Research
What the market actually asks for
A live corpus of job postings, measured rather than assumed. It is why the item banks test what they test — and it is research, not a number on your report.
What the corpus shows
The requirement is real. Its size is not settled.
Measured two ways from the same corpus — a raw-text scan and an extracted-skill scan. They disagree by roughly six-fold, which is the honest headline: the direction is unambiguous and the level is not.
1.4–8.1%
of the 998,166 postings dated 2026-07, measured on 2026-08-11, carry an explicit AI-usage requirement, read two ways. Over the whole corpus of 2,151,213 active postings measured the same day, the extracted-skill instrument reads 1.62%, and that is a different denominator rather than a different answer.
79.7%
of the 34,743 postings carrying an AI-usage requirement, measured on 2026-08-11, sit in Technology, with finance and consulting growing fastest.
~2×
growth in the raw-text requirement rate from 2026-02 to 2026-07, over 445,637 and 998,166 postings, both measured on 2026-08-11.
300
postings per arm to be hand-adjudicated before we publish a point estimate. That work is not finished.
Coverage
What we publish, and what we admit
The honest version of the coverage table: where a bank is grounded in enough postings to carry a benchmark, and where it is not yet.
| Occupation | SOC | Bank | Selectable today |
|---|---|---|---|
| Software engineering | 15-1252 | Live | Yes |
| Financial analysis | 13-2051 | Live | Yes |
| Management consulting | 13-1111 | Live | Yes |
| Data & analytics | 15-2051 | Live | Yes |
| Product management | 11-3021 | Live | Yes |
| Marketing | 13-1161 | Live | Yes |
| Journalism | 27-3023 | Live | Yes |
| Healthcare revenue cycle | 29-2072 | Live | Yes |
| Legal operations | 23-2011 | Live | Yes |
| Accounting & audit | 13-2011 | Live | Yes |
| Underwriting & claims | 13-2053 | Live | Yes |
| Supply chain | 13-1081 | Live | Yes |
Being exact, because this is the kind of table people buy against: no report carries a benchmark percentile. The corpus grounds what the banks test; nothing in the product compares a candidate against it, and there is no scheduled re-derivation job.
The honest part
What a posting corpus cannot tell you.
It cannot tell you whether this person will be good at your job, and no report compares a candidate against it. It grounds the assignments in what the occupation required this quarter. At best it says how the demands on an occupation are moving — a narrower and far more defensible claim than a percentile implies.
It also describes what employers *wrote*, not what they do, and not who answered. That gap is why our two instruments disagree by roughly six-fold, and why we would rather show you the bracket than pick the end of it that sells better.
https://olive.is/benchmarks · Edition 2026 Q3How a requirement is counted
A posting counts only on an explicit usage requirement — "must use", "expected to leverage", a named tool as a duty. A mention of AI in the company blurb does not count.
The research
Read the measurement itself
The corpus, both instruments, the disagreement between them and every caveat we know of — written up rather than summarized into a number.