What the New Anthropic Economic Index Shows Us About How People Actually Use AI

What the New Anthropic Economic Index Shows Us About How People Actually Use AI

The Anthropic Economic Index connector went live on July 22, 2026, giving anyone the ability to query Anthropic's usage measurement in plain language. I put a structured set of questions to it covering substitution patterns, skill concentration, geographic diffusion, and time savings, working against the May 2026 period, which is the latest published data.

Several results reward attention. Others matter mostly because of what they rule out.

Information retrieval sits above software

The ranked task list is dominated by search and reference work. Searching electronic sources, databases, or repositories for information accounts for 4.95 percent of sampled work conversations. Searching standard reference materials to answer patrons' reference questions accounts for 3.74 percent. Recommending and advising on products and services accounts for 2.25 percent, and answering user inquiries about software or hardware operation accounts for 1.98 percent. Writing or modifying programs to meet customer requirements appears fifth at 1.47 percent.

Anyone who followed the earlier Economic Index reports will find that ordering surprising. In the January 2026 report covering November 2025 usage, modifying software to correct errors alone represented roughly 6 percent of consumer conversations.

The explanation is measurement scope rather than a collapse in coding activity. Anthropic's March 2026 report documented coding work migrating out of Claude.ai and into first-party API traffic, where Claude Code's agentic architecture splits a single piece of work into many smaller calls that get classified as separate tasks. The connector draws on Claude chat and Cowork usage with no API traffic included. Coding did not shrink. It moved to a surface this feed does not observe.

That has a practical consequence for anyone citing these numbers. The visible task distribution describes what people bring to a conversational interface. The work most likely to be automated outright is running somewhere else.

The top 50 tasks together account for roughly 48 percent of sampled work conversations, and they cluster tightly around search, advise, answer, write, and edit. Physical work, supervisory work, and judgment exercised under professional liability are close to absent.

The occupational gap against real employment

The categorical breakdown makes the skew concrete. Computer and Mathematical work accounts for 23.8 percent of observed usage. Arts, Design, Entertainment, Sports, and Media accounts for 13.6 percent. Educational Instruction and Library accounts for 12.8 percent. Those three categories together exceed half of everything the Index sees.

At the other end of the distribution, Construction and Extraction accounts for 0.10 percent, Building and Grounds Cleaning and Maintenance 0.13 percent, Transportation and Material Moving 0.34 percent, and Farming, Fishing, and Forestry 0.04 percent.

Joining this to Bureau of Labor Statistics employment data sharpens the picture considerably. The May 2025 Occupational Employment and Wage Statistics release put construction and extraction employment at 6.4 million workers, or 4.1 percent of total national employment. Computer and mathematical occupations employed 5.26 million, which works out to roughly 3.4 percent of the national total.

So computer and mathematical work is over-represented in Claude usage by a factor of about seven relative to its employment share, while construction and extraction is under-represented by a factor of roughly forty. The Index has no wage or employment fields of its own, so this comparison is a join the analyst performs rather than a finding Anthropic publishes.

Coursework is the sharpest international divide

The clearest signal in the geographic data has little to do with occupational categories. It shows up in the split between work, personal, and coursework conversations.

Algeria registers 47.1 percent coursework. Tunisia registers 51.5 percent. Indonesia 45.9 percent, Bolivia 43.3 percent, Peru 41.9 percent. Against that: Japan at 5.9 percent, Canada at 8.3 percent, the United States at 8.9 percent.

A country where more than half of AI conversations are students doing coursework is at a different point on the diffusion curve than one where the figure is under one in ten. It also means the occupational task shares for those countries describe something closer to a curriculum than a labor market.

This divide deserves treatment on its own terms rather than as a footnote to adoption rankings. If a cohort of students in North Africa and Southeast Asia is learning to work with these systems during their education, the labor market effects arrive later and through a different mechanism than the professional adoption visible in high-income countries.

