How to use Claude Cowork for your business alongside your existing systems

This guide explains how to use Claude Cowork for your business. It starts with what the tool is, then moves through installing it and choosing its first tasks. Claude Cowork is Claude working directly with your files, folders and apps. It reads and edits them, and it produces real outputs such as documents and spreadsheets. Anthropic's help centre describes it as bringing the agentic capabilities of Claude Code, its coding tool, to knowledge work beyond coding, with no terminal required.

"Agentic" means you describe a task and Claude plans and carries out the steps, while you steer along the way. A chat works differently: you ask a question and it gives an answer. In Cowork, one request can mean opening several files, drafting a report and saving it to a folder.

Cowork is available only on paid plans: Pro, Max, Team and Enterprise. On desktop it runs on all four. On web and mobile it is in beta for Pro, Max and Team, and for Enterprise when an owner turns it on. The help centre's getting-started page also says Cowork is being merged into Claude itself, so Claude will decide whether your request needs a quick answer or a task. That change is reaching Pro and Max first, so the Cowork label may soon disappear from your menus.

If more than one person will use it, look at Team or Enterprise. On those plans an administrator controls several settings covered later in this guide, including which permission modes staff can use.

Cowork became generally available on macOS and Windows through the Claude Desktop app on 9 April 2026. The steps are the same on both systems. Install Claude Desktop, sign in with a paid account, then start a Cowork task from inside the app.

Two points often catch people out. First, a UK setup guide from April 2026 said Cowork was only available in the desktop app. That is no longer the full picture. The help centre now lists web and mobile in beta. It also says Cowork can run your sessions remotely in the cloud, again in beta, so your sessions and files stay with your Claude account on desktop, web and mobile.

Second, cloud sessions change where your files are processed. If Claude opens local files through the desktop app during a cloud session, that content is processed on Anthropic's servers rather than staying on your computer. Each task runs in an isolated, temporary environment created for that one session. The environment cannot reach your home or company network, and it is removed when the session ends.

On cost, the author of that April guide put Claude Pro at £15 a month in the UK. Use that as a rough starting point only. Check Anthropic's current price for the plan you need before you budget for a whole team.

Spend your first hour on four settings. Between them, they decide how useful Cowork is and how safe it is to run.

1. The folder. The first thing Cowork asks is which folder to work in. Create a dedicated working folder and copy in only the files the task needs. Do not point it at your whole drive.
2. Global instructions. These are rules Claude reads at the start of every session. Set them under Settings → Cowork → Global Instructions. Include your house style, your company name and anything Claude must never do.
3. Connectors. A connector is a link to a tool you use, such as email or a calendar. Add them under Customize → Connectors. Start with one connector and add more later.
4. Skills and plugins. A skill is a set of instructions for a recurring job. A plugin is a package that bundles skills, connectors and sub-agents for a role, team or company. You can browse plugins under Customize → Plugins.

The same April guide notes that Cowork includes built-in skills for Word documents, PDFs, spreadsheets and presentations. It also says you can build custom skills without code by describing the job in plain English. The Claude Marketplace lists 337 plugins. Several are labelled Anthropic verified, including an official Slack plugin that works from Cowork.

Install fewer plugins than you think you need, because each one adds tools Claude can use. On the Enterprise plan, you can turn on skill scanning, which checks skills and plugins for malicious content when they are installed.

Cowork has three permission modes. They control when Claude asks for your approval before it acts, for example before it uses a connector. Choose the mode per task, not once for everything.

In every mode, Cowork needs your explicit permission before it permanently deletes any files. On Team and Enterprise plans, the administrator decides whether Automatically approve is available to staff. It is available by default, so make a deliberate decision about whether to leave it on.

At Svennis we start each client team in Manual mode with a single working folder. We move a task to Automatically approve only once someone has reviewed its output on real work. That order costs a little time at the start and saves arguments later about what Claude was allowed to do.

Good first tasks read a lot and write very little. Each one should leave a person checking the output before anything goes outside the business.

1. A morning briefing

The author of the April setup guide built a briefing that connects to five tools and runs automatically. It is a scheduled task: a task that runs on a timer using the instructions, connectors and files you have already set up. His guide said your computer had to be on and connected. The help centre now says scheduled tasks run in the cloud, so they do not need your computer to be awake or the desktop app to be open.

2. Inbox triage with drafts only

The same author has Claude sort incoming email into four tiers. He reports that the Gmail connector creates drafts only and never sends without an explicit instruction. The Microsoft 365 connector now has write tools as well, so it can draft, send and organise email. A Microsoft Entra administrator must first consent to the new permissions, and an admin must turn them on for the organisation. Teams remains read-only.

3. Turning a folder into a document

Put meeting notes, an export and last month's figures into the working folder. Then ask for a summary report or a spreadsheet. Cowork's built-in document skills handle the formatting, and you review the file before sharing it.

