Verellax measures the distance between what organisations say about AI and how they have designed the work around it.
What Verellax is
Verellax measures the distance between what organisations say about AI and how they have designed the work around it. It reads three public records side by side: the job specifications employers publish; what their executives say in results, interviews, and keynotes; and the corporate signals in their filings and investments. Each role is read on whether it involves AI, whether AI sits at the centre of the work, and whether the role is accountable for what the model produces. Verellax does not survey anyone. Every figure is read from what an organisation has already published and carries its date, and every projection names the measures it is built on. Organisations use the Verellax Diagnostic to see where their roles stand against their sector, and against the minority that has designed the work furthest, before they redesign it.
Edition One reads 57,134 roles advertised in Q3 2026 (1 July to 30 September) by 284 employers in seven markets. The index between editions re-reads the live register every Monday. Every figure we publish carries its population, its basis and its date in the same block.
Measure the work before you redesign it.
How we measure
Every AI-central role is then read on four tests. Accountable is one of them.
Works with the model.
Can overrule it.
Is accountable for what it produces.
Keeps decisions over it.
A market publishes only with 150 or more live specifications and the other eligibility floors the method page names. Under 30 specifications a cell prints its count alone; under 150, its share prints with the count and base beside it. A sector publishes only with 10 or more companies. Nothing is estimated to fill a cell.
Read how we measure and how we correct.
Edition One
Edition One published on Friday 2 October 2026. AI is involved in more than one in three roles, but almost four in five AI-central roles do not say who is accountable for what the model produces.
It reads 57,134 roles advertised in Q3 2026 by 284 employers in seven markets. Each role is read on whether it involves AI, whether AI sits at the centre of the work, and whether the role is accountable for what the model produces. Every AI-central role is then read on four tests. It names no employer. Among its findings:
- Roles designed for people who use AI say who is accountable a third as often as roles designed to govern it.
- Financial services says who is accountable for the output of the model more than twice as often as advertising agencies.
- Fewer than one in ten AI-central roles say who can overrule the model.
- The US involves AI in almost twice the share of roles that Australia does, but all three markets say who is accountable at about the same rate.
- Sectors design the same function in very different ways.
Read it free at verellax.com/index/edition-one. Journalists can have the PDF without the sign-up form: write to hello@verellax.com.
On request, any journalist can have a cut the edition does not print, for their market, sector or function, read on the edition's own base, and first access to the founder to talk it through.
Quote the edition for the quarter. Figures on the index, markets and sectors pages are live reads that re-read every Monday, so they will differ from the edition; carry the as-at date with any live figure and the two can never be confused.
Assets
Exhibits from Edition One
Five exhibits, one per finding, exactly as Edition One prints them. Each carries its source line; reproduce it whole under the terms below.
Exhibit 1 · PNG 1654 wide · source line in the imageDownload
Exhibit 2 · PNG 1654 wide · source line in the imageDownload
Exhibit 3 · PNG 1654 wide · source line in the imageDownload
Exhibit 4 · PNG 1654 wide · source line in the imageDownload
Exhibit 5 · PNG 1654 wide · source line in the imageDownload
Wordmark and mark



The wordmark ships as PNG masters; the vector recut is in production and will be posted here. Clear space is built into every master: the placed image runs three times the intended mark height. Do not retype the wordmark, recolour it, or set it in another face. The stop is a brand device; in running copy the name reads Verellax, capital V, no stop.
The lockup, mark and wordmark together
The mark sits left of the word at one and a half times the wordmark height, one half of that height of air between them. Use the lockup where one image has to carry both; use the wordmark alone where the mark already appears on the surface.
placeholder
PNG 1600 × 2000 · supplied
Founder and bio
Supplied by the founder. Use it whole or not at all; do not paraphrase it into a claim the site cannot support.
Contact and terms
Press enquiries, interviews with the founder, and requests for a cut the edition does not print: hello@verellax.com.
Cite the edition as "Verellax Index, Edition One, 2 October 2026", and a live figure as "Verellax workforce intelligence, week ending" with its date. Where you reproduce a figure, carry its population and its basis with it, as the exhibits do.
No named employer carries a figure. Peers in a sector are pseudonymised and labelled as such, in the form “Peer C, insurer”.
Reproduction
Verellax exhibits may be reproduced in editorial coverage, in full and unaltered, with attribution in the form "Verellax Index, Edition One, 2 October 2026", or for a live figure, "Verellax workforce intelligence, week ending" with its date.
Each exhibit carries its source line in the image. Reproduce it whole: do not crop, recolour, retype or otherwise alter it, and do not remove the source line.
Where you quote a figure rather than reproduce an exhibit, carry its population and its basis with it, as the exhibits do.
This grant is for editorial use. It does not cover redistribution of the underlying record, resale, or inclusion in a commercial product. For anything else, write to hello@verellax.com.
A custom press grant, not CC BY 4.0. CC BY permits adaptation, and adaptation includes cropping, which the source line cannot survive. The same grant is stated on Terms.