Automation share varies more by geography than the global number suggests

Globally, 51.4 percent of conversations are classified as augmentation, where the person stays involved in the work, and 48.6 percent as automation, where the person directs Claude to complete it. These are conversation styles rather than employment outcomes, and the connector is careful to say so.

The country-level spread runs about 16 points. Zimbabwe sits at 57.6 percent automation, Mongolia 57.4 percent, Ghana 54.7 percent, Kenya 54.6 percent. Norway sits at 41.3 percent, Denmark 42.3 percent, Belgium 42.7 percent.

The pattern matches what Anthropic reported in earlier work: countries with lower usage per capita lean toward directive, delegate-the-whole-task interaction, while high-adoption countries lean toward iteration. Anthropic's own analysis found the relationship survived controlling for each country's task mix, which argues against dismissing it as a composition effect.

The connector data suggests the mix still explains a meaningful share of the variation. Countries with heavy coursework usage show more automation, and coursework is more directive by nature. Both things can be true. Treating the spread as evidence of national AI maturity would be reading more into it than the data supports.

The usage rankings need population held in mind

The Anthropic Usage Index divides a geography's share of Claude usage by its share of working-age population. Australia leads at 6.40, followed by Singapore at 5.81, Switzerland at 5.02, Luxembourg at 4.85, New Zealand at 4.84, and Canada at 4.13.

The United States sits at 3.87 while accounting for 20.2 percent of all observed usage. India sits at 0.30 while accounting for 7.1 percent.

Small-population geographies rank high on this measure by construction, so the leaderboard rewards Luxembourg and penalizes India in ways that have nothing to do with adoption intensity in absolute terms. There is no GDP or income field in the Index, so the question of whether adoption tracks income per head is another join the analyst has to make.

Concentration within countries adds a dimension the ranking misses. Every covered country shows Computer and Mathematical as its top category, but the share varies. Cambodia registers 32.9 percent, Tunisia 31.4 percent, Egypt 30.5 percent, and Israel 30.4 percent. Australia registers 21.7 percent, the United States 21.1 percent, and the United Kingdom 20.2 percent. High-usage countries show flatter distributions across occupational categories, while low-usage countries concentrate in software and coursework. Israel is a useful check on any simple income story, sitting at high income with high concentration.

The strongest single number, with a caveat attached

A classifier estimates that tasks which would take a person working alone roughly 5 hours ran to roughly 40 minutes of conversation with Claude. That is a ratio of 7.5 to 1.

The bucketing matters. Both figures are automated estimates derived from conversation content and grouped by order of magnitude, so the ratio is directionally meaningful and precisely wrong. It gives an anchor without giving a dollar figure, since there is no wage data in the Index to multiply against.

It also measures conversation time rather than end-to-end task completion, which excludes review, revision, and the cases where the output was discarded.

What the data cannot support

Several questions returned nothing usable, and the boundaries are worth stating clearly. The published dataset covers April and May 2026 with no internal trend series, so it cannot show whether automation is rising or which tasks moved fastest. O*NET tasks carry no seniority tier, so entry-level exposure cannot be separated from senior exposure. There is no API traffic and no enterprise use case field, so questions about how firms deploy these systems fall outside the feed.

The framing caveat is the one that will determine whether an analysis holds up. The accurate reading is that AI is being used for tasks commonly performed by a given occupation. The data does not establish that people in that occupation are the ones doing it. Someone asking about lesson planning may not be a teacher.

And this is Claude usage. Extending any of it to AI adoption broadly means extrapolating from one vendor's conversation sample.

The useful conclusion from a session with this data is narrower than the headlines it will generate. Diffusion is heavily concentrated in a small set of knowledge-work categories, the international divide runs primarily through education rather than employment, and the work most exposed to full substitution is migrating to interfaces this measurement does not cover. Reading the next release against the underlying dataset rather than the conversational layer is where the more interesting analysis will come from.