AJII scores 1,016 occupations. The search box accepts about sixty-five thousand job titles, and those are not the same statement.
The AI Job Impact Index is an initiative of Fikr Foundation. It publishes a single exposure score from 0 to 10 for a job title, plus seven factor scores underneath it, and it is free to use. Its public site says the index covers "over 65,000 job titles," and in one place describes the research as covering "65,000+ occupations." Its own methodology report says something narrower and more accurate: "AJII covers 1,016 job titles 'O*NET job titles'."
Both numbers are real. Only one of them is the unit of analysis.
The ONET-SOC 2019 taxonomy, maintained by the US Department of Labor, contains exactly 1,016 occupations. Version 30.2 of the ONET database ships an Alternate Titles file with 57,543 rows and a Sample of Reported Titles file with 7,953 rows — 65,496 lay names, all of them pointers into those same 1,016 occupations. So "65,000+" describes how many ways you can spell your job before the index recognises it. It does not describe how many jobs were independently scored.
We are stating this ourselves because the alternative is that somebody else states it first. A labour economist reading the site and then the report would find the gap in under a minute, and would be right to.
What the number does and does not buy you
The honest claim is about the interface, not the resolution.
AJII's scored resolution is the O*NET occupation. That is the same resolution as the AI Occupational Exposure index, and coarser than Eloundou and colleagues, who score at the task level beneath the occupation. Anyone positioning AJII as offering finer granularity than those frameworks is describing a product that does not exist.
What AJII does that they do not is resolve the words a person actually types. AIOE is a CSV keyed on Standard Occupational Classification codes, published on GitHub for researchers. Eloundou's exposure data is an academic appendix. ILO Working Paper 140 is a working paper. None of them will tell an eighteen-year-old in Amman what "graphic designer" means for her. That distribution gap is real and it is worth something. It is just not a methodological advantage, and it should never be sold as one.
What no exposure index can see
Every index in this field measures the same thing: how much of a job overlaps with what the technology can do. None of them measures what happens to the worker. The authors say so themselves, in their own papers, and are routinely quoted as though they had not.
Eloundou and colleagues state that their measure is a proxy for potential economic impact and is agnostic between labour-augmenting and labour-displacing effects. The World Bank's 2025 paper covering 25 low- and middle-income countries — 12% of workers highly exposed in the low-income group, 15% in the lower-middle group — says plainly that exposure "does not equate to job loss." The ILO's 2025 revision, built on 1,640 surveyed workers and 52,558 individual assessments, moved away from an automation-versus-augmentation binary and toward four gradients, precisely because the binary was misleading. That index finds one in four workers globally in an occupation with some exposure, and a spread from 11% of employment in low-income countries to 34% in high-income ones.
Then the outcome data arrived, and it made the distinction decisive rather than academic.
The Stanford Digital Economy Lab's analysis of ADP payroll records through June 2026 finds employment of 22-to-25-year-olds in AI-exposed occupations sitting 19% below where it would be had it tracked their less-exposed peers. The authors are careful: these are descriptive indicators, not causal estimates. But the fifth of their six findings is the one that matters here. Declines concentrate in occupations where AI substitutes for human tasks. Where AI complements the worker, employment is flat or rising.
A single directional risk score cannot carry that distinction. Ours included. The seven-factor breakdown mitigates it — a user can see whether a score is driven by routine content or by data availability — but the headline number, and the Low/Medium/High band on top of it, point one direction only.
Where ours is genuinely different
There is one axis where AJII is not playing catch-up, and it is worth being specific about it.
Two of the seven factors are F6, Economic Feasibility and Labor Dynamics, and F7, Social and Regulatory Acceptance. In plain terms: even if a machine can do the work, is it cheaper than the person, and will anyone allow it?
That is exactly the hole Daron Acemoglu identified in the academic exposure literature. His task-based macro model bounds AI's total factor productivity effect at no more than 0.66% over ten years, and it gets there by insisting on the two terms the exposure indices omit — the share of tasks actually affected, which he puts at 4.6% over a decade, and the average cost saving on those tasks, 14.4%. AIOE contains no cost term. The Eloundou rubric contains no cost term and no regulatory term. Frey and Osborne, whose 47% is still the most-quoted number in this field, had engineering bottlenecks but no economics at all.
AJII scores cost and regulation as first-class, visible dimensions. On that specific axis its conceptual frame is closer to Acemoglu's than the indices it sits beside. That is the defensible claim, and it is the only one of its kind we can make.
What we cannot defend
Three things, stated because they are true.
The weights are not published. The report says the seven factors "have a diffrient weight" and are combined by weighted aggregation. It does not say what the weights are, and AJII's terms of use prohibit attempts to reverse engineer the underlying models or datasets. Compare: AIOE publishes its full application-to-ability matrix and dataset. Eloundou published the rubric. ILO WP140 published its survey instrument, its sample construction and its expert reconciliation process. The practical consequence is that no outside researcher can replicate, audit or falsify an AJII score. There is a real tension here — the proprietary logic that makes the index commercially useful is the same logic that makes it academically uncitable — and it is a choice, not an oversight.
There is no published validation. No inter-rater reliability. No reported correlation against AIOE, against Eloundou, or against the ILO gradients. No back-test against observed employment change, which is now possible, because the ADP series provides an outcome variable that any exposure index can be tested against. A single correlation table would close most of this gap and would cost very little.
O*NET is a US instrument, and we publish from MENA. Every framework in this field inherits that problem. It bites hardest for the one organisation positioned to fix it. Public sector employment was around 40% of the total in Jordan, and 72% to 87% among employed nationals in Saudi Arabia, Kuwait and Qatar — figures from Assaad and Barsoum, published in 2019 on data from the 2000s and 2010s, and the most recent comparable series we could find. Our own report states that clerical and routine processing occupations score highest on AJII. In a labour market where the state is the largest employer and public administration is clerical-heavy, that puts the exposure in exactly the segment that is politically hardest to restructure. Meanwhile the region's youth employment-to-population ratio is 18.5%, the lowest of any world region and half the global rate. None of that is in the score. AJII publishes no crosswalk to the international occupational classification and no regionally adjusted figures.
What to do with a score like this
If you are building a product on top of an exposure index — ours or anyone's — four rules.
Never cite a single exposure number as a forecast. It is not one. The paper you took it from says so. Quote the caveat in the same sentence as the number, or leave the number out.
Ask which factor moved the score. An occupation scoring high on routine content and an occupation scoring high on data availability need different products. A composite hides that; the decomposition is the useful part.
Segment on substitutes versus complements, not on risk. That is the split the best available outcome data says predicts employment change. A reskilling product built around "high risk jobs" is targeting the wrong variable.
Measure in the market you sell into. Every occupational input behind every index named here was collected in the United States. Nobody has adjusted one of them for an Arab labour market. If your business depends on the answer, the answer does not currently exist, and producing it is a better use of six months than citing somebody else's.
AJII's own report puts it more plainly than any of our marketing does: it is "a diagnostic instrument not a predictive oracle," its outputs are "informational only" and "not absolute facts." Writing that into a PDF is cheap. Publishing the weights and a correlation table against three competing frameworks is not cheap, and it is the thing that would make the index worth citing.
[FIKR TO CONFIRM: whether AJII is being optimised as a proprietary commercial asset or as a citable public instrument, and, if the answer is the latter, the timetable for publishing factor weights and a validation table against AIOE, Eloundou et al. and ILO WP140.]