Most occupations in this index are easy to place. Either current systems can produce the work or they cannot. Stonemasonry sits at zero. Telemarketing sits at 73.3%. Neither is ambiguous.
Forty-eight occupations are not like that. They score high on capability and high on friction simultaneously, which puts them in the band we call assisted: the machine can produce the work, and something structural keeps a person in the loop anyway.
| Occupation | Capability | Friction | Exposed |
|---|---|---|---|
| Paralegals and Legal Assistants | 0.66 | 0.52 | 33.6% |
| Intelligence Analysts | 0.66 | 0.54 | 32.4% |
| Genetic Counselors | 0.63 | 0.60 | 27.2% |
| Dietitians and Nutritionists | 0.63 | 0.54 | 30.9% |
| Radiologists | 0.62 | 0.60 | 28.0% |
| Psychiatrists | 0.58 | 0.61 | 25.5% |
Radiology is the case worth examining, because it is the one where the capability question was settled earliest and most publicly, and where the friction has held anyway.
The capability side is not in dispute
Reading images to inform a diagnosis scores 3 out of 4 on our capability scale, which in the rubric means working quality output requiring human review and sign-off. That rating is not controversial and has not been for some time.
If capability were the whole story, radiology would sit near the top of this index. It sits at the 34th percentile.
What the friction is made of
Radiologists' tasks carry an accountability rating that is among the highest in the index. A diagnostic read is not a document. It is a clinical act attributed to a licensed individual, entering a medical record, carrying liability, and in most jurisdictions legally required to come from a person holding a specific registration.
Verification cost compounds it. A missed finding is expensive, dangerous and frequently invisible until much later. Our rubric scores that combination at the top of the scale, and it is the single strongest brake on deployment in medicine generally.
Those two frictions do not weaken because a model improves. They weaken when a regulator or a professional body decides they should, which happens on a timescale measured in years and after evidence that this index does not attempt to generate.
Why the assisted band is the interesting one
There is a tempting conclusion here, which is that licensure protects jobs and that professions with a registration requirement can relax. That is not what the data supports.
What the assisted band describes is a job changing shape while the headcount holds. If a system produces a working-quality first read and a radiologist's role shifts toward adjudicating, prioritising and taking responsibility for the cases that matter, the work has changed enormously without a single job being eliminated. The index would still show 28%.
This is also why we refuse to publish a single exposure number per occupation and call it a risk score. Radiologists and stonemasons both score low. They are low for completely different reasons, and only one of those reasons could change with a regulatory consultation.
The honest limit
We do not model adoption, and nothing here predicts what any health system will do. What the index gives you is the shape of the problem: where capability already exists, where friction is doing the work of holding the line, and which kind of friction it is. For the forty-eight occupations in this band, the answer to "will AI take this job" is almost certainly no, and the answer to "will this job look the same in five years" is almost certainly also no.
Figures from release v2026.Q3. Capability and friction are judgments against a published rubric; the banding rule and the inter-rater reliability of every dimension are on the methodology page.