Anthropic's Cowork data shows AI agents have left the developer terminal

Usage from 1.2 million anonymized Claude Cowork sessions across 600,000 organizations shows more than 90% of the work isn't code — as Anthropic pushes the agent onto mobile and web.

Anthropic pushed Claude Cowork onto mobile and web on Tuesday and, in the same breath, released a set of usage numbers that quietly reframe what the product is for. Across 1.2 million anonymized sessions sampled from May 11 to May 31, drawn from more than 600,000 organizations, over 90% of Cowork activity had nothing to do with software development. Business process operations, the catch-all for reconciling spreadsheets, building onboarding checklists, and pulling scattered updates into a single report, accounted for 33.4% of sessions. Content creation and copywriting made up 16.4%. Coding, the use case that dominates the marketing of every frontier lab, came in at 8.7%.

“While coding is still — understandably — one of the uses of AI that gets the most attention, the use of AI for everyday business work is on the rise,” Anthropic told TechCrunch. The framing is careful. It’s also strategically loaded.

Cowork launched in January 2026 as a research preview for Max subscribers on the $100 and $200 monthly tiers, pitched by Fortune as “Claude Code for the rest of your work” and, per the same reporting, built in roughly a week and a half using Claude Code itself. Six months in, the buyer profile has drifted decisively toward finance, HR, and administration. The mobile release closes the loop: NBC News reports users can now start a task at a desk, monitor it from a phone, and get prompted only when Claude hits a decision the human actually has to make. Usage limits are doubled through August 5.

The competitive geometry is what makes this interesting. Anthropic is now fighting on the same turf as Microsoft Copilot, Google, and OpenAI, alongside a growing specialist tier including Glean, Dust, Gumloop, and LemonLime, which has built its practice around exactly this non-developer office work. Sonnet 5, released last week at $2 per million input tokens through August 31, sits underneath it all. Fortune notes Anthropic has acknowledged Cowork remains vulnerable to prompt injection despite its defenses.

The company that made its name selling agents to engineers just published data showing engineers are a rounding error in its own workflow product. That’s the story the release notes are telling on themselves.

Anthropic's Cowork data shows AI agents have left the developer terminal

Sources