Type a job title into an AI risk calculator and it gives you a number. The number describes someone in Ohio.
That is not a complaint about one product. It is a description of every exposure index published anywhere, including the one Fikr Foundation publishes. All of them are built on the same American database, and almost nobody who quotes their output knows that.
An exposure index answers a narrow question: how much of this job overlaps with what the technology can already do? To answer it, you first need a machine-readable description of what the job involves. Only one such description exists at scale, and the United States Department of Labor maintains it.
The database everybody is standing on
O*NET is the Occupational Information Network. It decomposes every occupation in the American economy into tasks, skills, abilities, work activities and work context, and it is free. The current taxonomy contains 1,016 occupations.
Follow the frameworks back and they all end up there. AJII scores those 1,016 O*NET occupations on seven factors. The AI Occupational Exposure index maps ten AI application areas onto 52 O*NET abilities. Eloundou and colleagues scored O*NET tasks against whether a language model could cut the time to do them in half, and got the widely quoted result that around 80% of the US workforce has at least a tenth of its tasks affected. Frey and Osborne's 47%, still the most-repeated number in this field, came from O*NET bottleneck variables in 2013. Anthropic's Economic Index, which measures actual usage rather than potential, classifies real conversations against O*NET tasks before rolling them up to occupations.
There is a real benefit to this. When four frameworks disagree, they are at least disagreeing about the same objects.
The cost arrives when the score crosses a border. A job title is a container, and what is inside it is set by the labour market, not by the label.
Two in five
Here is the size of the gap, in one statistic.
Around 40% of employment in Jordan is in the public sector. In Algeria it is 31%, in Egypt 25%. Among employed nationals the Gulf figures are higher still: Qatar 87%, Kuwait 86%, Saudi Arabia 72%. Set those against the comparators in the same study — Indonesia 8%, Turkey 15%, Malaysia 17%, China 29%.
A caution on vintage, because it matters. Those figures come from Assaad and Barsoum, published in 2019 on data from the 2000s and 2010s. They are the standard citation in this literature and they are the most recent comparable series we could find, but they are not current, and anyone using them should say so.
Now put that next to what the exposure indices actually find. Clerical work is the single most exposed occupational category in every framework that reports one. The ILO's 2023 index put 24% of clerical tasks at high exposure and another 58% at medium — meaning most of the work in that category overlaps to some degree with what a language model does well. Public administration is clerical-heavy almost by definition.
So in Jordan, the exposure lands on the segment that employs two in five workers, that young graduates have historically queued for, and that is politically the hardest thing in the country to restructure. In the United States, the same occupational scores land somewhere else entirely, because in the US economy the graduate queue does not form outside a ministry. The score cannot represent that difference. It was not built to.
The same index, a sixteen-fold spread
The ILO ran the arithmetic country by country, and the result is the clearest available proof that an exposure score is not a property of a job.
In August 2026 the ILO applied its own generative-AI exposure index to detailed occupational distributions for each country — that is, it asked what share of a given country's actual jobs sit in the most exposed categories. Globally, 6.1% of youth jobs. Then the spread: 14.3% in high-income countries, 7.6% in upper-middle, 3.1% in lower-middle, 0.9% in low-income. By subregion, Northern, Southern and Western Europe reach 14.9% and Northern America 14.3%, while the Arab States sit at 5.8% and Northern Africa at 4.0%.
Same index. Same scoring logic. A gap of roughly sixteen times between the top and bottom income groups, produced entirely by which jobs exist where and in what proportion.
Read that carefully, because the obvious reading is wrong. Low exposure is not good news. The Arab States and Northern Africa also carry the highest youth unemployment rates in the world, 26.2% and 22.6%, with roughly a third of young people not in employment, education or training. The region is less exposed to AI because it has fewer of the jobs AI touches, which is another way of saying it has fewer of the jobs that pay well.
