Chapter 2 · Tasks and jobs

What people bring to AI is not who they are.

Anthropic classifies observed Claude conversations into occupational task groups. That shows activity on one platform. It does not show users' actual jobs, and it does not measure what AI does to employment.

The verified version of the thesis. In Anthropic's sample of U.S. Claude.ai conversations (May 2026), 21.1% were classified to computer and mathematical tasks, the largest of 22 groups. This describes activity on one platform, not users' occupations or the effects of all AI.

U.S. Claude.ai conversations by occupational task group

Bars: May 2026. Grey: April 2026.

Observed

Computer and mathematical tasks lead at 21.13%, followed by arts, design and media (12.91%) and education (11.92%). Construction (0.14%) and farming (0.05%) are at the bottom.

Computer and Mathematical21.1%
Arts, Design, Entertainment, Sports, and Media12.9%
Educational Instruction and Library11.9%
Sales and Related11.6%
Office and Administrative Support7.6%
Business and Financial Operations6.3%
Management5.8%
Healthcare Practitioners and Technical4.3%
Life, Physical, and Social Science4.2%
Community and Social Service3.2%
Architecture and Engineering2.9%
Personal Care and Service1.3%
Production1.1%
Legal1.0%
Installation, Maintenance, and Repair0.8%
Healthcare Support0.7%
Food Preparation and Serving Related0.6%
Protective Service0.4%
Transportation and Material Moving0.4%
Building and Grounds Cleaning and Maintenance0.2%
Construction and Extraction0.1%
Farming, Fishing, and Forestry0.1%
Current April 2026

Caveat. Shares describe what sampled conversations were about, mapped to O*NET tasks and SOC groups. They are not users' jobs, not all AI use, and not economic effects. Shares sum to 98.67% because small cells are suppressed.

UnitPercent of sampled conversationsGeographyUnited StatesObservedApr and May 2026 (monthly)SourceAnthropic Economic Index (2026-06-26 release)RetrievedRelease dated 2026-06-26 (file supplied, SHA-256 9be20dd4…)
Inspect the evidence
What this chart does not show
  • Users' actual occupations or industries.
  • AI used inside other software, machines or industrial systems.
  • Whether the task was completed faster or better.
  • Other AI products (ChatGPT, Gemini, Copilot).
Lineage: source to transformation to chart
  1. aei_claude_ai_2026-06-26.csv
  2. Filter geo_id=USA, category_name=soc_occupation, hierarchy_level=1, metric_id=pct
  3. No reweighting

Text description. Computer and mathematical tasks lead at 21.13%, followed by arts, design and media (12.91%) and education (11.92%). Construction (0.14%) and farming (0.05%) are at the bottom.

Link

Developer API traffic by occupational task group (global)

Observed

API traffic leans further to computer and mathematical tasks (28.17%).

Computer and Mathematical28.2%
Arts, Design, Entertainment, Sports, and Media12.6%
Office and Administrative Support11.9%
Management6.3%
Educational Instruction and Library6.2%
Business and Financial Operations4.6%
Life, Physical, and Social Science4.3%
Architecture and Engineering4.0%
Sales and Related3.7%
Personal Care and Service1.9%

Caveat. 1P API excludes Claude Code and third-party clouds. Global only; no U.S. split published.

UnitPercent of sampled API trafficGeographyGlobalObservedMay 2026SourceAnthropic Economic Index (2026-06-26 release)RetrievedRelease 2026-06-26
Inspect the evidence

Text description. API traffic leans further to computer and mathematical tasks (28.17%).

Link

What U.S. Claude.ai conversations produced

Observed

Explanations and documents lead; code fixes are a small share of consumer chat.

Explanation or answer17.9%
Advice or recommendation15.1%
Document or report13.9%
Analysis or summary6.0%
Email or message5.9%
Plan or strategy4.4%
App or website3.3%
Data or spreadsheet2.8%
Marketing or social content2.4%
Script or snippet2.2%
Code fix or debug2.1%
Resume or job application2.0%

Caveat. An artifact is the most prominent output, classified by a model. Producing a document does not mean it was used.

UnitPercent of conversations, by primary outputGeographyUnited StatesObservedMay 2026SourceAnthropic Economic Index (2026-06-26 release)RetrievedRelease 2026-06-26
Inspect the evidence

Text description. Explanations and documents lead; code fixes are a small share of consumer chat.

Link

Work

41.3%

of U.S. conversations classified as work

Personal

49.7%

classified as personal

Automation

49.6%

directive or feedback-loop pattern (AI does the task)

Augmentation

50.4%

learning, iteration or validation (human leads)

U.S., Claude.ai, May 2026. Self-reported productivity in Anthropic's linked survey (~9,700 respondents, not representative): faster 86%, broader scope 82%, higher quality 69%. Self-reports are kept separate from measured outcomes.

Workers

Occupations behind the headlines, May 2025.

Employment and pay from the BLS occupational survey. This is a level, not a trend: the BLS API serves only the latest year, and the historical files refused this client (logged in the gaps register). So this page does not claim that any occupation grew or shrank because of AI.

Radiologists are the canonical test. Since 2016 they have been cited as the job AI would replace first. In May 2025 BLS counted 26,770 radiologists earning a mean $381,530 a year. The trend that would settle the question is the next acquisition.

BLS Occupational Employment and Wage Statistics (OEWS), national, all industries. Reference May 2025. Retrieved 2026-11-18. Only the latest year is served by the BLS API. Earlier years (for trends) need the OEWS flat files, which returned HTTP 403 to this client.
OccupationSOCEmploymentMean annual wageMedian annual wage
Home health and personal care aides31-11204,305,810$36,120$35,800
Registered nurses29-11413,379,720$101,420$97,550
Customer service representatives43-40512,595,750$46,590$44,770
Heavy and tractor-trailer truck drivers53-30322,062,040$59,710$58,640
Software developers15-12521,687,890$148,100$135,980
Accountants and auditors13-20111,449,500$94,750$83,680
Elementary school teachers25-20211,388,390$72,650$63,970
Secondary school teachers25-20311,065,210$76,320$72,040
Electricians47-2111757,220$71,490$63,190
Plumbers, pipefitters, steamfitters47-2152465,840$72,170$63,800
HVAC mechanics and installers49-9021409,670$64,780$61,010
Paralegals and legal assistants23-2011392,880$69,700$62,890
Physicians, all other29-1229342,720$262,040$265,930
Nurse practitioners29-1171323,040$137,300$132,300
Network and computer systems administrators15-1244314,340$103,680$99,130
Machinists51-4041287,050$59,320$58,750
Data scientists15-2051262,440$126,800$120,230
Radiologic technologists and technicians29-2034230,490$83,840$80,110
Graphic designers27-1024197,830$70,560$62,960
Medical records specialists29-2072194,720$56,790$51,140
Family medicine physicians29-1215107,510$255,820$244,180
Interpreters and translators27-309152,060$66,360$60,170
Technical writers27-304245,500$96,970$90,390
Radiologists29-122426,770$381,530$420,860

Task exposure is not job loss. An occupation can have many AI-suitable tasks and still grow.

Observed task use is not automation. Most U.S. conversations are split roughly evenly between automation-style and augmentation-style patterns.

A new task is not a new job. Roles are called new here only with employment evidence that persists over time.