There are now several published measures of AI exposure by occupation. They disagree with each other, they disagree with this one, and there is usually no way to find out why.
The reason is the unit of analysis.
An occupation is not a thing a machine can do
"Accountant" is not a task. It is a bundle of several dozen tasks, some of which moved years ago, some of which moved last year, and some of which will not move this decade. Rating the bundle produces a single number that is true of nobody's actual day and cannot be traced back to any specific claim.
If you disagree with a score like that, there is nothing to point at. You can say the number feels too high. You cannot say which part of it is wrong, because the score was never assembled from parts.
What we rate instead
ONET decomposes 923 scored occupations into 18,838 task statements, and links every one of those tasks to a controlled vocabulary of 2,087 detailed work activities. The linkage is ONET's, not ours, and it covers 100% of tasks in this release.
We rate the 2,087 activities. Tasks inherit from the activities they map to. Occupations are aggregated upward from their tasks, weighted by O*NET's own importance and relevance ratings. No occupation is ever rated directly.
Two things follow from that choice.
Reproducibility. Rating 18,838 free-text task statements consistently is not possible for any group of people. Rating 2,087 standardised activity statements is, and the result can be re-checked, which is how we were able to measure inter-rater agreement at all.
Traceability. Every occupation score opens into task scores, and every task score opens into the activities behind it, and every activity carries six numbered ratings against a published rubric. A disagreement can always be reduced to a specific claim about a specific activity.
Why the separation of capability from friction matters as much
The second structural choice is splitting the score in two.
Most indices emit one exposure figure. That figure silently combines "can a machine produce this" with "would anyone let it", and those questions have completely different answers and completely different timescales.
We rate capability on its own, 0 to 4. We rate five frictions on their own, 0 to 3 each. Exposed is capability times the absence of friction. Assisted is capability times friction. Untouched is the absence of capability. The three sum to 100% by construction.
That is why this index can say something like: radiology has high capability and high friction, so it is 28% exposed and will change shape rather than disappear. A single combined number cannot express that sentence at all.
The cost of doing it this way
It is slower. Rating 2,087 activities carefully took a calibration pass plus seven independent raters plus a reliability check, and it has to be partially redone every quarter.
It also produces numbers that are harder to headline. "60% of this job is exposed" is a worse headline than "AI will destroy 40% of jobs", and it is the only one of the two that means anything.
Release v2026.Q3, built on O*NET 31.0. The rubric, the rating process and the reliability figures are on the methodology page.