The World Bank found the same inversion from a different direction. Across 25 low- and middle-income countries covering three and a half billion people, high exposure reached 12% of workers in the low-income group and 15% in the lower-middle group — and exposure was higher for women, for urban workers and for the more educated. The paper also notes that lack of electricity access limits effective exposure, which is a sentence worth sitting with. Its authors state plainly that exposure "does not equate to job loss."
The two factors that are country-specific are scored once
This is where we have to be specific about our own instrument.
AJII scores seven factors. Two of them 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 paying the person, and will anyone permit it?
Those two factors are genuinely the strongest thing about the index. Daron Acemoglu's central objection to the academic exposure literature is that it ignores whether automation is profitable. His model bounds AI's ten-year effect on total factor productivity at no more than 0.66%, and it gets there by insisting on the two terms everyone else drops: the share of tasks actually affected over a decade, which he puts at 4.6%, and the average cost saving on those tasks, 14.4%. The AI Occupational Exposure index contains no cost term. Eloundou's rubric contains no cost term and no regulatory term. Frey and Osborne had engineering bottlenecks and no economics at all. AJII scores the thing Acemoglu says is missing.
It scores it once, in one country's wage structure and one country's regulator.
Whether automating a task is cheaper than the worker depends on that worker's wage. Whether it is permitted depends on a ministry, a labour law, and a public-sector union. A data-entry clerk in Amman and a data-entry clerk in Zurich currently receive the same F6. They should not.
The institution decides, not the technology
One more piece of evidence, because it settles the question of whether a better score would fix this.
Two studies measured what happened to real workers after generative AI arrived. In Denmark, Humlum and Vestergaard linked adoption surveys to administrative labour records covering 25,000 workers and 7,000 workplaces across eleven exposed occupations, and found nothing: a precise null on earnings and hours, ruling out effects larger than 2% two years after adoption, with users saving an average of 3% of their time. On a large online freelance platform, Hui, Reshef and Zhou found the opposite — a 2% drop in jobs and a 5.2% drop in monthly earnings in the most affected occupations.
Both results are real, and the occupations overlap. The difference is not the technology. It is employment protection, firm-specific capital, and how fast demand reallocates when a task gets cheaper. Danish payroll employment has all three. A freelance platform has none.
That is the finding an exposure score structurally cannot carry, and it is the one that decides what happens to a person. It also has an obvious regional reading: two-thirds of MENA youth employment is paid work, and the vast majority of those young paid workers are informal. The region's youth employment-to-population ratio is 18.5%, the lowest of any world region and half the global rate. Neither the Danish answer nor the platform answer transfers cleanly, and there is no way to interpolate between them.
What a founder should do about it
Say which country produced the number. If you quote an exposure figure in a MENA deck, name the occupational structure behind it. "Fourteen percent in high-income countries, four percent in North Africa, same index" is a stronger sentence than any single percentage, and it survives the question that follows.
Ask for the crosswalk. National labour statistics across the Arab world are kept in ISCO, the international classification. Every index named here is kept in O*NET-SOC, the American one. AJII publishes no crosswalk between them, which means an Arab labour ministry cannot join our scores to its own data. The ILO's index is ISCO-native, which is precisely why the ILO could produce the country table above and we could not.
Build the local denominator before you build the product. If your business depends on how many people in Amman or Cairo hold an exposed job, that number does not currently exist. Producing it is a better use of six months than citing somebody else's.
Sell to the segment, not to the score. The public sector is 40% of Jordanian employment and the most clerical-heavy part of it. That is a procurement target, a reskilling target and a political constraint all at once, and none of those three facts is visible in a risk band.
The honest position on our own index is that it is a US instrument published from Amman. Every competing framework has the same problem. The difference is that we are the ones who could fix it.
[FIKR TO CONFIRM: whether AJII will publish an ISCO-08 crosswalk and country-adjusted F6 and F7 scores for Jordan, Egypt, Saudi Arabia and the UAE, and on what timetable.]