Key Takeaways

  • Anthropic launched mobile and web access for Claude Cowork and shared data showing most usage is for business operations, not software development
  • The move positions Cowork against rival agentic tools while highlighting rising demand for AI support in everyday productivity tasks
  • Security concerns and geopolitical pressure create added complexity as Anthropic deepens its enterprise footprint

Anthropic’s latest expansion of Claude Cowork is landing at a moment when enterprises are rethinking where AI provides the most practical value. The company has brought Cowork to mobile and web, broadening a product that had been confined to the desktop environment. The shift changes how knowledge workers can delegate and oversee tasks throughout the day. It also lands alongside new usage data that suggests the broadest demand for agentic AI is not in engineering teams but in the messier universe of operational workflows.

Anthropic analyzed 1.2 million anonymized Cowork sessions across more than 600,000 organizations between May 11 and May 31. Business process and operations tasks, such as reconciling spreadsheets or compiling updates, accounted for 33.4% of all sessions. The next largest category was content creation and copywriting at 16.4%. Software development represented only 8.7%. These numbers echo the pattern described by research firms like Gartner, whose work has pointed to office productivity and collaboration as the top areas where enterprises see early gains from generative AI. It also aligns with broader industry observation on generative AI’s first wave, including reporting from outlets like Reuters and Fortune.

The company framed this activity as the work around the work, meaning all the drafting, structuring, and organizing that sits between the core responsibilities of a job. Anyone who has spent time preparing a status update or merging multiple sources into a single slide deck probably understands the appeal. It also answers a question some executives quietly ask: do AI agents need to write code to drive operational efficiency, or can they focus on routine, cross-functional tasks that clog up the week?

The new mobile and web features support that second scenario. Sessions now sync across devices, and Cowork can execute scheduled tasks even if the user is offline. Decisions that need human judgment get flagged to the user’s phone. Together, these capabilities resemble an assistant that operates in the background. The mobile and web versions run server-side, a notable change because the desktop version relies on a local virtual environment to carry out tasks.

Anthropic also unified the interface across chat and Cowork so projects and artifacts persist across surfaces. The company doubled Cowork limits through August 5 to nudge adoption.

Behind the scenes, there is a clear competitive logic. Anthropic already has traction with developers through Claude Code, but the larger market sits with millions of professionals who live in spreadsheets and email threads. Tools like OpenAI’s agent-infused ChatGPT or Google’s Gemini in Workspace aim at similar territory. The shift also tracks with commentary noted in independent learning platforms such as DataCamp, which has described rising interest in AI companions that assist with documentation, analysis, and operational tasks. A related analysis on Elephas describes Cowork as an early example of a persistent AI agent that fits directly into daily knowledge work.

Claude Sonnet 5, launched last week, powers Cowork with improved reasoning and task durability. Claude Tag, released two weeks earlier, offers a multiplayer model for Slack channels. Anthropic’s head of product for Claude Code and Cowork described this distinction in Fortune, noting that the Code, Cowork, and chat experiences are largely single-player, while Tag is designed for multi-user collaborative contexts. Combined, they provide an enterprise entry strategy that pulls both individuals and teams into the Claude ecosystem.

Security researchers at Armadin in early July called out what they described as a full sandbox escape on Windows, detailed by outlets like SiliconANGLE. The vulnerability required local code execution, and Anthropic noted it did not qualify as a standalone security issue. Still, the situation rekindled discussion about how agentic AI interacts with local environments. Shifting Cowork’s execution to the server for mobile and web reduces exposure to local risks, but it raises different questions about prompt injection, data handling, and oversight, especially for workflows that inspect email threads or compile reports autonomously.

Geopolitical tension adds another layer of complexity. CNBC reported that Alibaba will ban employees from using Anthropic tools starting July 10 after Anthropic accused the company of carrying out a large-scale distillation attack. Meanwhile, Anthropic has committed to a $19 billion, 20-year data center lease with TeraWulf in Kentucky, according to Reuters. These investments suggest confidence in enterprise demand, although they arrive amid cross-border friction that complicates global adoption.

One curious detail in Anthropic’s own report is the admission that its classification pipeline changed around May 11. That explains why the analysis window is short. It also hints at how difficult it is to map the long tail of knowledge worker activity into neat categories. Marketing, HR, and finance work, for example, blend into the business operations bucket. Personal assistance accounted for 3.8% of sessions.

Enterprises are increasingly adopting AI tools to assist with everyday operational workflows. This expansion highlights an industry shift toward AI agents that support the small, connective tasks required to keep organizations moving forward.