Claude Cowork has had its own browser since 26 August 2026. This Anthropic agent no longer just answers in a conversation: it opens a folder on the user's computer, carries out a task in several actions and delivers a concrete result for the business. Claude Cowork changes the nature of the work you hand to an artificial intelligence tool. This guide explains how it works, where its limits lie and, above all, what no article on the subject covers yet: how to measure real usage of this agent once it is deployed in a team.
Understanding Claude Cowork: definition, how it works and how it compares with Claude Chat and Claude Code
A simple definition: the office agent built into Claude
Claude Cowork is a feature of the Claude app that turns the conversational assistant into an agent able to act directly on an organisation's files and on-screen documents. In practice, the tool accesses a working folder through an authorised connector, follows a task plan and works in the background to produce a usable result, with no prior technical training needed for a non-developer user in the company. Each task given to the agent stays tied to a specific folder, on the computer or through a connected application. Anthropic builds Claude Cowork on the same technical foundations as Claude Code, with capabilities called skills, a connector system and controlled access to the information needed to extend what the agent can do at the office. It is this agentic architecture, designed for companies and their teams, that sets Cowork apart from a simple chat that can write a text on request.
What is the difference between Claude Cowork and Claude Chat?
Claude Chat answers a question in a conversation. Claude Cowork carries out a task end to end through a series of actions. Where the chat waits for a new instruction after each reply, Cowork opens a session, acts on files or a website, then comes back with a deliverable ready to approve. The shift is about control: the chat keeps the user in the loop at every step, while Cowork acts with broader autonomy over the length of a session. On the settings side, Cowork logs each step in a tracking note you can review afterwards, something a simple email exchange does not allow.
What is the difference between Claude Cowork and Claude Code?
Claude Code remains the developers' tool, designed for writing and fixing code in a terminal or development environment. Claude Cowork aims at the opposite: office tasks that do not involve programming languages, such as writing reports, sorting files and preparing presentations. More than 90% of Claude Cowork usage has nothing to do with software development, according to figures shared by Anthropic (Blog du Modérateur, 2026). The two products share the same agentic engine, but Cowork is aimed at people who do not code, whereas the architecture of a Claude Code harness remains tailored to engineering.
When did Claude Cowork launch and how has it evolved?
Claude Cowork launched as a research preview in late January 2026, restricted to a small circle of macOS users. Powered by the Claude model, the tool then gained autonomy month after month. On 25 August 2026, Anthropic unified memory between Claude and Cowork: what one learns from a conversation now applies to the tasks delegated to the other. A few days later, Claude Cowork got its own built-in browser, able to read a site, click and type text on the user's behalf.
Analysing the request and building an action plan
A Cowork session always starts with a careful reading of the request. Claude restates the goal, identifies the files or tools it needs, then builds a visible action plan before acting. This scoping step limits drift: the user sees what the agent is about to do and can adjust the instruction before the task actually starts.
Splitting the work into subtasks and coordinating sub-agents
For a complex task, Claude Cowork breaks the work into subtasks and can delegate some of them to specialised sub-agents working in parallel. This coordination makes it possible to handle in a single session what would take several exchanges with a standard chat, for example reconciling data from several files while another sub-agent writes the summary.
Running in the background, on the web or locally
Once the plan is approved, Cowork executes without waiting for confirmation at every line. The work can run in the background while the user moves on to something else, on local files stored on the computer or on a website opened in the built-in browser. This ability to keep going without constant supervision is what most clearly sets Cowork apart from a standard conversational assistant.
Delivering a result ready to review, not an answer to rework
Claude Cowork does not return plain text to copy and paste. The result takes the form of a document, a report or a file edited directly, along with a history of the actions taken. The user reviews a deliverable rather than rephrasing a question, which changes the nature of the review work asked of teams. The diagram below sums up the four steps of this cycle.
Comparison table: file access, execution, autonomy, target user
The three Claude products share the same engine but meet different needs. The following table sums up what sets them apart and, above all, what to measure once one of them is deployed in a team.
| Criterion | Claude Chat | Claude Cowork | Claude Code |
|---|---|---|---|
| File access | One-off reading via upload | Read and write on connected folders | Read and write in a code repository |
| Execution | One reply per exchange | Long session with an action plan | Runs commands in the terminal |
| Autonomy | None, the user stays in the loop | High, with configurable approval | High within the code scope |
| Target user | General public | Non-developer business teams | Developers |
| What to measure once deployed | Volume of active conversations | Real usage rate vs. activated licences | Adoption rate per technical team |
That last row is the one no vendor details. It belongs to the company deploying the tool, not to the vendor building it.
What this technical distinction means you need to measure once the tool is deployed
Three products, three usage profiles, three distinct management questions. Lumping them together skews any adoption analysis: a Head of AI who counts Claude licences without distinguishing Chat, Cowork and Code ends up with a figure that cannot be used. Across the AI estates we analyse, 20 to 40% of paid AI licences go unused, all tools combined. This holds for a Cowork rollout too: buying a plan guarantees neither adoption nor return on investment. The question is not how many accounts are active but which real uses generate value and which ones sit idle.
