Opens in a new tab
needsahuman.

Can AI do it? How we measure capability coverage

Capability Coverage is the share of a job's working time that AI can handle today, on a scale of 0 to 100. A coverage of 40 means AI can handle tasks that make up about 40% of the job's weighted task time. It is the first of our three scores and the biggest input to…

Capability Coverage is the share of a job’s working time that AI can handle today, on a scale of 0 to 100. A coverage of 40 means AI can handle tasks that make up about 40% of the job’s weighted task time.

It is the first of our three scores and the biggest input to the headline Still needs a human score.

Step 1: List the job’s tasks

Every occupation in O*NET has a list of task statements, such as “Prepare, examine, or analyze accounting records” or “Explain medical procedures to patients”. O*NET also rates each task for importance and for how often it is done.

We give each task a weight of importance × frequency, then scale the weights so they add up to 100% for the job. A task done daily that matters a lot counts for much more than a yearly task that matters a little.

In our current data, 923 occupations have task lists, covering 18,838 tasks. Occupations without task statements, such as “all other” catch-all codes and military codes, are not scored.

Step 2: Score each task

Each task gets a score from 0 to 100 built from three kinds of evidence:

Task score = 0.40 × observed + 0.35 × theoretical + 0.25 × measured

If one kind of evidence is missing for a task, the other weights are scaled up so they still add to 1. The dataset records which components were present.

No task has a measured value yet (see below). So in the current scores, every task is scored from observed use and its theoretical rating alone, and their weights are scaled up from 0.40 and 0.35 to about 0.53 and 0.47.

Observed: what people use AI for

How much a task already shows up in real AI use at work. We average two open sources:

  • The Anthropic Economic Index. Anthropic publishes, task by task, how much of the work in each O*NET task appears in conversations with its Claude assistant, and whether that use is automation (AI does the task) or augmentation (AI helps a person do it). We use its task penetration measure.
  • Microsoft Research, Working with AI. Microsoft researchers measured which work activities people successfully use Bing Copilot for, and how much of each activity the AI covered. We use their completion × scope measure wherever the activity shows up in real use, and join it to O*NET tasks through O*NET’s own activity links.

Both sources measure use of AI chat assistants. Neither measures what AI could do, and neither sees automation that runs through other software. Usage is low on most tasks, even ones AI could do end to end, and that holds coverage down. That is why observed use is only part of the score, and one reason no job’s coverage is above 62 today.

Theoretical: what current AI could do

We rate every O*NET task against a published rubric (version r1, 2 October 2026). It asks what a frontier AI system can do with the standard tools of October 2026: documents and spreadsheets, running code, a browser, voice, reading images and documents, and connections to ordinary business software. It does not rate robots, which are handled in Step 3, and it ignores cost, regulation and liability, which show up under “What’s stopping it”.

LevelScoreWhat it means
T00No meaningful part can be done by AI: the task needs physical presence or handling, in-person care, or authority that cannot be handed to software.
T130AI can help with a minor part, such as look-ups, notes or first drafts. The core stays with the person; time saved is under about a third.
T265AI can do most of the task with a person reviewing, approving or handling exceptions. Time saved is more than half.
T390AI can do the task end to end at the quality of a typical qualified worker, with spot checks only.

The top level scores 90, not 100, because even those tasks keep some review. Each task is also put in one capability area (writing, analysis, coding, vision, speech, planning, physical, care), which feeds “Which AI skills matter” on the job page.

The approach follows published research that rated O*NET tasks for AI exposure (Eloundou et al., “GPTs are GPTs”, 2023), updated because tools such as a browser and code execution are now standard.

Who rated the tasks. One AI model, Anthropic’s Claude Sonnet 5.5, rated all 18,838 tasks under the rubric on 3 October 2026, up to 25 tasks at a time from one occupation. None was missing or refused. Across all tasks, 36% were rated T0, 26% T1, 26% T2 and 12% T3.

How we checked them. Before the full run, 573 tasks across 26 jobs were rated separately under the same rubric. The two runs agree exactly on 81% of those tasks and are within one level on 99.7%, with no lean either way. Task statements that appear word for word in more than one job got the same level 88% of the time, and never more than one level apart. Both runs were done by Claude, so this shows the rubric is applied consistently. It does not show the ratings are right. No person has rated a sample yet. We publish the rubric, the model, the run date and every rating with the dataset so others can check them.

Measured: how AI does on real work

Where a benchmark tests AI on work that matches a task, such as OpenAI’s GDPval or the Remote Labor Index, we will map the benchmark result to that task, and use the job family’s average, marked as imputed, where a job has none. No measured values are in the scores yet. Until they are, observed use and the task ratings carry this part of the weight.

Step 3: Cap the hands-on work

Software cannot lay a pipe or lift a patient. For tasks that are mainly physical, the task score is capped at the level robots have reached for that kind of work. We take the physical flag and a robot tier for each task from Anthropic’s robot exposure data (“What work can robots do?”, September 2026). Each tier has its own cap.

Robot tierMeaningCap
E0No robot is known to do this task in any setting15
E1Robots do it only in a purpose-built cell, such as a factory station25
E2Robots do it in structured workplaces built for people, such as warehouses35
E3Robots do it even in unstructured, everyday settings50

A physical task with no tier is capped at 35. The cap values are our judgment, reviewed each release.

The cap only ever lowers a task’s score. Robots never raise it. So coverage measures software AI, with robotics as a ceiling on hands-on work. Factory automation and other machinery do not show in it.

Step 4: Add it up

The job’s coverage is the weighted average of its task scores. We also sort each task into one of three groups for the page:

Task scoreShown as
60 or moreAI does this now
30 to just under 60AI helps
Under 30Still needs a human

Because observed use is low on most tasks, few reach 60. Across the October 2026 release, on average 5% of a job’s task time is “AI does this now”, 34% “AI helps” and 61% “Still needs a human”.

Alongside coverage, each page shows the automate vs augment split: of the AI use we can see, how much is AI doing the task and how much is AI helping a person do it.

How to read it

  • Under 20: AI touches little of this job today.
  • 20 to 50: AI handles a real slice of the work, usually the writing, searching, summarizing or checking.
  • Over 50: most of the job’s working time is in tasks AI can handle.

High coverage does not mean the job disappears. It means the work is changing. Whether AI does that work well enough is the next question, Is it better than a person?

Known limits

  • Assistant use is not all automation. Scanning software, machine translation, imaging tools and self-checkout are not in the usage data, so jobs such as data entry and translation may look safer than they are.
  • Robots only cap. Robot data holds hands-on tasks down but never adds to coverage.
  • No benchmark results yet. The measured part of the score is empty for every task.
  • Coding is under-counted in the usage sources, because some coding tools are outside them.
  • O*NET task lists describe the job as surveyed, which can lag behind how the job is done now.
  • Task ratings are judgments by one AI model. A model following a rubric is consistent, but the rubric is ours and no person has checked a sample yet. The rubric is cautious about work with people, such as care and therapy, and can be generous on real-time safety-critical work, such as air traffic control, because it ignores regulation. We publish every rating so others can check it.

Sources

  • O*NET 31.0 database, US Department of Labor, Employment and Training Administration, CC BY 4.0. onetcenter.org
  • Anthropic Economic Index, including the robot exposure data, CC BY. Hugging Face dataset
  • Microsoft Research, “Working with AI: Measuring the Occupational Implications of Generative AI”, CC BY 4.0.
  • Eloundou et al., “GPTs are GPTs” (2023), for the task-rating approach.
  • Our task ratings (rubric r1, Claude Sonnet 5.5, 3 October 2026), published with each release on the open data page.