Claude Cowork's Biggest Use Is Office Busywork (33.4%), Coding Just 8.7%: Interface Shape Decides How People Use AI
- Anthropic sampled 1.2 million Claude Cowork conversations from May 11–31, 2026, covering more than 600,000 organizations, and sorted them into 20 work categories.
- The top category is Business Process & Operations (consolidating scattered updates, building onboarding checklists, reconciling spreadsheets, etc.), at 33.4% — more than double the runner-up.
- Second place is Content Creation & Copywriting (drafts, slides, posts, proposals, etc.), at 16.4%; together the two categories account for roughly half of all usage.
- Software Development is only 8.7%. Anthropic's explanation: developers write more code in Claude Code (terminal), while Cowork (chat interface) handles the communication and coordination work that surrounds a role.
- Anthropic notes in the methodology section: the sample comes from rate-limited sampling, not a full-population share; the 20 categories have no standalone Marketing, Finance, or HR category; and about 5% is personal, non-work usage.
What Is Cowork Actually Used For
Anthropic published a blog post on July 7, 2026, disclosing a sampled categorical analysis of 1.2 million Claude Cowork conversations from May 11–31.
Many people assume an AI assistant capable of working continuously on its own gets used mostly for coding. But in the data, Software Development is only a small slice. What actually drives Cowork usage is office busywork — the kind with no clear job-title ownership that still fills everyone's calendar.
What Cowork Is, and How It Grew
Claude Code launched in 2025, originally built for developers to use in the terminal. But plenty of people who'd never written code started using it too — to organize folders, delete duplicate files, write spreadsheet formulas. For them, the terminal was still an intimidating black box.
So Anthropic built Claude Cowork: taking Claude Code's agentic, work-continuously-on-its-own capability and moving it into the chat interface people already used. Since launch, its capabilities have kept expanding outward.
Which Category Takes the Biggest Share Among the 20
These 1.2 million conversations were sampled from May 11–31, 2026, covering more than 600,000 organizations, and sorted into 20 work categories. Here's the full ranking:
There's a clear gap between the top two and everything else: Business Operations alone takes a third, and adding Content Creation, third-place Software Development comes to less than a fifth of the two combined. Below is Anthropic's original chart, which you can treat as the source citation. The official chart only plots the top-ranked categories and lumps the smaller ones into "Other"; the finer breakdown above (Legal & Compliance, Customer Support, etc.) comes from the blog post's body text and the methodology notes at the end.
Why These Two Categories, of All Things
Why do Business Operations and Content Creation take the top two spots? These two categories share something in common: they're rarely the core responsibility of any specific job title. They're more like the supporting work that revolves around a role — stitching together scattered information, pushing projects forward. Anthropic gave it a name: connective work.
Connective work refers to the supporting work built around a role that ties scattered information together and pushes projects forward — writing status updates, building slide decks, reconciling spreadsheets, pulling together onboarding checklists. This work rarely appears in anyone's job description, yet almost every job comes bundled with it. A spreadsheet pulls scattered data points into one place for comparison and tracking; a slide deck explains an idea to an audience with varied backgrounds; an onboarding checklist helps a new hire pick up the unwritten know-how of a company.
It's like the oil between a machine's gears: it doesn't directly generate the core power, but without it, the whole machine grinds to a halt.
Anthropic gives three concrete job examples showing that people hand this kind of work to Cowork in order to save time for the parts that actually require professional judgment.
Put side by side, all three examples follow the same formula:
Anthropic's own framing of this division of labor: people use Cowork to assemble scattered information into a usable shape, freeing themselves up to exercise their professional judgment. By this account, what AI takes over first is the padding layer surrounding professional judgment; the judgment itself stays in human hands.
Drawing It as an Iceberg
The visible, job-title work is just the tip of the iceberg above the waterline; below the surface — accounting for nearly half of all usage — is the connective work nobody wants to claim.
Put these two big categories alongside all the rest on one bar, and it's almost an even split.
Why Don't Coders Code in Cowork
Software Development is only 8.7% in Cowork — that looks small. Developers are, of course, still using AI to write code; they've just kept coding in a different tool: Claude Code in the terminal. The same person does two different kinds of work across the two interfaces.
- Write code
- Debug
- Ship to production
- Write status updates
- Build slide decks
- Coordinate communication
Interface shape, to a large degree, determines how people use AI. The same model gets used to write code when placed in a terminal, and to write status updates and build slide decks when placed in a chat box. The Decoder's coverage reads the "low coding share" as evidence for this argument.
This argument has a rare condition working in its favor here: Cowork's origin makes it something like a naturally controlled experiment. The official blog states plainly that Cowork simply moved Claude Code's autonomous, continuous-work capability into the chat interface, unchanged. The underlying capability is identical — the only thing swapped out is the shell. So the usage distribution across the two entry points becomes a direct comparison of just how much power the interface holds:
What gives the interface this much power? Break it down, and two forces are at work simultaneously.
