How it works

In simple words

For each job we take the official list of tasks from the U.S. Department of Labor (O*NET). Then we check each task against real data on what people actually use AI for. Your result is the share of the tasks you ticked that show up in that AI data. It shows overlap, not replacement.

The details

Tasks: O*NET 31.0 task statements; “core” tasks are pre-ticked and count double in your score. We shorten task wording for readability (we cut long lists after “such as” and swap a few formal words, e.g. “utilize” → “use”).

AI usage per task: Anthropic Economic Index, release 2025-03-27, file task_pct_v2.csv (share of sampled Claude conversations mapped to each O*NET task). High = top 15% of non-zero shares (≥2.876%), medium = above median (≥0.544%), low = any other recorded usage, none = not found.

Matching: (1) exact wording match; (2) if none, a “close match”: TF-IDF cosine similarity on content words against all Anthropic task texts, accepted only at ≥0.66. Threshold chosen by manual review: matches below ~0.66 often linked different tasks. Coverage across our 1095 tasks: exact 321, close 51, not found 723. “Not found” can mean the task wasn’t in the sample — it is not proof AI can’t help.

Traffic light: green under 25%, amber 25–49%, red 50%+. These cut-offs are our own choice for readability.

Microsoft data: the “Working with AI” occupation applicability score (Bing Copilot, 2024) is kept in our data for reference.

Plans and prompts: made by fixed rules from each job’s data (tasks, O*NET skills, in-demand software). No AI model writes them. They are guidance only.