AJII vs the Anthropic Economic Index: What an Exposure Score Can and Cannot See

Two indices try to answer the same question from opposite ends, and almost nobody has put them side by side, because until recently no single organisation had a reason to.

The Anthropic Economic Index measures what people are actually doing with AI. It reads real usage and reports which jobs' tasks are being handed to a model, in which countries, in what proportions, month by month.

AJII, the AI Job Impact Index, is an initiative of Fikr Foundation. It measures what could be done, and what would stop it. It scores an occupation from 0 to 10 on how much of the work overlaps with what AI can do, and then on whether automating it would be cheap enough and permitted.

One is a photograph of the present. The other is an estimate about a future. Both are built on the same American occupational codes, which means they are joinable, which means one of them can be tested against the other. Fikr publishes one of the two. That is the only reason we can write this piece with the specifics in it, and it is also why the honest version has to be harder on ours than on theirs.

What observed usage can see

Anthropic samples conversations from Claude.ai and from its own API, runs classifiers over them without human beings reading transcripts, and maps each conversation to a task in O*NET, the US Department of Labor's occupational database. From the task it derives the occupation, and weights occupations by US employment and wage statistics. Six reports have been published since February 2025. The latest, in June 2026, covers 10 April to 10 June, samples hourly rather than in weekly blocks, and classifies the type of thing produced across more than thirty output categories.

The findings that come out of this method are of a kind no survey can produce.

Ninety-three percent of Claude conversations produce an artifact — a document, an explanation, a piece of code. Explanations are the largest single category at 17%, documents and reports 15%, guidance 11%. Personal use rises from about 35% of conversations on weekdays to just under half at weekends. In the March 2026 edition, the ten most common tasks had fallen to 19% of all traffic from 24% four months earlier, meaning usage is spreading out rather than concentrating. Coursework fell from 19% of conversations to 12% while personal use rose from 35% to 42%. Forty-nine percent of jobs have now seen at least a quarter of their tasks performed using Claude at some point.

And it revises itself in public, which is the strongest thing about it. The January 2026 edition cut its own estimate of AI's contribution to annual labour productivity growth from 1.8 percentage points to between 1.0 and 1.2. The March edition moved its estimate of how long the rest of the world will take to reach US per-capita adoption from two-to-five years out to five-to-nine.

What it cannot see is exposure. Usage measures where Claude has already been pointed, which is a function of who holds an account, what they bought it for, and what the model happens to be good at. A high task share in an occupation is evidence of adoption, not of vulnerability. It sees no other company's AI. It sees nothing at all about occupations whose workers do not use Claude. A conversation is not a completed task, and no employment outcome appears anywhere in the dataset.

What a theoretical exposure score can see

AJII scores 1,016 O*NET occupations. Its search box accepts about sixty-five thousand job titles, and those two numbers are not the same statement — the O*NET database ships 57,543 alternate titles and 7,953 reported titles, 65,496 lay names pointing into the same 1,016 occupations. We wrote a whole article about that gap in March, and it remains the first thing an economist will check.

The score runs 0 to 10 and decomposes into seven factors, each published alongside the total. Five of them describe the work: routine content, creative and problem-solving demand, physical dexterity, human interaction, and whether the data and technology exist. Two of them describe the world the work sits in: F6, Economic Feasibility and Labor Dynamics, and F7, Social and Regulatory Acceptance. Even if a machine can do this, is it cheaper than the person, and will anyone allow it?

Those two factors are the substantive case for the index, and the case is stronger than the marketing makes it. Daron Acemoglu's objection to the entire academic exposure literature is that it scores technical capability and ignores whether automation pays. 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 everybody else omits — 4.6% of tasks affected over a decade, at an average cost saving of 14.4% on those tasks. The AI Occupational Exposure index has no cost term. Eloundou and colleagues, whose rubric produced the widely quoted finding that around 80% of the US workforce has at least a tenth of its tasks affected, have no cost term and no regulatory term. AJII scores both, visibly, as first-class dimensions.

What it cannot see is adoption. An F6 of any value is a judgement about what a rational firm would do, not an observation of what firms are doing. It has no time series, so it cannot tell you whether exposure is rising. And F6 and F7 are, of all seven factors, the two that are most obviously country-specific — wages and regulators do not travel — yet they are scored once, on a US baseline. A data-entry clerk in Amman and a data-entry clerk in Zurich currently receive the same F6.

The comparison, stated plainly

That last row is the one to sit with. Anthropic ships roughly 664 megabytes of open data under a CC-BY licence. AJII's report states that its seven factors carry different weights and are combined by weighted aggregation, without saying what the weights are, and its terms of use prohibit attempts to reverse engineer the underlying models. No outside researcher can replicate, audit or falsify an AJII score. That is a deliberate choice with a real commercial logic behind it, and it is also the reason the index is not currently citable in academic work. Both things are true at once.