Getting set up: Claude Cowork connectors, onboarding and pricing
Native connectors: Slack, Microsoft 365, Google Drive, Zoom MCP
Claude Cowork plugs into the tools teams already use. Available connectors include:
- Slack, to read and draft messages in authorised channels
- Microsoft 365, to open and edit Word, Excel or PowerPoint documents
- Google Drive and Gmail, to access files and email
- Zoom via an MCP connector, added at general availability on 9 April 2026
Each connector is switched on individually, and the agent only sees the folders and accounts that have been explicitly authorised.
The plugin system and internal marketplaces
Beyond connectors, Claude Cowork accepts plugins that add specific business capabilities. A company can build its own plugin for a specific internal project or pick from a marketplace of community-shared plugins, much like the plugins that extend Claude Code to other models. This modular system explains why two Cowork deployments can have very different scopes: a marketing team's setup looks nothing like a legal team's.
The built-in browser: driving the web without leaving Cowork
Since 26 August 2026, a task that needs a website opens a browser directly in Cowork's side panel. Claude navigates it, reads the content, clicks on elements and types text on the user's behalf. This feature widens the scope of action well beyond local files: researching a supplier, checking a price or booking an appointment online becomes a task you can delegate like any other.
Agent Skills: tailoring Claude to a specific business need
Agent Skills are reusable instructions that teach Claude Cowork how to handle a recurring type of task in a given line of work, unlike a plugin, which connects an external tool. A skill can encode the expected format of a sales report or the structure of committee minutes. Once created, it applies to every similar new task without anyone having to explain the context again.
Which plans include Claude Cowork?
Claude Cowork is available on Claude Pro, Max, the Team plan and Enterprise, with usage levels rising from one plan to the next. The Pro plan covers regular individual use and the Max plan multiplies the available capacity for heavy use, while the two team tiers add the administration features needed for professional use: access control, per-group spending limits and technical oversight. Pricing is set by Anthropic and changes regularly, so check it at the source before any budget decision.
Installing and setting up a first Cowork session
Access to Claude Cowork goes through the Claude Desktop app, which you use to launch a first local session. After signing in to their account, the user picks a working folder, switches on the connectors they need, then writes their first task in natural language. An installation guide specific to each organisation is still useful, if only to set the authorised folders and the default approval mode before the first team-wide use. Planning a short internal training session on this guide keeps internal clients from discovering the usage rules after the fact.
Web, mobile and desktop: which devices does Cowork run on?
Long limited to the desktop app on macOS, Claude Cowork is now available on the web and on mobile. Users can start a task from their computer at the office, follow its progress from their phone on the move, then come back to the office to approve the result. This continuity across interfaces matches an increasingly mobile way of doing office work.
Session and quota limits to know about
Every Cowork session uses quota, even more so when auto-approval mode chains actions without interruption. Data processed during a session is deleted after 30 days by default. A complex multi-step task can therefore cost more quota than standard conversational use, a point to factor into any full-cost calculation before rolling it out across a team.
And the full-cost question goes beyond the quota of a single session.
Have you just deployed or authorised Claude Cowork in your team? Before extending access to the whole organisation, there is a factual way to find out what this usage really produces. Idun Group analyses the logs and exports of the AI platforms you already administer to map real usage, without installing anything on workstations. Getting a report on your real usage is the first step before any decision to expand.
Security, data governance and limits of Claude Cowork
Access control and administrator roles (RBAC)
Administrators of a Team or Enterprise account can define who is allowed to open a Cowork session and on which folders. This role management, added at general availability on 9 April 2026, comes with per-group spending limits. An administrator thus keeps control of the account without having to approve each task individually: the autonomy given to the agent is set upfront, by profile, rather than at each use.
Which documents and files can Claude Cowork read or edit?
Claude Cowork only reads folders explicitly connected by the user or administrator. Once access is granted, the agent can read, create and edit text files, Excel spreadsheets, presentations and other common office formats. There is no automatic access to the rest of the computer or the company network: the scope remains what the user has explicitly opened.
Data retention and shared memory between Claude and Cowork
Since 25 August 2026, what Claude learns about a user in conversation also applies to the tasks given to Cowork, and vice versa. This shared memory makes delegated tasks more relevant, but it shifts the data governance question: information given in chat can now influence a task the agent runs several days later. The confidentiality question deserves to be put to the data protection officer before any team-wide rollout, backed by an internal note that sets out which version of the rules applies.
What Claude Cowork does not say about its own usage
No official Claude Cowork documentation answers the question that matters most to a finance department: is the tool actually used, by whom and for what measurable benefit? Anthropic describes what the agent can do, not what its real usage produces once it is rolled out to a team.
The gap between declared adoption and real scale-up sits exactly in this blind spot: adopting an agent is not enough to know what it brings in. It is a question of measurement, not of product features.
