Opens in a new tab
needsahuman
Search ⌘KGet the data
Guide · updated 3 October 2026

How we score jobs, in plain English

A plain-English walk through how NeedsAHuman decides whether AI could replace a job: tasks, three questions, evidence grades and ranges.

Key points

  • We break every job into its tasks and score each task, weighted by how much time it takes.
  • We ask three questions: Can AI do it? Is it better than a person? When could it be replaced?
  • The answers roll up into one score, Still needs a human, from 0 to 100. Higher is safer.
  • Every quality score has an evidence grade, and every timeline has a range.
  • All our data is public, and we update it every quarter.

Step 1: Break the job into tasks

A job is not one thing. A nurse checks vital signs, gives medicine, writes notes, comforts patients and talks to doctors. AI might help with the notes but not with the rest.

So we start with each job’s task list from O*NET, the US Department of Labor’s occupation database. The current version, O*NET 31.0, was released in August 2026. Its taxonomy “includes 1,016 occupational titles, of which 923 represent O*NET data-level occupations,” meaning the ones with full task data.

O*NET rates each task for how important it is and how often it is done. We use those ratings to weight each task, so the work that fills most of a day counts most.

We currently score 923 jobs, in our 2026-Q4 data release.

Step 2: Ask “Can AI do it?”

For each task, we look at three kinds of evidence.

What people actually use AI for. Two large studies track this:

  • Anthropic’s Economic Index looks at how people use its Claude assistant at work, task by task, and whether AI is doing the task or helping a person do it.
  • Microsoft Research studied “200k anonymized conversations with Microsoft Bing Copilot” from 2024 (Tomlinson et al.) to see which work activities people use AI for.

What AI could do. We rate each task against a published rubric: could today’s AI cut the time it takes by at least half, at the same quality?

How AI does on real work. Benchmarks that test AI on genuine work tasks.

We blend these into a score for each task, then add up the tasks. That gives the job’s coverage: how much of its working time AI can handle.

Physical tasks get special treatment. AI chat assistants cannot fix a pipe, so for hands-on tasks we cap the score at what robots can actually do today.

Full details on coverage.

Step 3: Ask “Is it better than a person?”

Being able to do a task is not the same as doing it well. So we ask how AI’s work compares with a qualified professional’s.

The best evidence comes from blind tests, where experts compare AI’s work with professionals’ work without knowing which is which. One example is OpenAI’s GDPval, which “covers the majority of U.S. Bureau of Labor Statistics Work Activities for 44 occupations across the top 9 sectors.” The experts grading it had “an average of 14 years of experience.” In its first results, “47.6% of deliverables by Claude Opus 4.1 were graded as better than (wins) or as good as (ties) the human deliverable.” Its authors note its tasks are “one-shot, not interactive,” which is not how most real work happens.

Tests like this exist for only a few dozen jobs. So every quality score comes with an evidence grade:

  • A: a blind, expert-judged test on this job’s own work.
  • B: a test on closely related work.
  • C: weaker evidence, such as benchmarks without a human comparison.
  • D: no usable evidence. We show “Not yet measured”.

Right now only a minority of jobs have grade A or B evidence. Most jobs are C or D, and we say so on the page.

Full details on quality.

Step 4: Ask “When could it be replaced?”

This is the hardest question, so we answer it carefully.

First, we set a high bar. A job only counts as replaceable when AI can do at least 90% of its task time, at least as well as a person, reliably, and at least half of employers have adopted it.

Then we run 10,000 simulated futures. Each one makes different assumptions about things nobody knows for sure:

  • How fast AI improves. We use research from METR, which measures how long a task AI agents can complete. In its original 2025 study, that length showed “a doubling time of around 7 months.” METR later reported faster doubling in recent years. Some of our simulated futures also assume progress slows down.
  • How long quality takes to catch up.
  • What slows things down, such as licensing, regulation, liability and customers wanting a person.
  • How fast businesses adopt AI. The US Census Bureau reports that business AI use “hovered between 17% and 20%” from December 2025 to May 2026.

The middle of those 10,000 futures is the median year. The middle 80% of them is the range. We always show both. If the median is after 2060, we say “Not foreseeable before 2060,” because nobody can honestly forecast further out.

Full details on the timeline.

Step 5: Roll it into one number

The Still needs a human score combines how much of the job AI can do with how well it does it. It runs from 0 to 100, and higher is safer. Each band has a one-word answer to “Will AI replace this job?”:

ScoreWill AI replace it?
80–100Nah.
60–79A little.
40–59Partly.
20–39Mostly.
0–19Largely.

We never show the score alone. The three answers sit beside it on every job page.

Full details on the headline score.

What we never do

  • We never say a job “will be eliminated”.
  • We never show a year without its range, or a quality score without its grade.
  • We never invent a statistic. Every number on a job page comes from the data.
  • We never let a commercial relationship change a score.

What our scores can’t tell you

  • Your job is not the average job. Your tasks, employer and sector may differ.
  • Usage data only sees AI assistants. Automation through other software, like scanning or self-checkout, does not show up, so we add evidence for those jobs where we can.
  • The future is uncertain. Our timelines are estimates, which is why the ranges are often wide.
  • Exposure is not demand. A job can be highly exposed and still grow.

Check our work

Everything is public:

If a score looks wrong to you, tell us.

Sources

How long will your job last?

Answer a few questions about how you actually do your job and get your own Still needs a human score, with the range, the risks and three things to do next.