They already disagree, and the disagreement is the interesting part

In March 2026 Anthropic published a companion paper introducing observed exposure — theoretical capability crossed with actual usage, weighting automation-style use above augmentation-style use. The occupation at the top of that ranking is computer programmers, at 75% coverage. Three quarters of what the job involves is demonstrably being done with AI, today, by real people.

And the labour-market signal underneath it is: nothing much. The paper finds no systematic unemployment rise among exposed workers since late 2022. What it does find is that hiring of 22-to-25-year-olds in exposed occupations slowed by about 14% after ChatGPT, and that the highly exposed workers are 16 percentage points more likely to be female, more educated, and earning 47% more than average — not the population that a "job at risk" framing conjures.

Stanford's payroll analysis, working from a different dataset entirely, lands in the same place from the other side. Employment of 22-to-25-year-olds in AI-exposed occupations sits 19% below where it would have been had it tracked their less-exposed peers, on ADP records through June 2026. The authors are explicit that these are descriptive indicators rather than causal estimates. Their fifth finding is the one that matters here: declines concentrate where AI substitutes for human tasks; where AI complements the worker, employment is flat or rising.

Put those together and the shape of the problem is clear. The occupation with the highest observed AI exposure on earth shows no unemployment effect, while the age band inside it shows a large one. A single directional risk score for "computer programmer" cannot represent that. Ours cannot. Nobody's can. The ILO reached the same conclusion from the methodology side and moved its 2025 index away from an automation-versus-augmentation binary toward four gradients, precisely because the binary was misleading.

The join nobody has run

Both indices key on O*NET-SOC codes. Anthropic publishes an occupational identifier per row; AJII publishes onet_soc_code alongside ajii_score and F1 through F7. They join on one column.

That makes one question answerable that neither index can answer alone:

Are people actually using AI most in the occupations AJII says are most exposed — and does that relationship look different in MENA than in the United States?

If observed usage tracks AJII score, the index is validated as a predictor and can say so with a correlation table. If it does not, the divergence is the finding, and the divergences are where the useful product ideas are. Anthropic also publishes an augmentation share per occupation, which gives a second axis: an occupation that is high-AJII and high-augmentation is a reskilling market, while high-AJII and high-automation is a displacement market. Those two need completely different products and the composite score cannot distinguish them.

We have not run it. That is the honest state of play, and it is the gap this article exists to close.

Three cautions belong in the design of that analysis, and anyone else attempting it should carry them too.

The usage index is renormalised between releases. Israel was reported at 7.0 in September 2025. In the May 2026 data, computed from the raw file, it reads 3.05 and ranks 21st, with Australia highest at 6.40. We could not find a published note reconciling the two. Values from different releases are not a time series.

A missing row means "not published", not "zero". Cells are suppressed below sample floors. Most small Arab markets will be missing on most cuts, and treating that as an absence of activity would invert the finding.

Both instruments are American underneath. Anthropic maps to O*NET tasks and weights by US employment statistics. AJII scores O*NET occupations. The ILO's index, which is built on the international classification instead, shows how much that choice costs: the same exposure logic yields 14.3% of youth jobs in high-income countries and 0.9% in low-income ones, with the Arab States at 5.8%. Neither of the two indices in this article can produce that table.

What a founder should do about it

Know which of the two questions your business depends on. If you sell reskilling, you need exposure — what is coming. If you sell a product people use today, you need usage — what is here. Quoting the wrong one at an investor is a tell.

Segment on substitution versus augmentation, not on risk band. That is the split the best available outcome data says predicts employment change. A product built around "high-risk jobs" is targeting a variable that has not been shown to predict anything.

Treat a single exposure number as a hypothesis. It is a modelled estimate with unpublished weights. The reason to use one is to decide where to look, not what to conclude.

Ask any index for its validation table. Ours does not have one. The AI Occupational Exposure index published its full matrix and dataset. Eloundou published the rubric. The ILO published its survey instrument, its sample construction and its expert reconciliation process, on a base of 1,640 surveyed workers and 52,558 assessments. That is the standard, and it is not an unreasonable one.

The comparison in this article is available to Fikr because Fikr owns one half of it. That is an advantage only if we use it to test the index rather than to promote it.

[FIKR TO CONFIRM: whether Fikr Foundation will run and publish the O*NET-SOC join between AJII scores and the Anthropic Economic Index occupational usage shares, including a correlation table and a MENA-versus-US comparison, and whether AJII's factor weights will be published alongside it.]