An analysis of more than one million Cowork sessions shows that about half of all usage falls into two categories: business operations and text-based work.
Anthropic has broken down how people actually use its chat agent, Claude Cowork. The analysis draws on a sample of 1.2 million anonymized sessions from May 11 to 31, 2026, spanning more than 600,000 organizations. An automated system assigned each session to one of 20 work categories. According to Anthropic, about half of Cowork usage goes toward what the company calls "the work around the work," tasks that are rarely anyone's core responsibility but come with virtually every office job.
The largest category is "Business process and operations" at 33.4 percent. Anthropic says this includes things like pulling scattered status updates into a single report, building onboarding checklists, or reconciling spreadsheets. Second is "Content creation and copywriting" at 16.4 percent, covering drafts, slide decks, posts, and proposals. Together, these two broad categories account for the "roughly half" Anthropic describes.
Anthropic describes these activities as "connective in nature." Spreadsheets pull scattered data into one place, slide decks translate a decision for audiences with varying levels of background knowledge, and onboarding checklists give new hires access to an organization's institutional knowledge.
The remaining categories are much smaller. Software development accounts for 8.7 percent, DevOps and infrastructure for 7 percent, research for 6.4 percent, data analysis for 5.8 percent, document processing for 4.1 percent, and sales operations for 4 percent. Personal assistance sits at 3.8 percent, education at 2.4 percent, and meeting analysis at 1.8 percent.
The low share of coding confirms Anthropic's thesis that the interface plays a major role in how people adopt AI models. Developers still use Claude Code to build, debug, and ship software, according to Anthropic. In Cowork, they handle the communication tasks that surround their work instead.
This pattern tracks with experience. After Claude Code launched in 2025, a surprisingly large number of non-technical users showed up in the terminal, organizing folders, removing duplicate files, and writing spreadsheet formulas. For others, the terminal stayed a black box. Cowork was designed to bring those same agent-like capabilities to a familiar chat interface.
Anthropic illustrates the idea with three examples. A lawyer can have documents formatted and filed to free up more time for legal analysis. A hiring manager uses it to schedule interviews and summarize feedback so they can spend more time on candidates and work samples. A team lead can have slides built to explain a tough decision, buying them more time to actually make it. The blank screen is often the first hurdle, and Cowork helps turn thoughts and information into a first draft, Anthropic says.
Anthropic itself notes that Business Ops classifies activities, not professions, and that there are no separate categories for marketing, finance, or HR. Those functions are likely folded into Business Operations, which Anthropic says probably helps explain why the category accounts for a third of all usage.
The study has other limits. The sample captures only a fixed maximum number of sessions per hour, so peak times are underrepresented. The figures are percentages, not absolute volumes. About five percent of sessions involve personal use like hobbies or chatting, so the dataset doesn't exclusively reflect work. Anthropic says it will keep updating the analysis.
A review paper by Meta, Stanford, and the University of Illinois places systems like these in a broader trend. Code is no longer just the end product but the operational layer AI agents run on. Claude Code and Cowork, along with OpenAI's Codex (also known as ChatGPT Work), are considered prominent examples of this shift.