Anthropic has been reading millions of AI conversations. Here's what they found.
The Economic Index is the best data we have on how AI is actually changing work. Most people aren't looking at it.
Most of what you read about AI and jobs is made up. Predictions from consultants, survey data from people who want your email, confident takes from people selling something.
Anthropic has been quietly doing something different. For over a year they have been publishing the Anthropic Economic Index, a research programme that looks at millions of anonymised Claude conversations and maps them against the US Department of Labor’s occupational taxonomy. In plain English, they are measuring what people are actually using AI for, broken down by job and task.
It is the most grounded picture of AI at work that exists in public. And it is telling a story that does not match the headlines.
Let me walk you through what it shows.
What the Economic Index actually is
The premise is simple. Every conversation on Claude.ai gets classified, automatically and anonymously, against O*NET, the huge taxonomy the US Department of Labor uses to describe what jobs involve. Software developer, paralegal, graphic designer, radiologist, and thousands more. Each occupation has a list of specific tasks, and each conversation gets tagged with whichever tasks it matches.
Stack all those tags together and you get a map. A view of which kinds of work are being done with AI, how much, and in what way.
Anthropic has released three reports since February 2025, each one sharper than the last. The latest extends the lens to API usage from developers and enterprises, not just consumer Claude. Together they form the closest thing we have to a real-time dashboard for AI’s footprint in the economy.
The jobs using AI, and the ones not
The first surprise is how concentrated it is.
Roughly 37% of all Claude usage maps to computer and mathematical occupations. That is software engineers, data analysts, and the people who write code for a living. Another 10% is arts, design, and writing. Education, business operations, and research make up most of what is left.
What is not there is striking. Physical work barely registers. Healthcare delivery, agriculture, construction, transport, cleaning, retail, hospitality, manufacturing. These sectors make up most of the economy and most of employment, and they show up as a rounding error in the data.
There is also a pattern inside the white-collar segment. Mid-to-high wage jobs dominate usage. Very high wage jobs, CEOs, partners, senior physicians, use AI far less than you would expect, probably because their time is spent in meetings, decisions, and relationships that do not translate well to a chat window yet.
So the hype version of “AI is coming for your job” is roughly 180 degrees wrong for most readers. The people using AI most heavily right now are the ones most likely to use it to do their job better, not the ones whose jobs are being replaced. The people not using it are either insulated by the nature of their work, or they have not yet encountered a workflow it fits.
The ratio that keeps shifting
The most important number in the entire Economic Index is not about which jobs. It is about how people use the tool.
Anthropic classifies every conversation on a spectrum from augmentation to automation:
- Augmentation means a human is iterating with the AI. Drafting, revising, brainstorming, stuck on a problem and working it out together. The human is still in the loop on the output.
- Automation means handing the full task off. “Do this thing, come back with the answer.”
In the first report in early 2025, usage leaned toward augmentation, roughly 57% to 43%. By the latest report the ratio has flipped. More than half of all tasks are now classified as automation, and the shift is even starker on the API side, where developers are stringing together full workflows rather than chatting.
This is the shift worth paying attention to. It is the single best signal of where AI is actually going. Not “an AI assistant that helps you draft an email” but “an AI that runs the whole email pipeline and only pings you if something breaks.” The direction has been one-way since the index started, and the gradient keeps steepening.
Where in the world this is happening
The third report added a geographic lens, and it is uncomfortable.
AI use is not evenly distributed. It is concentrated in a small number of places. Washington DC, California, Massachusetts, and a handful of other US states have per-capita Claude usage multiples higher than the national average. Internationally the story is similar. Ireland, Singapore, Israel, and the UK punch well above their weight. Large parts of the world barely show up.
The uncomfortable bit is that this mirrors other forms of economic advantage. The places that were already rich in knowledge work, venture capital, and research infrastructure are also the places adopting AI fastest. If AI turns out to be a productivity multiplier on knowledge work, and the data so far suggests it is, the gap between AI-dense regions and the rest of the world looks likely to widen, not narrow.
This is not a prediction. It is the current trendline plotted forward a few years.
What the index does not show
A few honest caveats before drawing conclusions.
The Economic Index only sees Claude. It does not see ChatGPT, Gemini, Copilot, Llama, or the dozens of smaller models people use at work. Claude has a distinctive user base, heavier on developers and knowledge workers than most, which probably inflates the software engineering share and suppresses the marketing and sales share relative to the true industry-wide picture.
It also cannot see intent. A conversation tagged as “legal research” might be a lawyer doing their job or a student learning the field. The index is good at describing what AI is being asked to do. It is less good at describing who is asking.
And it lags reality. These are conversations that have already happened. A report published this month is reflecting usage from a few months ago, which in AI terms is a long time.
With those caveats logged, there is no better source of ground truth on this question in public. Anthropic is effectively running the statistics office for a corner of the economy that did not exist three years ago.
The pattern across all three reports
Read them together and a clear shape emerges.
AI adoption is real but narrow. It is deepening fast inside knowledge work and barely touching the rest of the economy. The nature of use is shifting from assistance to delegation. The benefits are accruing to specific geographies and specific job categories, not evenly.
The mental model “AI is a new technology, like the internet, that will gradually spread everywhere” is probably wrong, or at least incomplete. The Economic Index suggests something more like “AI is a new kind of employee who is absolutely everywhere in some industries and entirely absent from others.” That has very different implications for policy, for careers, and for where the economic surplus ends up.
The question nobody has a good answer to yet is whether this stays concentrated, or whether it eventually spreads. The optimists argue the current narrow footprint is just early-adopter dynamics and will broaden as tools mature. The pessimists argue the footprint mirrors structural features of which work is digitisable and will stay narrow for a long time.
The Economic Index does not settle that argument. It gives both sides better data to argue with.
What any of this means for you
Bringing it back to ground.
If you work in a sector that barely shows up in the data: You probably have more time than people are telling you. Physical work, regulated professions, hands-on roles. AI is not sweeping through these today and will not next quarter either. But the productivity gap between your sector and the AI-heavy sectors is widening, which affects relative wages and investment over time. Worth noticing, not worth panicking over.
If you work in a sector that is right in the middle of the data: Software, writing, design, analysis, ops. The question is not whether AI changes your work. That is already happening. The question is which side of the augmentation / automation flip you end up on. People building workflows and pipelines are doing fine. People still chatting with a model one turn at a time are about to be outcompeted by people who are not.
If you run a business: The people on your team using AI heavily and the people who barely touch it are doing work of very different quality per hour. The gap between them is bigger than any single-tool decision you are making. Worth measuring before you worry about which vendor to buy.
If you are watching from the sidelines: The Economic Index is one of the most important pieces of ongoing research in AI right now, and almost nobody is reading it. The next couple of reports, particularly if the automation ratio keeps climbing, will tell us whether we are in a gradual transition or a fast one.
The value of the index is that it replaces opinion with measurement. Most AI coverage is someone’s vibe about the future. This is a record of the present. That alone makes it worth an hour of your attention every time they ship a new one.
See you next week.
Stephen
Sources and further reading
- Anthropic, Economic Index landing page
- Anthropic, first Economic Index report (Feb 2025)
- Anthropic, second Economic Index report
- O*NET, occupational taxonomy