How the playbooks are built

Everything on Task by Task comes from the public O*NET 31.0 database. Our classification of each task is transparent and listed in full below.

1. The data: O*NET

O*NET is the U.S. Department of Labor's open database of occupations. For each of 1016 occupations with task data it lists the real tasks workers do: 18,838 task statements, most of them rated by workers themselves for importance, frequency and relevance.

O*NET also links every task to Detailed Work Activities (about 2,000). These roll up to 332 Intermediate Work Activities, which roll up to 41 Generalized Work Activities. That shared structure is what lets us reason about 18,000+ tasks consistently.

2. Classifying work activities

We hand-labelled all 332 Intermediate Work Activities with a play (the kind of help AI gives, like “Write it up” or “Crunch the numbers”) and a tier:

• AI-ready: AI can produce a usable first pass (drafting, summarizing, scheduling, record keeping, routine analysis). The worker reviews and owns it.
• AI-assist: AI speeds up prep, drafting or thinking, but the core of the task needs the worker's expertise, context or relationships.
• Stays human: physical, hands-on, care, in-person or accountable-judgment work. AI can help around it (checklists, rehearsal, write-ups) but not do it.

3. Refining each task

A task can link to several activities. For each task we pick the activity whose wording best matches the task, and break ties toward the more human label so we don't over-promise. Then we apply five rules to the task's own wording:

• Physical wording (“Lift…”, “Weld…”, “Clean…”, “by hand”, “bedside”, “using measuring devices”) makes a task human-led even when its activity is mostly desk work. Applied to 62 tasks.
• Desk wording (“Prepare reports…”, “Compile…”, “Schedule…”) upgrades an AI-assist documentation or planning task to AI-ready. Applied to 223 tasks.
• Document wording (“Prepare reports…”, “Write correspondence…”, “Develop manuals…”) puts the task in the writing play whatever activity it's filed under. It's AI-ready unless the task also involves delivering or presenting, which keeps it at AI-assist. Applied to 56 tasks.
• In-person wording (“Plan and supervise…”, “Prepare and deliver…”) keeps a mixed task at AI-assist rather than AI-ready. Applied to 128 tasks.
• Health privacy: AI-ready tasks involving patients or medical information are capped at AI-assist, because they need approved, compliant tools. Applied to 144 tasks.

Tasks that involve people's information, safety, legal requirements or money get extra “watch out” notes.

4. Human review

Rules get most tasks right, but not all. So we also reviewed by hand:

• Every one of the 2,076 Detailed Work Activities that tasks link to. 145 of them got a more specific label than their parent activity (for example, “Read blueprints to determine equipment setup” is AI-assist, not AI-ready, and “Make travel arrangements for others” is AI-ready even though its parent is in-person customer help).
• Blind audits of random 60-task samples. Before the activity review, about 92% of sampled tasks had the right tier; after it, 93% (56 of 60). The misfits were idiosyncratic and are corrected individually (9 tasks so far).

5. Emerging tasks

O*NET flags 338 tasks that are new to an occupation or newly reworded. They have no activity links or worker ratings yet. Reworded tasks inherit the activities of the task they replace. For new tasks, we score candidate activities from the five most similar existing O*NET tasks plus direct matches on activity names, then every one of the 292 new tasks was checked by hand (115 corrected; for example, operating a drone is treated as hands-on work throughout). They appear on each job's page but don't count toward its bar.

6. Occupations without task lists

O*NET publishes tasks for 923 of its 1,016 occupations. The other 93 are mostly “All Other” catch-alls and military occupations. Their pages list the work activities most shared across the specific occupations they group together (for military occupations, the closest civilian jobs, mapped by hand), classified the same way, with links to the full playbooks.

7. Monthly updates from real AI usage

Once a month the backend reads the newest Anthropic Economic Index release, which reports for each O*NET task how often people use Claude for it, and how: whether AI completed the task directly (“automation”) or the person iterated with it (“augmentation”), and how autonomously AI worked (1–5).

• A task moves from AI-assist to AI-ready when, in every recent month, AI completed it directly in at least 65% of conversations with an average autonomy of at least 3.0.
• It moves from AI-ready to AI-assist when automation is at most 30% and autonomy at most 2.6.
• Labels set by the human, physical, in-person or health-privacy rules, and hand-reviewed labels, never move. Tasks absent from the data don't move: absence isn't evidence.

The same run checks onetcenter.org for a newer O*NET database; the build downloads the newest one automatically. Every change is listed on the updates page.

8. Weighting and ordering

Each task's weight is importance × frequency × relevance, using O*NET's worker ratings (importance 1–5, frequency on a 7-point scale from “yearly or less” to “hourly or more”, and the share of workers who say the task is part of their job). Tasks are ordered by weight, and the “Where AI fits” bar is the weighted share of each tier. It's a guide to where AI can help, not a prediction of job loss.

9. Results

Across all 18,838 O*NET tasks: AI-ready 2,647 AI-assist 6,923 Stays human 9,268

10. Limitations

• Labels describe what generally available AI chat tools can help with in 2026, not specialised automation or robotics.
• O*NET describes the typical job. Your workplace, tools and rules may differ, so you're the judge of what fits.
• Prompts are templates. They need your specifics and your review. Never paste confidential or personal data into tools your employer hasn't approved.
• Rule-based refinement can misjudge individual tasks. If one looks wrong, the activity mapping shown at the bottom of each task card tells you why.

The plays

All 332 activity labels

The full lookup table. “Tasks” counts the tasks whose primary activity is this one.