Source: ai-research/claude-blog-how-people-use-cowork.md — Anthropic first-party blog post, claude.com/blog/how-people-are-using-claude-cowork.
Anthropic sampled 1.2 million anonymized Claude Cowork sessions from May 11-31, 2026, across more than 600,000 organizations, and classified them into a 20-category taxonomy of work. The headline finding: software development is a small share of Cowork usage (8.7%), while business process/operations (33.4%) and content creation/copywriting (16.4%) together account for roughly half. Anthropic’s framing is that Cowork carries “the work around the work” — the connective, cross-role tasks that keep projects moving but rarely appear in anyone’s job title. This is the wiki’s first article built on first-party Cowork usage-share data rather than feature announcements or tutorials, and it’s the data source cited directly by web-beta launch post.
Key Takeaways
- Business process/operations is the single largest category at 33.4% — pulling scattered updates into one report, building onboarding checklists, reconciling spreadsheets. Anthropic attributes the size to how many different roles (finance, HR, admin) touch this kind of task.
- Content creation/copywriting is second at 16.4% — drafts, slide decks, posts, proposals. Framed as solving the “blank page” problem; also spans multiple roles (marketing, comms, biz dev, project management).
- Software development is 8.7% of Cowork sessions — deliberately small by design, per Anthropic’s framing: developers reach for Claude Code to write code, and use Cowork instead for the connective, communications-focused work that surrounds their role (the same category most other roles use it for).
- The full named breakdown (8 of 20 taxonomy categories are given a number): business process/operations 33.4%, content creation/copywriting 16.4%, software development 8.7%, DevOps/infrastructure 7%, research/intelligence 6.4%, data analysis/business intelligence 5.8%, document processing/extraction 4.1%, sales/revenue operations 4%, personal assistance 3.8%, education 2.4%, meeting intelligence 1.8%. The remaining ~9 categories in the 20-category taxonomy are unnamed in the post (see Open Questions).
- Anthropic’s interpretive thesis: Cowork usage is “connective.” Spreadsheets pull disparate data into a comparable context; decks convey a decision across audiences with different context levels; onboarding checklists transfer institutional knowledge. The worked examples given are a lawyer using Cowork for document formatting/filing (freeing time for legal judgment), a hiring manager using it to schedule meetings and synthesize interview feedback (freeing time for candidate conversations), and a team lead using it to build the slide deck that explains a decision (freeing time to actually make the call).
- Methodology is sampling-based, not a full-traffic count. Sessions are captured at a capped rate (a fixed max sessions/hour), not a fixed percentage of traffic — so every percentage in the report is a share of the sample, not a share of absolute Cowork volume. Classification ran through an automated system; a privacy-preserving analysis tool kept all session content anonymous, and no individual session was read by a human analyst.
- Cowork launched in January (2026, per the post’s own phrase “since we launched Claude Cowork in January”), extending Claude Code’s agentic capabilities into the same chat interface non-technical users already used — the post explicitly recalls that non-technical users started using Claude Code itself in unexpected ways after its 2025 release (organizing folders, deduplicating files, writing spreadsheet formulas), while others found the terminal “a literal black box.” Cowork is presented as the answer to that second group.
- Anthropic frames this as the first of a recurring series — “we plan to continue publishing data as usage grows and shifts, and we’ll report on what changes over time.”
The 11 named categories (of 20 in the full taxonomy)
| Category | Share | Anthropic’s examples |
|---|---|---|
| Business process & operations | 33.4% | Consolidating scattered updates into a report, onboarding checklists, reconciling spreadsheets |
| Content creation & copywriting | 16.4% | Drafts, slide decks, posts, proposals |
| Software development | 8.7% | (not detailed — contrasted against Claude Code) |
| DevOps & infrastructure | 7% | (not detailed) |
| Research & intelligence | 6.4% | (not detailed) |
| Data analysis & business intelligence | 5.8% | (not detailed) |
| Document processing & extraction | 4.1% | (not detailed) |
| Sales & revenue operations | 4% | (not detailed) |
| Personal assistance | 3.8% | (not detailed) |
| Education | 2.4% | (not detailed) |
| Meeting intelligence | 1.8% | (not detailed) |
Business process/operations and content creation/copywriting together are described as “roughly half of all usage” (33.4 + 16.4 = 49.8%, consistent with the post’s own rounding).
Why it matters for a marketing agency
This is direct, dated evidence for where Cowork’s usage-weight actually sits, and it corroborates two things the wiki already argued from tutorials and case studies rather than aggregate data:
- Cowork for Marketing’s eight use cases (strategy decks, campaign performance analysis, competitor reports) sit squarely inside the two largest measured categories here (business ops + content creation) — the tutorial’s use-case selection wasn’t arbitrary, it tracks where actual usage concentrates.
- Anthropic’s own marketing-ops case study (weekly metrics report, event-build pipeline) is a worked example specifically inside the “business process & operations” 33.4% bucket — reporting consolidation and event/campaign setup are exactly the kind of “pulling scattered updates into a single report” and multi-system reconciliation this category describes.
For an agency deciding where to invest Cowork automation effort first, this data argues for reporting/consolidation and content-drafting workflows over anything code-adjacent — which is also where an agency’s actual day-to-day time goes, unlike a software team.
Try It
- Audit your own Cowork usage against this taxonomy. If most sessions are landing in categories outside the top two (business ops, content creation), that’s either a sign of an unusual workflow or a sign you haven’t yet found the highest-leverage use case for your team.
- Prioritize automation-building effort by category size, not by novelty. A reporting-consolidation workflow (33.4% territory) is statistically more likely to generalize across a team than a bespoke research workflow (6.4% territory) — though both are legitimate, size the investment accordingly.
- Read this alongside the marketing-ops case study for a concrete worked example inside the largest measured category, and Cowork for Marketing for the broader use-case menu.
- Watch for the promised follow-up reports. Anthropic states it will keep publishing this data as usage shifts — worth adding to the wiki’s watchlist as a recurring source.
Related
- Claude Cowork (Product Overview) — the product this usage data describes
- Getting Started with Claude Cowork — the onboarding walkthrough; its mobile/web-beta update cites this usage report directly as justification for expanding beyond desktop
- How Anthropic’s Own Marketing Operations Team Uses Claude Cowork — a worked example inside the largest measured category here
- Claude Cowork for Marketing — use-case tutorial whose focus areas this data corroborates
- Claude AI — topic landing, including Claude Code for the explicit code-vs-Cowork usage contrast this post draws
- AI Industry Research — home for other first-party/independent usage and benchmark studies this report is methodologically adjacent to
Open Questions
- 9 of the 20 taxonomy categories are unnamed. The post explicitly lists only 11 (8 named with a percentage, plus the 3 “all other” examples): business process/operations, content creation/copywriting, software development, DevOps/infrastructure, research/intelligence, data analysis/business intelligence, document processing/extraction, sales/revenue operations, personal assistance, education, meeting intelligence. The remaining ~9 categories and their shares are not given.
- The “Limitations” section did not survive extraction. The live post has a “Limitations” header under “Additional details about our research” but the body content wasn’t captured by the fetch tool (only an invisible placeholder character came through). Given the methodology section’s own caveat about capped-rate sampling, the limitations likely address sampling bias and/or taxonomy coarseness, but that is inference, not extracted text — re-fetch with a rendering-capable tool if the exact stated limitations matter.
- No comparison to a prior period. This is presented as a first snapshot; there’s no month-over-month or launch-to-May trend data yet, despite Anthropic’s own framing that “its uses are evolving quickly.”