Anthropic Analyzed 1.2 Million AI Agent Sessions: The Top Use Case Is the Office Work Nobody Wants
Half of all Claude Cowork usage goes to status reports, checklists, and slide decks — what Anthropic calls 'the work around the work.' The data says a lot about where AI actually helps.
What do people actually do with an AI agent once they have one? Anthropic just published an unusually concrete answer. The company analyzed 1.2 million anonymized sessions of Claude Cowork — its agent for everyday office work — from more than 600,000 organizations, collected over three weeks in May. The winner, by a wide margin: the mundane organizational work that comes with every office job but is rarely anyone’s actual job.
The two biggest categories tell the story. “Business process and operations” leads at 33.4 percent — think pulling scattered status updates into one report, building onboarding checklists, or reconciling spreadsheets. “Content creation and copywriting” follows at 16.4 percent: drafts, slide decks, posts, proposals. Together that’s half of all usage, which Anthropic sums up as “the work around the work.” Everything else is smaller: software development at just 8.7 percent, research at 6.4, data analysis at 5.8.
That low coding number is worth a second look. It’s not that developers don’t use AI — they use Claude Code, Anthropic’s terminal-based coding tool, for the technical work, and turn to Cowork for the communication tasks around it. The interface, in other words, shapes the use: same underlying models, completely different jobs, depending on whether they’re wrapped in a terminal or a chat window.
What’s behind the numbers? Anthropic calls these tasks “connective” — spreadsheets pull scattered data together, slide decks translate decisions for different audiences, checklists pass institutional knowledge to new hires. It’s the tissue that holds organizations together, and almost nobody lists it as their core skill. That makes it the natural first target for delegation. A fair caveat on the data itself: the taxonomy classifies activities, not professions — there’s no separate marketing, finance, or HR category, which probably inflates that big Business Ops number. The sample also caps sessions per hour, underrepresenting peak times, and percentages say nothing about absolute volume.
What this means for you: If you’ve been wondering what to actually try with an AI agent, this is a map drawn from 1.2 million real sessions: start with the report you keep postponing, the checklist that lives in your head, the slides that explain a decision already made. That’s where hundreds of thousands of organizations are quietly finding the value — not in sci-fi automation, but in finally offloading the work around the work.
Sources
Study: 41 Percent of Long LinkedIn Posts Are AI-Written — the Highest of Any Platform
AI-detection company Pangram scanned over a million social media posts. One in four long posts across platforms is AI-generated — on LinkedIn it's four in ten. Here's how to read those numbers.