The terminal requires typing commands, which naturally keeps out people who don't code; the chat box has zero barrier — anyone can walk in. Different people coming in means different work getting done. The official blog tells this origin story itself: after Claude Code launched, plenty of people who'd never touched a terminal pushed through anyway to organize folders and write spreadsheet formulas, but for most people, the terminal remained an intimidating black box. Cowork was built for them.
A tool's appearance carries a built-in assumption about "what it's for." Take the same motor: put it in a drill, and everyone uses it to bore holes; put it in a blender, and everyone uses it to juice. Nobody juices with a drill. The motor hasn't changed — the shell is what dictates the use. The terminal looks like a place for technical work; the chat box looks like a place for talking and writing. Usage follows appearance.
One boundary needs to be honestly flagged: what this data can directly prove is only that "usage distribution differs between the two interfaces." It can't on its own distinguish whether the same person changes their usage when they switch interfaces, or whether the two interfaces simply attract two different crowds. Anthropic's blog offers the former explanation: the same developer writes code in Claude Code, and once in Cowork, does the communication work that surrounds their development job.
A Bigger Shift Behind This
Zoomed out, this usage data is part of a broader shift already underway. A survey paper from Meta, Stanford University, and the University of Illinois argues that code is shifting from being the final product delivered to users to being the operational layer where AI agents actually run.
Code was the final product, written for people to read and delivered to users.
Code is the foundation layer where AI agents run and get work done.
The paper treats Claude Code, Claude Cowork, and OpenAI's Codex (also known as ChatGPT Work) as representative examples of this trend. Cowork's data is one concrete slice of this larger thread.
A Few Things Worth Noting About This Data
This is just a snapshot — Claude Cowork is still new, and its usage patterns are shifting fast. Before reading these numbers, there are a few things worth knowing; they determine how far these percentages can and can't be read into.
Available Now
Claude Cowork is now available to all Claude users. To get started, check out the official product page, or take Anthropic's "Introduction to Claude Cowork" course.
Anthropic says this data is meant as a reference point for anyone figuring out how to bring AI into their daily work, and that it will keep publishing new data as usage evolves.
What developers do in Claude Cowork is communication-driven connective work that surrounds every role — including software engineering itself.Anthropic, "How people are using Claude Cowork"
What Claude Cowork Does Most Isn't Coding — It's the Office Busywork Nobody Wants to Claim
Anthropic sampled 1.2 million conversations — office busywork takes the top spot, coding ranks only third. One page with charts explains why.
↓ Read it in one page · includes one animated chart
Claude Cowork is an AI assistant: no need to touch the dark command-line window programmers use (called a "terminal") — regular people just type, and it works continuously through a task on its own. It launched in January 2026, and now runs on desktop, web, and mobile. This time Anthropic wanted to nail down something nobody had data on: what people actually use it for.
The result runs against most people's intuition.
An AI that can work on its own ≈ a coding tool — surely programming is the main use.
The top spot is office busywork (Business Process & Operations) at 33.4%, Content Creation is second at 16.4%, and coding ranks only third at 8.7%. The top two together are already close to half.
Coding — the most attention-grabbing use case — turns out to be the smaller slice. Why do those particular two categories pull so far ahead? They share something in common.
Business Operations and Content Creation share something: neither is the core responsibility of any single job title — both are supporting work that revolves around a role, stitching scattered information together and pushing projects forward. Anthropic gave it a name, "connective work": writing status updates, building slide decks, reconciling spreadsheets, pulling together onboarding checklists — rarely written into anyone's job description, yet almost impossible for any job to shake off.
There's also an intriguing contrast: the same developer keeps coding in the command-line Claude Code, but once inside the chat interface Cowork, does work-related communication odds and ends instead. Anthropic's takeaway: what a tool looks like, to a large degree, determines what people use AI for.
"What share" sounds abstract. Spread out 100 conversations, and it's obvious at a glance:
All these percentages come from Anthropic's own rate-limited sampling (capped per hour, reflecting share only — not total volume or growth rate), classified by an automated system with no manual per-conversation review and no third-party replication; the sample is also mixed with about 5% personal, non-work conversations.
isn't that for coding?
everyone's
first reaction
it's flipped!
nobody escapes it
busywork + copy ≈ half
Debugging · deploying
- × All figures are from Anthropic's own sampling
- × Capped hourly — shares only, not totals
- × No third-party replication · ~5% personal chats mixed in
is the unclaimed busywork
but never your job description
Members can copy or download the full Markdown, and let their AI subscribe and ask.
Found this useful? Leave your email — every new explainer, the moment it's out. Unsubscribe anytime.