Anthropic put its knowledge-work agent on phones and web, then published numbers showing that coding is under 9% of what people actually use it for.
Answer capsule: Anthropic expanded Claude Cowork, its Claude Code-style agent for general knowledge work, from a desktop-only app to web and mobile, rolling out first to Max subscribers. Alongside the launch it released usage data from 1.2 million sessions showing that business operations (33.4%) and content creation (16.4%) dominate, while software development accounts for just 8.7%. The move reframes the agent market away from developer seats and toward the whole org.
Anthropic sells Cowork as the descendant of a coding tool. Then it published a study proving that almost nobody uses it to code. That is not an accident, and it is not an embarrassment. It is the pitch.
What Happened
On Tuesday, Anthropic brought Claude Cowork to web and mobile, extending a product that launched as a desktop-only app back in January. The rollout is in beta and starts with Max subscribers, the $100-a-month tier, with Pro, Team, and Enterprise access promised in the coming weeks.
The functional change is continuity across devices. You can start a task at your desk, watch its progress from your phone, and collect the finished output later, even with the laptop closed. Scheduled tasks now run in the cloud rather than depending on a machine staying awake. Anthropic’s example: set Monday’s client prep for 6 a.m., and Claude works through the email threads, transcripts, and recent news, builds the briefing doc, and leaves a follow-up email drafted but unsent for you to review over coffee.
Bundled with the launch was something more revealing than the feature list: a study of 1.2 million anonymized Cowork sessions from more than 600,000 organizations, sampled May 11 to 31. Business process and operations was the biggest category at 33.4%. Content creation and copywriting followed at 16.4%. Software development came in at just 8.7%. More than 90% of what people do with Cowork has nothing to do with writing code.
The Backstory
Cowork was born inside Anthropic’s coding-agent narrative. Claude Code built the company a serious developer reputation, and Cowork launched in January as “Claude Code without the code,” a way to hand the same multistep agent to people who never touch a terminal.
That framing was smart for a launch and limiting for a business. Developers are an obvious, well-defined, willing-to-pay audience, but they are also a capped one. Every AI lab is already fighting over the same engineering seats. The far larger prize is everyone else in the building.
Anthropic has been circling that prize for months. Two weeks before the mobile launch, it dropped Claude Tag, an always-on shared @Claude that lives inside Slack channels and acts as a team member the whole channel can summon. The enterprise appetite was already there: Anthropic recently passed OpenAI in business-paid AI subscription share, reaching 41% in May. Cowork on mobile is the next brick in the same wall.
The Plan
Read the two launches together and the strategy is a pincer. Cowork captures the individual: one person delegating tasks that follow them across desk, phone, and browser. Claude Tag captures the team: a shared identity embedded in the channel where work is coordinated. Between them, Anthropic is trying to own both the private workflow and the group conversation, the two surfaces where knowledge work actually happens.
The usage data is the marketing weapon. By publishing that coding is 8.7% and business ops is 33.4%, Anthropic is sending a direct message to enterprise buyers: budget agent seats for finance, HR, comms, and operations, not just engineering. It coined a phrase for the dominant use case, “the work around the work,” the connective admin that spans nearly every role but sits in no one’s job description. That framing turns a coding tool into an operations tool without changing a line of the product.
The Business Model Angle
This is a total addressable market expansion dressed as a product update, and it is worth being precise about why.
A coding agent’s TAM is bounded by the number of developers. An “everything agent” that reconciles spreadsheets, drafts proposals, builds onboarding checklists, and prepares client briefings has a TAM bounded by the number of knowledge workers, which is an order of magnitude larger. Anthropic just told the market, with its own data, that Cowork already lives in that bigger category.
Mobile and web are what unlock it. The desktop-only requirement was a friction tax: it capped the user base at people willing to install and run a heavy local app. Web and mobile remove the install, remove the “laptop must stay open” constraint, and let the agent bill for background work that runs while you sleep. Every one of those changes widens the funnel and lengthens the time the product is earning its keep.
There is a deeper strategic logic underneath. If frontier models are commoditizing, and the recent flood of capable Chinese open-weight models suggests they are, then the raw model stops being a durable moat. The moat migrates to distribution and workflow: whoever owns the surface where work gets done keeps the customer even when a rival ships a marginally better model. That is the same instinct now pushing enterprises to ration their AI spend and swap in cheaper models for routine tasks. Owning the workflow is how a lab stays sticky in a world where the model itself is increasingly interchangeable.
Anthropic is not alone in seeing this. OpenAI is running the identical play with Codex, expanding a developer tool into general enterprise work. Google has its always-on Spark agent. Startups are crowding in, with one raising $75 million to embed an AI coworker into Slack and Teams. The office productivity surface is the next battlefield, and everyone knows it.
The Risk
Three cracks are worth watching.
First, the data is Anthropic’s own, and it is rate-capped, automatically classified, and drawn only from Max subscribers, a paid tier that may not represent the broader base. The numbers are directionally useful and self-servingly convenient at the same time. Treat “coding is 8.7%” as a signal, not gospel.
Second, security. Giving a phone remote control over a desktop agent that can touch local files and a browser is a genuinely sharp edge. A manipulated instruction or a phishing link could trigger actions that are hard to undo. Security researchers have already published a claimed sandbox escape in Cowork on Windows; Anthropic disputes its severity, but the broader worry, that background agents on nontechnical users’ machines create visibility gaps, is not going away.
Third, the moat is not exclusive. If owning the workflow surface is the strategy, Anthropic is racing OpenAI, Google, Microsoft, and a wave of funded startups for the same real estate. Being first to publish the data is not the same as being first to lock the seat.
Quick Questions
What is Claude Cowork?
Anthropic’s AI agent for general knowledge work, built on the same technology as Claude Code but aimed at non-developers. It handles multistep tasks like building reports, reconciling spreadsheets, and drafting communications.
Who can use the mobile and web version?
Max subscribers first, in beta. Pro, Team, and Enterprise plans are due in the coming weeks.
What did the usage data actually show?
Across 1.2 million sessions, business operations led at 33.4% and content creation at 16.4%, while software development was only 8.7%. Over 90% of use was non-coding.
How is this different from Claude Tag?
Cowork is a single-player tool for individual task delegation across devices. Claude Tag is a multiplayer, shared identity that lives in a team’s Slack channel.
The Business Model Analyst Take
The headline is the mobile launch. The story is the reframe. Anthropic just used its own telemetry to reposition a coding tool as an operations tool, and it did it by publishing a number that looks unflattering on the surface and is strategically brilliant underneath. “Coding is only 8.7%” is not a confession. It is a flare fired at every CFO who thinks agent budgets belong to the engineering org.
For operators, there are two lessons worth stealing. The first is that publishing your own usage data can be a positioning move, not just transparency theater. Anthropic turned a research post into a category-defining argument about where AI value lives. If your product’s real use case is drifting away from your original pitch, the data behind that drift may be your best marketing asset, not something to bury.
The second is about where moats are relocating. When the core product commoditizes, and models are commoditizing fast, defensibility moves to the surface you occupy in the customer’s day. Cowork on mobile, Claude Tag in Slack, Codex creeping into non-dev work: these are all land grabs for the same thing, the spot where the work already happens. Win that spot, let context and habit pile up, and become the tool a team cannot cleanly walk away from. The model is not the moat anymore. The workflow is.
Based on reporting by Rebecca Bellan for TechCrunch, with usage data from Anthropic.
