What is Claude Cowork? An independent guide for managers and HR

What is Claude Cowork?

Cowork runs on the same agent technology as Claude Code, Anthropic’s developer tool, but needs no terminal. You describe an outcome, such as a summary of 300 open-text survey comments, and Claude plans the steps, runs code and hands back an Excel file with working formulas, a deck or a formatted document. It shows its plan and its progress at each step, and you can step in and redirect it mid-task. Tasks run in an isolated environment on Anthropic’s servers, so they keep going when you close the laptop. For local files, Claude goes through the desktop app and reaches only the folders you connect, so the app has to stay open (Anthropic).

In January 2026 I gave Cowork a folder of unstructured people analytics data from nine sources. Fifteen minutes later it had built a leadership deck on turnover, pay equity risks and recommendations. The video shows the older interface.

How is Cowork different from Claude chat, Claude Code and AI agents?

Use chat for a quick answer or a single draft. Switch to Cowork when the job spans several files or steps, or will take a while.

Cowork is an AI agent in the plain sense of the word, and I think it’s a good first one for most teams. Unlike the agents IT or a vendor builds into HR systems, you set up the task, read the plan and approve or stop actions, so the person who knows the work is the one watching it. My test of ChatGPT’s Operator agent on HR tasks is in AI agents for HR.

How much does Claude Cowork cost?

Cowork is included in every paid plan at no extra cost, but tasks draw on your plan’s usage, and Anthropic says Cowork uses up limits faster than chat. US list prices before tax, November 2026 (Anthropic).

Watch Settings > Usage for the first weeks before you decide who needs a premium seat. To make the usage last, Anthropic suggests grouping related work into one task, starting a new conversation for unrelated work and telling Claude when you only need a quick answer. Team and Enterprise include Anthropic’s Data Processing Addendum, with Standard Contractual Clauses, in the terms. Where Claude fits next to Copilot and ChatGPT is covered in Copilot vs ChatGPT.

What happens to your data in Cowork?

Five things HR, IT and your privacy lead should know before anyone connects a folder.

• Cloud sessions run on Anthropic’s servers. That includes local files Claude opens through the desktop app. Each session gets its own temporary environment, which can’t reach your company network (Anthropic).
• Business data is stored in the US, even though traffic may be processed in Europe, Asia or Australia (Anthropic).
• Personal plans can train on your content. With training on, chats can be kept in de-identified form for up to five years (Anthropic). Turn training off under Settings > Privacy, and keep employee and candidate data out of personal accounts.
• Session transcripts need Enterprise. Team and Enterprise can stream Cowork events to your own monitoring through OpenTelemetry, which Anthropic says doesn’t replace audit logging. Pulling Cowork transcripts through the Compliance API requires Enterprise (Anthropic).
• Local sessions stay on the laptop. On Enterprise, cloud sessions are off until an owner turns them on, so the work runs on the user’s computer, with code in an isolated virtual machine. Admins can’t manage or delete that local history centrally (Anthropic).

If data must stay in the EU or UK, IT can run the desktop app through Amazon Bedrock or Google Cloud in that region, with conversations stored on the device but without web, mobile or cloud sessions. Either way, write down which data never goes in. There’s a one-page policy outline in my AI governance framework.

How to set up Claude Cowork

1. Open Claude. To work with folders on your computer you need the desktop app for macOS or Windows from claude.com/download, with Linux in beta. Web, mobile and the Chrome side panel work too, though on Enterprise web and mobile stay off until an owner turns them on (Anthropic).
2. Select Cowork in the message box. If there’s no choice between Chat and Cowork, you have the merged layout and describe the task directly.
3. Choose Manual in the mode selector. Claude then asks before it acts. Auto, where Claude screens each action itself, is the default on Pro, Max, Ultra and Team. Skip checks nothing, and I wouldn’t use it with HR data.
4. Write global instructions once. They apply to every task and sit under Settings > Cowork, or under Settings > Your Claude in the new layout.
5. Connect one folder and write a brief. Put only what the task needs in the folder. Name the result, the files to use, what Claude must not touch and the format you want back, then read the plan before it starts. The examples below follow that pattern, the same one that makes a good prompt.
6. Check the output against the source. Spot-check numbers and quotes. You own whatever you pass on.

