ Limitation: Survey-based (subject to bias); aggregate sector-level only; no occupation scores
McKinsey Global Institute
โ Strength: Detailed task analysis; considers adoption pace scenarios
โ ๏ธ Limitation: Proprietary methodology; not freely available; less frequent updates
Limitations
- Risk scores reflect task-level automation potential, not definitive job loss predictions.
- Adoption speed depends on regulation, cost, and organizational inertia โ none of which are modeled.
- The Frey/Osborne component may overestimate risk for some service occupations.
- Layoff signals are noisy and may reflect economic cycles rather than automation.
- Scores are updated periodically as new data becomes available, not in real-time.
- Our model assumes current technology trajectories. Breakthrough capabilities (or regulatory freezes) could shift scores significantly.
- Geographic variation within occupations is not captured โ a bank teller in NYC faces different automation pressure than one in rural Montana.
- We don't model wage effects. An occupation can have a low displacement risk but still see significant wage compression from AI-driven productivity gains.
Academic References
Frey, C.B. & Osborne, M.A. (2017). "The Future of Employment: How Susceptible Are Jobs to Computerisation?" Technological Forecasting and Social Change, 114, 254-280.
Nedelkoska, L. & Quintini, G. (2018). "Automation, skills use and training." OECD Social, Employment and Migration Working Papers, No. 202.
Eloundou, T., Manning, S., Mishkin, P. & Rock, D. (2023). "GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models." arXiv:2303.10130.
Acemoglu, D. & Restrepo, P. (2019). "Automation and New Tasks: How Technology Displaces and Reinstates Labor." Journal of Economic Perspectives, 33(2), 3-30.
Autor, D.H. (2015). "Why Are There Still So Many Jobs? The History and Future of Workplace Automation." Journal of Economic Perspectives, 29(3), 3-30.
Brynjolfsson, E., Mitchell, T. & Rock, D. (2018). "What Can Machines Learn, and What Does It Mean for Occupations and the Economy?" AEA Papers and Proceedings, 108, 43-47.
Webb, M. (2020). "The Impact of Artificial Intelligence on the Labor Market." Stanford University Working Paper.
Felten, E., Raj, M. & 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.
Updates & Versioning
The current methodology is v2.1 (July 2026). Major updates occur when new BLS projections are released (typically every 2 years) or when significant new research warrants weight adjustments. All historical scores are preserved for longitudinal comparison.
Version History
v2.1 (July 2026) โ Updated GenAI exposure ratings to reflect 2025 LLM capabilities; added 2024โ2034 BLS projections.
v2.0 (January 2026) โ Added GenAI Exposure Index as 5th component; reweighted from 4 to 5 factors.
v1.1 (March 2025) โ Expanded from 702 to 925 occupations via O*NET task similarity imputation.
v1.0 (September 2024) โ Initial 4-component model launch.
Frequently Asked Questions
How often are AI risk scores updated?
Scores are updated quarterly. The layoff signal component (15% of the score) updates continuously as new WARN Act filings and news reports are processed. Major recalibrations happen when BLS releases new employment projections (every 2 years) or when significant new research warrants weight adjustments.
Why don't you use a single methodology like Frey/Osborne?
No single model captures the full picture. Frey/Osborne was published in 2017 โ before LLMs existed. The OECD approach handles task-level nuance better. BLS projections capture real market dynamics. Layoff signals show actual displacement happening now. By combining five approaches, we get a more robust and current assessment than any single model provides.
Can I reproduce your scores?
Yes. All input data sources are public. We publish our normalization formulas, weight assignments, and SOC crosswalk tables. The only proprietary element is our WARN Act filing aggregation (since each state publishes in different formats), but the raw filings are public records. See our API documentation and downloadable datasets for full transparency.
How do you handle occupations not in the Frey/Osborne study?
Frey & Osborne scored 702 of the 840+ SOC-coded occupations. For the remaining ~125 occupations in our 925-occupation dataset, we use O*NET task similarity matching: we find the 3 most similar scored occupations (by task profile cosine similarity) and use their weighted average probability. These imputed scores are flagged in our data exports.
Why do some scores differ from other AI risk tools?
Different methodologies produce different results. Tools using only Frey/Osborne will give different scores than our composite approach. Our layoff signal and GenAI exposure components are unique โ they capture developments since 2022 that static academic models miss. We explain every component and its weight so you can evaluate our approach yourself.
Explore the Data
See our methodology in action across 925 occupations.