Method

How the number is calculated

Last updated: 14 August 2026

The check on the homepage shows one number: how much of your work is within reach of current AI. It is built from two things, a published measure of your occupation and the six decisions you make during the run. This page is the whole calculation, including the parts we are least sure about.

1. Where the occupation number comes from

The source is the O*NET 30.3 Database, published by the US Department of Labor, Employment and Training Administration under a Creative Commons Attribution 4.0 licence. For every occupation it rates 41 generalized work activities for importance, on a scale from 1 to 5, from surveys of people who actually do the job.

We take the 15 activities each occupation rates most important as its core. Each of those activities carries a weight from 0 to 1 for how much of it current AI systems can carry out. The occupation's exposure is the share of core importance sitting on activities at or above 0.6.

exposure = sum of importance on core activities within reach
          ÷ sum of importance on all core activities × 100

This is deliberately the same shape as the IAB Job-Futuromat, which computes automatable core tasks divided by all core tasks for about 4,000 German occupations. If your job appears there, comparing the two numbers is a fair test of ours.

Why a fixed core of 15 and not a cutoff. An importance cutoff gives different occupations different denominators. A cook rates only a handful of activities above the line, so a single activity moves the number by twenty points; our first build put cooks at exactly zero, which was arithmetic rather than truth. A fixed core keeps occupations comparable.

2. The range, and why there is one

Every occupation is scored three times, with a core of 12, 15 and 18 activities. The range you see is the spread across those three runs, never narrower than four points. A wide range means the answer depends on where you draw the line around a job, and you should trust it less.

3. Which activities count as within reach

This is our judgement, and it is the weakest part of the model. It is written down as data, one weight and one reason per activity, and it is versioned in the repository so it can be argued with and corrected rather than hidden.

The weights follow one line: current systems are strong at producing and processing language and data, and weak at anything that needs a body in a room or accountability that someone carries. Writing things down scores 0.90. Analysing data scores 0.80. Making decisions scores 0.40, because a system proposes well and carries none of the consequences. Coaching people, handling conflict and physical care sit at 0.15 and below.

4. How your own decisions move it

The six decisions are scored the same way the full check scores them, by the same domain weights, giving a human edge from 0 to 100. That edge then moves your occupation's baseline:

replaceability = baseline × (1 − 0.35 × (edge − 50) ÷ 100)
bounded to within 15 points of the baseline

So the occupation dominates and you are not stuck with it. An average run leaves the number where it started. A strong run pulls it down by up to 15 points, a weak one pushes it up by the same amount. Six decisions cannot tell us more than that, so the model does not pretend they can.

5. What this does not claim

It is not a prediction. It does not say your job will be automated, when, or by whom. The Job-Futuromat puts the same warning on its own numbers, and it is right.

It is not about your employer. Whether a task that can be automated will be automated depends on cost, regulation, union agreements, customer expectations and management appetite. None of that is in here.

It is not a measure of you beyond the six decisions you made. It is a measure of the work, plus a small, bounded read on how you handle six specific situations.

It uses US occupational data for occupations everywhere. Work is organised differently in different countries, and a German electrician's day is not identical to an American one's.

6. What we do not show you

You will not see a claim that some percentage of other people picked the same answer, or that you rank above a share of your peers. We could generate both. We do not have the data to support either yet, and a number nobody can check is worth nothing on a page about honesty.

7. Sources

  • O*NET 30.3 Database, US Department of Labor, Employment and Training Administration, used under CC BY 4.0. Humans Verified has modified it by grouping occupations into families and scoring their work activities.
  • IAB Job-Futuromat, Institut für Arbeitsmarkt- und Berufsforschung, for the method's shape and as a second opinion on German occupations.
  • Felten, E., Raj, M. and Seamans, R. (2021), Occupational, industry, and geographic exposure to artificial intelligence, Strategic Management Journal 42(12), 2195 to 2217, as related work. Its values are not used here.
  • Anthropic Economic Index, for a current read on which tasks people actually hand to AI.

Take the check