Global instructions to start from. Adjust the brackets.

I work in HR at a [US/UK] company. Use [American/British] spelling and our terms, such as [PTO or annual leave].
Never write out Social Security or National Insurance numbers, health information or an individual's pay, even if they appear in the files.
Don't invent numbers, quotes, dates or terms. Tell me when the material isn't enough.
Deliver Word, Excel or PowerPoint files and list the files you used.

Four ways to use Claude Cowork at work, step by step

Adapt these examples to your own files and terms. Start with the first one, which needs no personal data.

1. Find contradictions in the employee handbook

Input. Your current HR policies as Word or PDF files.

Read all the policies in this folder and list in Excel any contradictions between documents, details that look outdated, and questions an employee couldn't get a clear answer to. Give the document, section and a short quote for each. Don't change the files or make legal interpretations. Flag items we should take to employment counsel.

Check. Go through the list with the policy owner, and edit the originals yourselves.

2. Theme the open-text answers in an engagement survey

Input. The comments in Excel, without names or email addresses, and with groups too small to stay anonymous removed.

Read comments.xlsx, our answers to "What would make your workday better?"
1. Group the answers into no more than seven themes and count them.
2. Pick two or three quotes per theme, unedited, leaving out anything that could identify a person.
3. Make five slides for the leadership team with the themes and three questions they should decide on.
4. Make a sheet with each answer and its theme so I can check the grouping.
Don't invent numbers that aren't in the file.

Check. Read ten random rows. Does the theme fit, and do the quotes appear word for word in the source?

3. Prep for your weekly leadership meeting

Input. Meeting papers in SharePoint through the Microsoft 365 connector, which a Microsoft admin approves once and which searches everything you can see. Read the connector advice below first. Keep personnel cases out, since Anthropic advises against scheduled tasks that reach sensitive files.

Every Thursday at 3 p.m., read the new papers in Leadership/Papers. Make one page per item with the decision needed, the key numbers with document and page, two questions I should ask and what it means for my team. Don't change or send anything. Deliver the brief as one Word document.

Schedule. Type /schedule in the task, or open Scheduled in the sidebar. Scheduled tasks run in the cloud while your computer sleeps, unless they need local files or apps (Anthropic). Read each brief next to the papers for the first few weeks.

4. Write the job ad and interview guide, but leave screening out

Input. The role profile and your job ad template.

Using the role profile and our template, create a job ad of no more than 300 words in our tone, an interview guide with eight questions tied to the requirements and what a strong answer contains, and an Excel scorecard rating each requirement from 1 to 4. Avoid wording that could put off applicants, such as words about age or gender. Don't invent pay, benefits or terms.

Check. Read the ad next to the role profile. Don’t upload CVs or ask Claude to rank candidates. In the EU, AI that filters applications or evaluates candidates is high-risk under the AI Act, with obligations from 2 December 2027 (EUR-Lex). New York City already bars employers from using an automated employment decision tool without a bias audit in the past year (NYC DCWP). What else to watch in hiring is in AI in recruiting.

How Cowork reaches your other systems with MCP connectors

MCP, the Model Context Protocol, is an open standard for connecting AI tools to other systems. Anthropic created it, and ChatGPT, Gemini and Microsoft Copilot have adopted it. In Claude the connections are called connectors, and you add them under Customize > Connectors. For each tool you can choose Always allow, Needs approval or Blocked (Anthropic). On Team and Enterprise, owners add connectors for the organization, including custom ones to your own MCP servers.

Before HR connects anything, I’d follow three rules.

1. Start read-only. The Microsoft 365 connector is read-only until you turn on write tools, and it shows Claude only what the user can already see (Anthropic). Keep it that way until you’ve seen how it gets used.
2. Don’t connect the HR system first. It holds pay, sick leave and ratings. Start with policies and other material every employee can already read.
3. Review your SharePoint permissions. The connector searches all of SharePoint with the user’s access and can’t be limited to certain sites. If personnel folders are open to more people than HR, Claude reaches them too.

In this five-minute video I set up a connector in Claude and ChatGPT. The longer walkthrough is in MCP and AI connectors for HR.

