Methodology

1. The core dataset

The backbone is “Working with AI: Measuring the Applicability of Generative AI to Occupations” (Kiran Tomlinson, Sonia Jaffe, Will Wang, Scott Counts, Siddharth Suri; Microsoft Research, 2025; arXiv:2507.07935). The researchers analyzed 200,000 anonymized conversations between U.S. users and Bing Copilot, classified each conversation against the U.S. Department of Labor's O*NET work-activity taxonomy, and measured three things per activity:

• Coverage: how often the activity actually appears in AI usage;
• Completion: how successfully the AI performs it when it does;
• Scope: how much of the activity the AI handles rather than a small slice.

From these, they published an AI applicability score for 785 occupations, along with per-activity metrics for 332 intermediate work activities (IWAs). Microsoft released the result files openly under CC BY 4.0 on GitHub. We use them as published: no re-weighting, no editorial adjustment of any score.

Why this dataset over the alternatives? Because it is measured behavior, not prediction. The famous 47% figure (Frey & Osborne, 2013) came from experts guessing "computerisation probability" per occupation; later task-based work (Arntz, Gregory & Zierahn for the OECD, 2016) cut that estimate to ~9% just by changing assumptions. Usage data has its own biases (see §6), but it starts from what AI demonstrably does, not what someone imagines it might do.

2. Occupation descriptions

Job descriptions on occupation pages come from the O*NET 30.0 Database (2025) by the U.S. Department of Labor, Employment and Training Administration, used under CC BY 4.0. O*NET® is a registered trademark of USDOL/ETA; this site is not endorsed by them.

3. From scores to bands

The applicability score is a continuous value. Publishing it raw invites false precision ("my job is 0.213!"), so we present quintile bands of the actual distribution across all 785 occupations: each band contains exactly one fifth of occupations:

Band	Score range	Meaning
Minimal	< 0.070	Bottom 20%: AI is rarely observed doing this occupation's activities
Low	0.070 – 0.117	20th–40th percentile
Moderate	0.117 – 0.171	40th–60th percentile
Elevated	0.171 – 0.250	60th–80th percentile
High	> 0.250	Top 20%: heavy overlap between the job's activities and what AI does well today

We also show a percentile (rank among the 785) and the full distribution histogram, so a score is always read in context.

4. The task table on each occupation page

Each occupation page lists its top work activities by O*NET importance weight (normalized to the shown set). Each activity carries the measured AI performance value, completion × scope × coverage from the user-goal perspective, displayed as one of four states:

Label	Value	Reading
High overlap	≥ 0.60	AI is frequently used for this and completes it well
Moderate	0.40 – 0.60	Regular use with mixed success or partial scope
Emerging	> 0 – 0.40	Observed, but limited success, scope or frequency
Not observed	0	This activity does not meaningfully appear in the 200K-conversation sample

That last state matters: 205 of the 332 measured activities show zero observed AI performance: most physical, supervisory and hands-on care work. When your page says "not observed," that's a real measurement, not a gap in our data entry.

5. Personalization & the task self-assessment

The six-question personalization on occupation pages applies small, fully-displayed percentile adjustments (each between −6 and +5 points, total capped at ±18) for factors the occupation-level data can't see: physical presence, legal accountability, relationship dependence, routineness, and your own AI adoption. The result is deliberately a range, not a number. The adjustment sizes are our editorial judgment, informed by the moderating factors named across the research literature. They are labeled as such and never blended invisibly into the source scores.

The task self-assessment uses ten task families; each family's value is the mean measured AI performance across its constituent activities in the Microsoft dataset (zeros included). Your result is Σ (time × family value) ÷ Σ time, placed against the occupation distribution for context and labeled indicative.

6. Limitations; read these

• Usage data reflects one assistant, one period. Bing Copilot conversations, U.S. users, collected January–September 2024 (published 2025). Enterprise agents, coding copilots and newer models are underrepresented; true capability is likely ahead of this data in fast-moving areas like software.
• Applicability ≠ displacement. The paper's own authors warn against reading the score as "probability of job loss." It measures overlap, which can mean augmentation, redesign, or replacement; economics, regulation and management decide which.
• U.S. taxonomy. O*NET describes U.S. jobs. Task content travels across borders reasonably well; labor-market outcomes don't.
• No wage or employment weighting. We rank occupations by exposure, not by how many people hold them or what they earn.
• Physical automation is out of frame. This measures generative AI (language, analysis, images); not warehouse robots or self-driving trucks. A trucker's low score here says nothing about autonomous vehicles.
• Editorial layers are labeled. Band names, personalization weights, family groupings and action plans are our design on top of the data, and each is called out where it appears.

7. Wider evidence we cite for context

• Anthropic Economic Index; ongoing measurement of Claude usage; ~52% augmentation vs 45% automation (Nov 2025). anthropic.com/economic-index
• World Economic Forum, Future of Jobs Report 2025; employer survey projecting ~92M jobs displaced / ~170M created by 2030.
• IMF (2024); ~40% of global employment exposed to AI; ~60% in advanced economies, roughly half of it complementary.
• Eloundou et al. (2023), GPTs are GPTs; ~80% of U.S. workers have ≥10% of tasks exposed to LLMs.
• Frey & Osborne (2013) and Arntz et al. (2016); the historical bookends of the prediction debate; cited as history, not as inputs.

8. Data vintage & updates

Component	Version	Integrated
Microsoft "Working with AI" result files	GitHub main (2025)	2026-07
O*NET Database	30.0 (2025)	2026-07

When materially better public data appears (new dataset releases, O*NET updates), scores are rebuilt from source and this table changes. No silent edits.

9. What this site never does

• No hidden model deciding your result; everything is static data plus arithmetic you can check.
• No made-up percentages of doom, no countdown clocks, no "your job dies in 2027."
• No accounts, no cookies, no tracking, no selling your inputs; there's nothing to sell; your answers never leave the browser. The only thing stored is an optional "recently viewed" list in your own browser's local storage.

10. Version history

Version	Date	Changes
V2	2026-07	Added 22 field pages with static rankings and band mix; side-by-side comparison tool with shared-activity analysis; CSV export of the rankings table; field-relative rank on every occupation page; per-page structured data (Occupation, FAQ); site-wide search in the header; strict Content-Security-Policy (scripts restricted to this site's own files); accessibility upgrades (combobox and quiz semantics, visible focus states).
V1	2026-07	Initial rebuild: 785 occupation pages on the Microsoft Research dataset, task-level tables, personalization quiz, adjacent moves, self-assessment, methodology.