Anthropic's first party analysis of 1.2 million Claude Cowork AI agent sessions shows business operations at 33.4%, ahead of content creation and software development combined.
Business process and operations work accounts for 33.4% of Anthropic's Claude Cowork AI agent sessions, the largest share of any task category in the company's first-party analysis of 1.2 million anonymized sessions between May 11 and May 31, 2026. Content creation and copywriting is second at 16.4%. Software development comes in third at 8.7%, followed by DevOps and infrastructure at 7%, research and intelligence at 6.4%, and data analysis and business intelligence at 5.8% (Anthropic).
The composition inverts the developer-first story that has dominated enterprise AI agent coverage since the category took shape. Cowork is Anthropic's AI agent product that automates multi-step workplace tasks, and the company's own framing is that operations work, not coding, is the dominant use case. Anthropic tied the data release to a separate product move: the expansion of Cowork to the web and mobile clients in beta, with the explicit goal of letting employees monitor long-running, asynchronous agent sessions from any device, send approval requests, and pick sessions up on a different machine (InfoWorld).
The web and mobile rollout is the surface trigger for the data, not the underlying finding. The finding is that organizations are routing Cowork into business operations work: reconciliation, vendor follow-up, drafting internal documents, scheduling, compliance paperwork, and the long tail of process-driven tasks that previously sat with operations teams. The web and mobile clients matter because that work is asynchronous. A multi-step agent job can run for hours while the user is away from a desk, and the new clients exist to let a human step back in to approve a step, redirect a workflow, or close a session from a phone.
Pareekh Jain, principal at Pareekh Consulting, framed the data in productivity terms. Operations delays have historically been a drag on CIO metrics, and a category of work that automates the delay-prone steps is a direct lever on cycle time. Bhupendra Chopra, chief revenue officer at Kanerika, took the frame further. The enterprise pattern emerging from Cowork is a shift from employees executing the work to employees supervising AI outputs and applying business judgment to the results, a change that affects hiring profiles and operating models as much as tool choice (InfoWorld).
Operations-heavy deployment puts a different pressure on agent governance than a developer-heavy one would. A 2025 survey of 445 IT and security leaders by Gravitee found 88% of organizations reporting at least one confirmed or suspected AI agent security or privacy incident in the past year, and only about 14% saying they had deployed agents with full security and IT approval (Gravitee). When the work being automated is internal business processes, the data the agent touches tends to be broader and more sensitive than code repositories, which puts the approval gap in sharper relief.
The mechanism by which a misaligned agent causes damage is also different in an operations context. Noma Security's "GitLost" disclosure, reported by InfoWorld, showed an unauthenticated attacker using a crafted public GitHub Issue to get an AI agent with private repository read access to exfiltrate and publish the contents, a prompt injection problem expressed as an authorization boundary failure, not a content filter failure (InfoWorld). The same shape of problem, an agent following instructions from a document it was handed rather than from its operator, lands in operations work on contracts, customer records, and financial files, not only on code.
Two limits belong on the finding. The 1.2 million sessions are Anthropic's own first-party data, useful for the shape of Cowork usage and not a market-wide measurement. Software development may also be split across other Anthropic products, including Claude Code, which would push some coding work out of the Cowork denominator. Cowork itself is in beta, and the web and mobile clients are in beta, so the usage mix can shift as the product moves toward general availability.
The next milestone to watch is general availability of the web and mobile clients, which will broaden the install base that feeds the usage data, and any subsequent Anthropic release that splits out enterprise-wide agent telemetry by department. Operations work at 33.4% of a beta product is a signal. The same number, on a GA product, would be a measurement.