Before you start
How to read a dot
Four real tasks, drawn big. Hover or tap one — you get the same card every dot on this page gives you.
A task’s label comes from O*NET, the Labor Department’s catalog of what jobs consist of. It names the work, not the person asking: tasks filed under “Healthcare Practitioners” are 88% personal use — patients, not doctors.
The field
All 2,713 tasks, seven ways
The same 2,713 dots re-form for each question. Search them, hovertap them, or jump between lenses in any order.
01
2,713 kinds of help
Every task Claude was seen doing in May, sized by how much of the month it took up. Most are specks. A handful are enormous.
The size scale is the same in every lens of this field, so the picture stays comparable as it re-forms.
02
The biggest job is looking things up
Not writing code. The two largest tasks are both searching for information, and a third retrieval task sits just below the biggest coding task. Together the three are 10.1% of conversations; the largest coding task ranks fifth, at 1.5%.
This describes Claude chat. Claude Code is in no published file, so it says nothing about how much code the model writes overall.
03
Only 43 in 100 are work
Forty are personal — health, money, dinner, feelings. Sixteen are school. Each task sits in the disc it leans toward, and each disc is sized by the share of conversations its own tasks carry.
Counted by conversation, never by person — one heavy user contributes many. The 43/40/16 split counts every conversation; the discs count the tasks, and work-leaning tasks are the bigger ones, so the work disc is larger than 43% alone.
04
Half do it for me, half do it with me
Every task is placed by how often it gets handed over whole rather than worked through together. Mentoring and tutoring stay shared. System monitoring gets handed over. Overall: 48.6 / 51.4.
That split has not left the 45–51% band in ten months — the one measure in this data that can honestly be drawn across time. It is below.
05
Ten tasks carry a fifth of it
The bright core is the 1,147 tasks that carry every measurable share. The pale ring is the 1,566 that are real but publish as 0.00% — present in the count, too small to weigh. Ringed in the middle: the ten tasks that carry 22% of all the published share.
06
In 87.6% of conversations, a person could have done it alone
Almost every task piles up at the right-hand end. People are not mostly asking for things they couldn’t do. They are asking for time.
This is a classifier’s judgement of each conversation, not a test of what anyone can actually do.
07
About five hours of work, handed back in forty minutes
Across all conversations the estimate is 4.73 hours alone against 40.12 minutes with help — roughly seven times less. Each task sits at its own ratio.
Read the units carefully: the published columns are hours for one and minutes for the other. Compared raw, they appear to say AI makes work slower. Both are estimates a classifier assigned, not stopwatch times.
An empty space is not proof of absence. Small cells are withheld for privacy, so a missing task means below the publication threshold — not “never happens.”
What comes back
The most common thing it hands back is an answer
500 dots, one for every 0.2% of conversations, sorted into five families of output (my grouping of the 32 labels the release publishes). HoverTap any dot.
1 dot = 0.2% of conversations · 500 dots = 100%
Explanations, documents and advice — the oldest kinds of knowledge work. Every code-shaped output combined (apps, fixes, scripts, configs, queries, ML systems) is 12.9% of Claude chat, less than plain explanations alone.
Two machines
The chat product and the API are different species
Same month, same classifier, two doors into the same model. HoverTap a row for what it means.
Each row: the share of that source’s conversations, 0–100%
Software wired to the API is given tasks to finish. People in the chat window work alongside it. The API side here is first-party API traffic with Claude Code excluded, so none of it is a picture of coding agents. There is no combined number either: the release publishes no volumes, so the two populations cannot honestly be averaged — an average of a 94% and a 49% population describes nobody.
Against the economy
The work people bring looks nothing like the workforce
Each occupation family’s share of U.S. conversations against its share of U.S. jobs. 1× would mean proportional.
Share of U.S. Claude conversations ÷ share of U.S. jobs. Left of the line = less Claude than jobs. Right = more.
Three orders of magnitude separate the ends. Each label names the work a task belongs to, not who typed it: tasks filed under “Healthcare Practitioners” are 88% personal use — patients, not clinicians. And this compares a conversation mix with a headcount mix — two different universes, with the jobs baseline from May 2023 against Claude data from May 2026. It measures what work gets brought to the machine, not who is being replaced.
121 countries
Where it is a school tool
Every country with published data. Across → how much Claude gets used per working-age person. Up → how much of that country’s own use is schoolwork.
Schoolwork runs from Tunisia at 51.5% down to Japan at 5.9%, with the United States at 8.9%. These are two separate facts and the chart keeps them apart: coursework share is measured inside each country’s own conversations, so Tunisia’s figure is a statement about its mix, not its volume — it cannot be read as Tunisian students using Claude more than American students. The usage index runs about 90× from Australia (6.40) down to Tanzania (0.07).
Ten months
The one line that survives
The share of conversations handed over whole, in all five published windows. HoverTap a point.
The five points are not evenly spaced in time and are not the same length: the first three are single weeks, the last two are whole calendar months. Everything else in this data breaks at the June 2026 release, which rewrote the classifier, the schema and the window length: individual task shares reshuffle so hard across that boundary that no trend may cross it. This measure does survive it — never leaving 45–51% — which is what makes it the only series here worth drawing. It is a statement about the published measure, not proof that behaviour held still.
Coverage
The catalog of American work has 18,796 tasks.
Claude was seen doing 2,713
Every circle is one task statement in O*NET. HoverTap a lit one.
- 2,713 observedTasks Claude was seen doing in May — 14% of the catalog. Bigger circle, bigger share of the month.
- 16,083 not publishedNo row for them in this release. That folds together work below the publication threshold and work nobody asked for.
- The 10 in the middle carry 22%Ten tasks account for 22 of the 94 published percentage points. They sit at the centre, inside the ring.
The 2,713 observed tasks are ordered by size from the centre out; the rest are one small speck each, so the whole catalog fits in a single disc.
Machine work is a thin, deep slice of that catalog rather than a broad one. The unlit dots fold two things together that this file cannot separate: work below the publication threshold, and work nobody asked for.
The full register
All 2,713 tasks, one ledger
Search it, sort it, open any row for its full profile.
month’s conversationsHanded over
wholePersonal
use
Every column but the first is a share of that task’s own conversations. † observed, but below the size the release will put a number on — not zero.