Say you run a small professional services firm. You want Claude to prepare your Monday: triage the inbox and summarise last week's hours by client. The steps below use settings named in the sources.

1. Create a folder called Cowork Monday. Copy in the timesheet export and the client list.
2. Open Settings → Cowork → Global Instructions and add: Write in British English. Save email as drafts; never send. Do not move or delete files. Work only in the Cowork Monday folder.
3. Under Customize → Connectors, add Gmail, or Microsoft 365 once your Entra administrator has consented.
4. Set the permission mode to Manual.
5. Ask: Read my inbox from last week. Sort it into four tiers: reply today, reply this week, file, ignore. Draft replies for the first tier. Then summarise hours by client from the timesheet in a spreadsheet.
6. Approve or refuse each action as Claude asks. Read every draft and open the spreadsheet to check the totals against the export.
7. When the output has been right on a few Mondays, make it a scheduled task.

If a run goes wrong and you delete the task, it disappears from your task history immediately. It is deleted from Anthropic's backend storage within 30 days. Remember that sessions cannot be shared with colleagues, although you can share individual artifacts, such as the spreadsheet, that you create in a session.

Anthropic's guidance on using Cowork safely relies on a few terms. A trust boundary is the set of sources you consider safe and under your control, such as your own files or company communications. Read tools let Claude access content, such as your inbox. Write tools let it act, for example by creating a calendar invite or deleting a file. Prompt injection means malicious instructions hidden in content Claude reads. For such an attack to succeed, Claude must be able to read information from outside your trust boundary and also be able to take actions that could harm you.

One rule sits above the rest. You remain responsible for every action Claude takes on your behalf, including purchases, messages sent and whatever a scheduled task does.

Cowork is personal and session-based. Sessions cannot be shared, and memory is off by default for Team and Enterprise organisations. Your CRM (customer relationship management system) works differently. It holds shared records that every colleague relies on. That difference points to the right division of labour: the CRM stays the record, and Cowork reads from it and drafts work for it.

Some vendors now ship a direct route. Salesforce in Claude launched in beta as a plugin that brings a seller's accounts, opportunities and pipeline into Claude, with 37 pre-built sales skills. It is available on all paid plans, but only to organisations Salesforce approves through its beta sign-up.

If you run Zoho CRM or another system, check Customize → Connectors and Customize → Plugins for what exists today. If your tool is missing, expect extra setup. The April guide's author found that Claude did not connect to TickTick or Starling Bank out of the box.

Day to day, Cowork helps most with the work around the record. That includes preparing for a client call from exported notes, drafting follow-up emails and turning a pipeline export into a board summary. Updates to the record itself are better made by a person or by a proper integration. We cover that approach in putting Claude inside the tools you already use.

Cowork suits one person's recurring desk work. It suits less well a process that many people share, runs all day and changes customer records. The table sets out the trade-off.

A service desk shows the difference well. Tickets arrive all day, several people handle them, and every ticket must reach the right queue. That work belongs in the helpdesk, for example Zoho Desk, with the AI built into the flow. It should not depend on one person's Cowork session. Our post on a Teams service desk in practice shows that set-up. The companion piece on routing email right first time explains why accuracy matters more than speed there.

For a UK business, the first question is where your data goes. In a cloud session, local files Claude opens are processed on Anthropic's servers. Decide in advance which folders hold personal or client data, and keep those out of Cowork unless your data protection review allows it. Our checklist for European firms using Claude sets out what to check. The page on AI law in the UK covers the wider rules.

Next, set your defaults deliberately. On Team and Enterprise plans, memory starts off and Automatically approve starts on. Write both choices into a short internal policy so staff know what applies.

The UK public sector offers a useful pattern. The government's Incubator for AI has open sourced Lex, a semantic search layer over almost all legislation data, and its API is public. Lex can be connected to agents such as Claude Code, which shares Cowork's architecture. The Incubator has also open sourced a Parliament MCP for navigating Hansard through existing language models. Both show the approach this guide recommends: the authoritative records stay in their own system, and the AI reaches them through a defined connection rather than holding copies of its own.

1. Pick the plan. Pro suits a trial by one person. Team or Enterprise suits a rollout, because an administrator controls the permission modes.
2. Install Claude Desktop on one Windows or Mac machine. Create a dedicated working folder.
3. Write global instructions that ban sending, moving and deleting until you trust the output.
4. Add one connector and run the Monday example in Manual mode.
5. Take Anthropic's free course. The Introduction to Claude Cowork on Anthropic Academy has 14 lessons and a quiz and takes about 2.5 hours. Claude Academy also offers an AI Fluency course for small businesses with nine lessons.
6. List the processes that are shared, such as a support queue or order handling. Mark them for a proper integration rather than Cowork.

When that list is ready, read how to build AI into the systems you already run. It will help you decide which processes need an integrated build and which a person with Cowork can handle well.