Anthropic analyzed 1.2 million Claude Cowork sessions from over 600,000 organizations (May 2026). Half of all usage falls into two categories: administrative processes (33.4%) and content creation (16.4%). That means: filling in reports from scattered updates, building onboarding checklists, reconciling spreadsheets, drafting presentations and documents.

Coding accounts for only 8.7% of sessions: developers use Claude Code to write code, and Cowork for the communication around it. Other categories (research 6.4%, data analysis 5.8%, personal assistance 3.8%) remain marginal.

Why this matters to you. The thesis of this site’s manifesto holds up in the data: AI strengthens where work is repetitive and ungratifying—the “work around the work.” It doesn’t replace the creative or strategic part, but frees up time for it. If you’re deciding what to delegate, these numbers tell you where AI delivers recognizable results right away: the connective and formatting tasks everyone has to do and nobody enjoys.

In detail

The context

Claude Cowork is Anthropic’s chat agent launched in 2025, designed to bring agentic capabilities into a familiar interface. Before this analysis, the assumption was that coding would be the main use case: instead, the data shows developers go to Claude Code for programming and use Cowork for everything else.

What the analysis actually says

The sample covers May 11–31, 2026. An automated system assigned each session to one of 20 work categories. The key numbers:

• 33.4% operational and administrative processes: status reports, onboarding, reconciliations.
• 16.4% content creation: drafts, slides, posts, proposals.
• 8.7% software development, 7% DevOps, 6.4% research, 5.8% data analysis.
• The rest distribute across minor categories (sales 4%, meetings 1.8%).

Anthropaic describes these activities as “connective”: they link scattered information, translate decisions for different audiences, give newcomers access to institutional knowledge.

Stated limitations

Anthropaic itself flags:

• The taxonomy classifies activities, not professions: there are no separate categories for marketing, finance, or HR—these likely sit under “Operational processes.”
• The sample has a ceiling on sessions per hour, so peak hours are underrepresented.
• About 5% of sessions are personal use (hobbies, chat): the dataset isn’t purely work.

How this connects to what we’ve covered

On July 6, Zuckerberg acknowledged that agents move slower than expected: building reliable agents for complex tasks is hard. These Anthropic numbers confirm it from the other side—agents work well today where the task is clear and repetitive, not where judgment or strategic creativity matters.

The lessons from the course on writing clear instructions start from this reality: an agent works well where you can describe “how to recognize a good result.” Vague or creative tasks remain human territory.

Practical takeaways

If you’re assessing what to delegate to an agent, start with the connective tasks in your day: recurring reports, summaries of long threads, first drafts for review. The data says that’s where AI delivers measurable value. Don’t hand it strategy or decisions: hand it the prep work that makes them possible.

Anthropaic will publish updates to this analysis. For now, these numbers confirm the right angle is augmentation, not replacement.