How we score AI exposure

Every occupation gets a score from 0 to 100. It estimates how much of the work in that job AI tools can take over today. A score of 70 means that, across the tasks the job is made of, roughly 70% of the working time could be removed by AI tools that already exist. It doesn't say the job will disappear. Most jobs keep the tasks that need a person in the room, a pair of hands, or someone who is accountable for the result.

We score tasks, then add them up. A job title is too broad to grade in one go, but "prepare tax returns" or "give medication to patients" is narrow enough to judge.

Where the tasks come from

The task lists come from O*NET 30.3, the US Department of Labor's database of occupations. It covers 923 occupations and 18,796 task statements, each written and rated by people who do the job. O*NET also rates how important each task is to the job, which we use as its weight.

We wrote task lists for 15 Malta and iGaming roles that O*NET doesn't cover, such as KYC analyst, VIP manager and live casino dealer. They're graded the same way. The research measures below don't cover these roles, so each one borrows the average external score of the three O*NET occupations closest to it. A KYC analyst borrows from compliance officers, fraud analysts and loan clerks.

How a task is graded

A language model reads the occupation and its full task list and grades every task on three things:

• Exposure, 0 to 100: the share of the time spent on that task that AI can remove today, for a worker who uses widely available tools well. That means chat assistants, coding copilots, document, speech and image AI, and ordinary software automation. We don't assume future breakthroughs or general purpose robots.
• Label: automate when AI can do the task end to end and a person only reviews it, augment when AI does part of it faster but a person leads, and human when AI helps little.
• Why: a reason of 12 words or fewer, shown on each page.

The prompt includes fixed reference points, so grades stay consistent across jobs. Entering data from paper forms scores about 85. Interpreting medical images scores about 40. Installing wiring in a building scores about 3.

The grader is Claude Haiku 5.5, run with no tools and a fixed output format. Every response is checked before it's kept. Every task must be present, with a valid label and a number from 0 to 100. A task that fails goes back to the grader alone, and the tasks that passed are kept. A label that disagrees with its number, such as "automate" at 30, is logged for review.

From tasks to a score

1. Task score. The weighted average of task exposure. A task's weight is its O*NET importance rating. Core tasks count in full and supplemental tasks count half.
2. External score. Three published research measures, each turned into a percentile rank across occupations and averaged:
   • OpenAI and University of Pennsylvania, "GPTs are GPTs" (the beta measure, averaging its human and GPT-4 ratings).
   • Microsoft Research, "Working with AI" (AI applicability score).
   • Anthropic Economic Index (observed exposure from real Claude usage).
3. Final score = 70% task score + 30% external score, rounded. When none of the external measures cover an occupation, the final score is the task score.

The bands are: 0 to 29 low, 30 to 49 moderate, 50 to 69 high, 70 to 100 very high. The mix on each page is the share of the job's task weight labelled automate, augment and human.

Employment and wages come from the BLS Occupational Employment and Wage Statistics (May 2025), and the growth outlook from the BLS Employment Projections. Where one BLS code covers several O*NET occupations, the main occupation gets the full count and its specialties (such as acute care nurses under registered nurses) show no separate figure. When there's no main occupation, we split the count evenly.

Checking the grades

• Second opinion. We regrade a random 5% of occupations with a larger model, Claude Sonnet 5.5, and measure how often the two agree. The results are below.
• Outlier review. Scores that look wrong against what we know about a job get flagged for a person to read: a nurse above 70, an electrician above 35, or data entry below 40. Each flag is checked against the task grades behind it.
• Human overrides. When a person finds a wrong grade, they fix that one task. The fix is recorded in the repository and the score is rebuilt without regrading anything else.

Spot check result

Claude Sonnet 5.5 regraded 46 occupations, a fixed random 5% of the 923, covering 1,013 tasks. It used the same instructions and never saw the first grades.

Measure	Result
Same label (automate, augment, human)	83% of tasks (Cohen's kappa 0.71)
Average gap in task exposure	5.6 points out of 100
Task exposure within 15 points	95% of tasks
Average direction of the gap	Sonnet 0.9 points lower
Occupation task scores, correlation	0.96
Occupation task scores, average gap	2.7 points

So the two models rank occupations almost the same way, and an occupation's task score moves by about 3 points depending on which model grades it. Single tasks vary more, which is why we publish the weighted score and not any one task's number as the headline.

Grading the same occupation twice with the same model also moves its task score by up to about 4 points, because the model's reasoning differs from run to run. We grade each occupation once, cache it, and never regrade it unless its tasks or the grading instructions change.

Hours, departments and industries

The business pages turn scores into hours. We treat a task's share of the job's task weight as its share of working time, because O*NET rates how important each task is but not how long it takes. A full-time year is 2,000 hours. So a task that carries 5% of an accountant's task weight, with an exposure of 80, frees about 2,000 x 0.05 x 0.8 = 80 hours a year per accountant.

Departments group occupations into ten business functions, such as finance and accounting or customer service. Most are assigned by SOC code. A few business and office occupations with no obvious function are assigned by the O*NET work activity they rate most important, such as selling, administration or checking compliance.

Industries use the U.S. Bureau of Labor Statistics' staffing patterns (May 2025): how many people in each occupation work in each industry. An industry's score is the average of its occupations' scores, weighted by how many of its jobs each occupation holds.

Malta's workforce

Malta publishes employment by occupation group, not by detailed job. We use Eurostat's Labour Force Survey for 2025, which counts 322,600 employed people in Malta. The 37 civilian ISCO-08 groups with a figure cover 320,500 of them. For each group, we take the O*NET occupations that the BLS crosswalk links to it and average their scores, weighted by US employment. Malta's overall score is the average of the group scores, weighted by Malta's employment in each group. The share of Malta jobs in each band follows the same weights.

This assumes a job in a group looks the same in Malta as in the US. That holds better for clerks, accountants and developers than for groups whose work differs by country, such as farming or public administration. iGaming jobs are spread across several groups in this data. The 15 iGaming roles we graded, and the Malta Gaming Authority's count of 15,039 full-time jobs in the sector in 2025, are reported alongside the groups and aren't added into them.

Limits

• The scores describe tasks, not people. Two accountants can do very different work.
• "Can remove time" isn't the same as "will". Adoption depends on cost, regulation, trust and how a company is organised.
• O*NET describes US jobs. The tasks carry over well to most of Europe, but pay and employment figures are US only.
• A language model graded the tasks. We check it with a second model and human review, but individual grades can still be off by 10 or 20 points. The ranking between jobs is more reliable than any single number.