ReMyndwhat people use ai for at work, by the numbers
most writing about ai at work is vibes. anthropic publishes an actual dataset, so this is a read of it.
the anthropic economic index matches the content of sampled claude conversations against task descriptions from o*net, the us department of labor's catalog of what each occupation actually does. the numbers below are from the may 2026 period, published under cc by 4.0, covering 121 countries and 718 occupations with published usage.
as of september 2026 this is the most recent release, snapshotted on june 24, 2026.
one caveat up front, because it changes how you should read everything else. this measures conversations matched to tasks. it does not measure who the users are, and it says nothing about the labor market.
it is also worth saying what makes these ai usage statistics 2026 unusual. most numbers you see quoted come from surveys, where people report what they think they do. this one reads the conversations.
the top task is searching, and it is not close
the single most common work task in sampled conversations worldwide is "search electronic sources, such as databases or repositories, or manual sources for information," at 4.95%.
second is "search standard reference materials, including online sources and the internet, to answer patrons' reference questions," at 3.74%.
put those together and the two highest ranked work tasks on the list, 8.69% of sampled conversations between them, are both retrieval. not generation. not reasoning. finding the thing.
| rank | work task | share |
|---|---|---|
| 1 | search electronic sources for information | 4.95% |
| 2 | search reference materials to answer questions | 3.74% |
| 3 | recommend and advise on products and services | 2.25% |
| 4 | answer inquiries about software or hardware | 1.98% |
| 5 | write or modify programs for customer requirements | 1.47% |
that ordering is the most useful thing in the dataset, and it is almost the opposite of how these products are marketed. the ads are about creation. the most common ai tasks are lookup.
it holds across vendors too. openai, google and microsoft all launched memory features in 2026 on the same premise, that the assistant should stop asking you to repeat context, which is a retrieval framing rather than a generation one. openai shipped its background memory system on june 4, 2026 and anthropic reworked claude's memory into editable entries on july 10, 2026.
what people are actually asking for
request topics tell the same story from a different angle.
content creation and copywriting leads at 22.72%, then education and learning at 13.23%, software development at 11.51%, and research and intelligence at 10.94%. those four are 58.4% of everything.
then look at the bottom of the list, because that is where the interesting gap is.
- knowledge retrieval and enterprise search: 3.61%. people want to search their own organization's knowledge, and mostly cannot.
- personal ai assistant: 2.86%. the thing every keynote promises is 1 in 35 conversations.
- conversation and meeting intelligence: 0.26%. essentially nothing, in a working world that runs on calls.
the last one is not a statement about demand. meetings are where decisions get made. the share is near zero because the data was never available to ask about, which is a supply problem dressed up as a preference.
augmentation vs automation is roughly a coin flip
51.38% of conversations are classed as augmentation and 48.62% as automation.
augmentation means the person stays actively involved in the task. automation means they direct the model to complete it. these are conversation styles rather than job outcomes, and the split has been the most misread number in the whole index.
the practical implication is about session shape. more than half of this usage is long, iterative back and forth where a person and a model work a problem together. that is not a query. it is a working session, and it produces exactly the kind of context that vanishes when the tab closes, which is the subject of where your ai conversations go.
the work, personal and coursework split
of classified conversations, 43.36% look like work, 40.2% look like personal life, and 16.45% look like coursework.
personal use being nearly as large as work use is the number most companies get wrong when they plan deployments. and coursework at 16.45% is a sixth of everything, which is worth sitting with given how much of the public argument treats student use as a fringe case.
by job category, conversations matched to computer and mathematical tasks lead at 23.8%. arts, design, entertainment, sports and media follows at 13.55%, then educational instruction and library at 12.79%, sales at 9.14% and office and administrative support at 7.89%.
the phrasing matters here. these are conversations matched to tasks commonly done by those occupations. many of the people in them do not hold those jobs.
how much time does it actually save?
the index carries an estimate, and it is worth quoting with its caveats intact.
tasks that a classifier estimates would take a person working alone roughly 5 hours ran to roughly 40 minutes of conversation. those are automated estimates from conversation content, bucketed by order of magnitude, not measured times. treat it as an order of magnitude claim rather than a stopwatch.
what the estimate does not include is the time spent finding the inputs. reconstructing what you already knew, hunting the document you had open last week, rewriting the context you gave a model on tuesday because thursday's session has never heard of it.
the gap this data exposes
read the top of the list and the bottom of the list together.
the number one thing people do with ai is search for information. the smallest categories are searching their own organization's knowledge and their own meetings. so the demand is overwhelmingly retrieval, and the supply of retrievable personal and organizational context is close to zero.
every one of those 4.95% searches runs against a corpus someone else assembled. none of them can search what you personally read, wrote, watched or decided. that is not a model capability problem, and a larger context window does not fix it. the record simply does not exist in a queryable form.
this is the gap remynd exists to close on a mac. it captures the focused window, runs ocr locally so what you looked at becomes searchable text, and keeps the archive on your machine. capture, ocr and storage stay local by default and you can exclude apps or sites entirely, though sign in and the cloud agent do reach the network, which we set out on the security page.
once that archive exists, the number one task in this dataset finally has somewhere useful to point. more on how that search works in practice in how to find something you saw on your screen, and on wiring it into agents in how to give an ai agent context.
how to read these numbers responsibly
four rules, because this dataset gets misquoted constantly.
- it is usage, not employment. the department of labor's o*net catalog supplies the task descriptions. the match is content to task, not user to job.
- it is a snapshot, not a trend. the published span runs april to may 2026. nothing here can show a share rising or falling.
- null is not zero. unpublished figures mean privacy minimums were not met, not that usage was absent.
- it is one product. claude's user mix is not the world's, and the software development share almost certainly reflects that.
with those held firmly, it is still the best public ai at work data anyone has released, and the best public picture of what people use ai for.
download remynd for mac and give the searching somewhere to land.