An open dataset estimating how much of a typical working week in 217 occupations current AI can already do, how much it can speed up, and how much stays with the person. It also has a five-year projection for each job.
These are the same numbers behind the job pages on humanedge.work. There you can also check your own job in about two minutes, for free and without signing up.
| File | Format |
|---|---|
occupations.csv |
One row per occupation, sorted from most to least exposed |
occupations.json |
Same data as JSON |
| Column | Meaning |
|---|---|
rank |
1 = most exposed of the 217 |
occupation |
Job title |
category |
Broad occupational group |
automatable_today |
% of a typical week that current AI can do without a person |
ai_accelerated_today |
% that AI speeds up, with the person still doing the work |
human_anchored_today |
% that stays with the person: judgement, trust, physical presence, accountability |
automatable_in_5_years |
Projected automatable share in five years if current capability trends continue |
human_edge_score |
0–100 summary of how well-positioned the role is (higher is safer) |
url |
Full breakdown for that job: which tasks are exposed and which AI tools to use |
The three _today columns add up to 100.
| Most exposed | Automatable today | Least exposed | Automatable today |
|---|---|---|---|
| Data Entry Clerk | 78% | Construction Labourer | 5% |
| Virtual Assistant | 75% | Sewing Machinist / Textile Worker | 6% |
| Transcriptionist / Court Reporter | 73% | Logging Equipment Operator | 7% |
| Administrative Assistant | 71% | ||
| Translator / Interpreter | 69% |
Each occupation is modelled as a mix of 37 work activities. Each activity has four parameters:
- how well AI does it today;
- whether AI replaces the person or amplifies them;
- how strongly the activity is anchored to a specific human;
- how fast the remaining gap is closing.
A job's scores are the weighted sum over its activities.
The model is checked against published research. The main check is against Microsoft Research's Working with AI: Measuring the Applicability of Generative AI to Occupations: the build fails if the model's ordering disagrees with their published high and low groups. The Anthropic Economic Index and Stanford Digital Economy Lab's Canaries in the Coal Mine? also informed it. These organisations are cited as sources only. They have not reviewed or endorsed this dataset.
The full method, including a list of the model's known weaknesses, is at humanedge.work/method.
- These are estimates for a typical person in each role, not forecasts for any individual. Seniority, client relationships, formal accountability and where the work happens all change a real person's position a lot.
- The five-year figures project current trends forward. Treat them as a direction, not a prediction.
- "Automatable" means the tasks, not the job. A role that is 60% automatable usually changes shape rather than disappearing.
Data: CC BY 4.0. Use it for anything, including commercially, with a link back to Human Edge.
Suggested citation: Human Edge (2026). AI exposure scores for 217 occupations. https://humanedge.work
Want your own number instead of the average? Take the free 2-minute assessment →