What Cowork cannot do well yet
Claude Cowork is still a fast-evolving product, which means instabilities documented by early users, particularly on long sessions or very complex tasks. A large share of agentic AI projects will be abandoned before they reach production, because of runaway costs and business value that was poorly scoped upfront. Treating Cowork as a finished product rather than as a still-young agent invites exactly this kind of disappointment. Every article comparing successive versions of the tool is a useful reference point for a client who wants to see what an update really contains before adopting it across a team.
The risk of deploying without a shared framework
Authorising Claude Cowork team by team without a common framework quickly leads to inconsistent usage, with each team choosing its own connectors and its own approval mode. AI agents are now in production in roughly one company in three, led by banking and insurance, with healthcare and the public sector well behind. A minimal framework (which folders, which default approval mode, which pilot team) keeps every deployment from starting from scratch. A clear project strategy sets these rules before any new plugin or API connector, with a dedicated test environment and a human who also remains accountable for the final decision.
These three rules require no extra tooling, only a decision made before the first deployment rather than after the first incident.
Claude Cowork in your existing AI estate: one more building block or a duplicate?
Claude Cowork never arrives alone. In an organisation that already uses Microsoft 365 Copilot, ChatGPT Enterprise, Gemini or internal agents, the real question is not what Cowork can do on its own but what it adds, duplicates or replaces in that installed estate. Enterprise applications increasingly embed a specialised agent. The trend has clearly accelerated over the past two years. Cowork therefore joins a fast-crowding landscape that nobody necessarily has a full view of. The same need (writing a report, processing a spreadsheet, finding information) can today be handled in three or four different tools, depending on each person's habits. Placing Cowork within this estate before extending it to a whole team avoids stacking one more licence on top of uses already covered elsewhere.
Concrete Claude Cowork use cases in the enterprise
Producing formatted documents and presentations
One of the most cited applications of Claude Cowork remains producing complete documents: a summary report, a formatted presentation or meeting minutes structured to a given template. Claude Cowork gathers the necessary sources, applies the expected formatting and delivers a file ready to review rather than a draft that has to be reworked from scratch.
Research, monitoring and information synthesis
With its built-in browser, Claude Cowork can carry out targeted web research, compare several sources and draw a structured summary from them. This use case is a useful replacement for a manual press review or time-consuming competitive monitoring: the agent goes through several pages, extracts the relevant information and presents it in a format the user has defined in advance.
Data processing and automating recurring tasks
Some tasks, rooted in a repetitive workflow, lend themselves particularly well to task automation with Claude Cowork:
- Reconciling data from several spreadsheets
- Scheduling recurring reports, ready for approval every week
- Analysing a batch of contracts or similar documents
- Sorting and filing documents according to a defined rule
This kind of repetitive task is precisely where time savings are easiest to measure, provided you know how many times the task is actually run.
Uses that remain out of scope
Claude Cowork is not designed to automate everything indiscriminately. Sending LinkedIn messages automatically, for example, falls outside the scope Anthropic intends and puts you at risk of having the account blocked or breaching third-party platforms' terms of use. Setting these limits before deployment prevents individual uses that would escape any shared framework.
What Claude Cowork changes for managing your enterprise AI
Measuring who really uses Claude Cowork, and for what
An activated licence is not usage. That is the starting point of any audit of real AI usage, and it is what we measure at Idun Group once an agent like Claude Cowork is deployed. Across the AI estates we analyse, 20 to 40% of paid AI licences go unused. Knowing who actually opens a Cowork session, for what type of task and how often completely changes the decision you make at renewal time. It is this measurement, rather than intuition, that feeds an AI portfolio strategy.
Knowing the full cost of widespread Cowork usage
The cost of a deployment is not limited to the price of the chosen plan. On top of it comes the consumption from tasks run in the background and from activated connectors (Microsoft 365, Google Drive, Zoom), each with its own usage volume. Targeted and measured in your organisation: a 30 to 50% reduction in the cost of the targeted uses, never the whole bill, never a result promised in advance. That figure cannot be decreed: it is measured line by line, agent by agent, before any renewal decision.
What a company puts in place once the measurement is done
Measuring is not enough. The steps that follow are ranked from least costly to most committing. They are three forms of tooling, never discipline rules. Publish shared assets first: a Cowork skill, project or template is created locally and stays private, so everyone rebuilds the same report template in their own corner. Publishing them across the organisation requires no budget. Connect the agent where the work already happens next: a use that forces people to leave their document space or their email does not take hold, whereas a connector to the everyday tool turns voluntary use into natural use. Pool agents last, on the use cases the measurement has identified as recurring: when twenty people each automate the same meeting report on their own, one shared agent replaces twenty private automations. It is the most committing lever. It is also the only one decided on data rather than intuition.
Industrialise, consolidate or cut: the decision that follows measurement
Market studies by Gartner and McKinsey estimate that 30 to 50% of AI initiatives do not deliver what they should. Once real Claude Cowork usage has been measured in your organisation, three decisions open up: industrialise what creates value, consolidate redundancies with the rest of the AI estate already in place, cut what consumes quota without ever really being used. It is this decision, not the tool itself, that determines the return on investment.
