How we measure
We estimate how much human work time AI can already save without sacrificing quality.
AI WORK SCORE
score = Σ (weight × saving)
Task weights add up to 100% of the workweek. The final score is calculated, not entered by hand.
30% of the week
80% saved with AI
24 points
What it measures
If a task takes 100 minutes and current AI can reduce it to 20 without sacrificing quality, the time saving for that task is 80%. The job is the sum of those tasks by how much of the week each one takes.
What it does not measure
It does not say you will be let go. It does not estimate how many jobs disappear. It is not “AI can help a bit”. A two-minute spellcheck does not make a job 80%.
Tasks
Tasks come from O*NET. O*NET does not publish what share of the week each task takes. Importance (IM) is not time; we do not use it as weight. Frequency (FT) says how often, not how long. We turn FT into relative time and normalise to 100%: that is our estimate, not an “O*NET weight”. One page = one O*NET code. “Junior developer” is Computer Programmers (15-1251.00), not Software Developers in general. The time saving is estimated separately and is never adjusted to fit the headline score.
Origin
Public: O*NET, ESCO, benchmarks (SWE-bench, LiveCodeBench…), datasets, papers, statistics. Public evidence must include a source. Estimate: no source is precise enough. Public + estimate = 100% of the week. Internal test evidence will only be included once sufficient results are available.
Confidence
High: most of the modelled workweek is backed by recent public evidence. Medium: a mix. Low: almost all estimate. It is not a scientific probability.
Sources
O*NET and ESCO describe the work. Benchmarks and papers say what AI can already do. Each page shows the O*NET and ESCO/ISCO code when we have it, and each task a “See evidence” link if there is a URL.
PUBLISHED
NOT PUBLISHED