Claude skills and plugins for workflows the whole team can reuse

A skill is a folder of instructions, sometimes with scripts and templates, that Claude loads only when a task calls for it. A connector gives Claude access to a system, and a skill teaches Claude how you do a task (Anthropic). A plugin bundles skills, connectors and commands into one install you can share. On Team and Enterprise, owners can provision skills for everyone.

How to create a skill

1. Create a folder called job-ad with a file named skill.md.
2. Start the file with a name of up to 64 characters and a description of up to 200, which Claude uses to decide when the skill applies.
3. Write the instructions below, ideally with an example of a good result.
4. Zip the folder and upload it under Customize > Skills, then + and Create skill, then Upload a skill.
5. Test it with a few requests. If Claude skips the skill when it should use it, rewrite the description (Anthropic).

---
name: job-ad
description: Writes job ads in our template and tone. Use when someone asks for a job ad or a posting for a role.
---
# Job ad
Ask for the role profile if it's missing. Never guess pay, benefits, hours or contract type.
1. Job title, then two sentences on why the role exists.
2. What you'll do, no more than six bullets.
3. Requirements and nice-to-haves, as things the person can do, not personality traits.
4. What we offer, only what's in the material.
5. How to apply. Write [FILL IN] if the closing date is missing.
Write "you" to the applicant and "we" for the organization, in no more than 350 words.

On a Mac, on Pro, Max, Ultra or Team but not Enterprise, you can also record yourself doing a task in Cowork and let Claude propose the skill. Anthropic’s HR plugin has commands for offer letters, onboarding, performance reviews and policy lookup (Anthropic on GitHub). Read its instructions and adjust them to your own policies before you share it.

What are the limits and risks of Claude Cowork?

In HR, the risk I’d spend the most time on is a polished document that nobody checked. The error shows up only after someone has decided on the wrong number.

• Errors that look right. A spreadsheet can carry a wrong formula, and a theme can rest on three answers out of a hundred. Ask for the source of every number and check a few yourself.
• Hidden instructions. Emails, web pages and outside documents can carry text that tries to steer Claude, known as prompt injection. Anthropic says the chance of an attack is still non-zero. Don’t let a task that reads unknown emails also send or change things.
• Screen control. Claude asks before it uses each app, but Anthropic’s safety guide says there is no sandbox between Claude and your screen. Block apps like payroll before you let it control the computer.
• Deleted files. Claude can delete files in connected folders, though only with your explicit permission (Anthropic). Work in a dedicated folder and keep a backup.
• Usage that adds up. On Team, owners can let people work past the seat limit with usage credits billed at API rates, and on Enterprise all usage is billed that way. Set spend limits for the organization and per member before a team rollout (Anthropic).

What to agree on before a team rollout

Your colleagues may already use Claude. In a Swedish survey of HR professionals, 29% use Claude at work, and about half of them (48 of 94) say their employer pays for it. Cowork isn’t on the free plan, so anyone using it without a company plan is on a personal plan, where training is on unless they turn it off. Find out which Claude accounts people use before anyone connects people data. And since 70% in the same survey don’t measure the effect of AI in HR, decide how you’ll measure before you start.

Source: AI & HR Barometern 2026, humAIn Academy. 329 HR professionals in Sweden who chose to take part, 8 June to 24 September 2026, probably more interested in AI than HR overall. It describes Swedish HR respondents, not US or UK workplaces, and the pay gap is not a measure of shadow AI.

Start with what you want Cowork to take off the team’s plate and which data that needs. The other decisions follow from that.

If people already use Claude on personal accounts, read Shadow AI before you write any rules. How to measure adoption more broadly is in AI enablement.

Claude Cowork or Copilot Cowork?

Microsoft has its own Cowork in Copilot for the same kind of work inside Microsoft 365. It needs a Microsoft 365 Copilot license with Cowork and usage-based billing turned on, and it asks before it sends email, posts in Teams or schedules meetings (Microsoft Learn). It can use Anthropic’s models, which sit outside the EU Data Boundary and are off by default in the EU, EFTA and UK (Microsoft Learn).

If you already pay for Microsoft 365 Copilot and your files live in SharePoint and OneDrive, I’d test Copilot Cowork first, since it works where your files already are, and ask IT whether Anthropic’s models are on for you. Claude Cowork is worth a look if your work involves files on your own computer or you don’t have Copilot licenses.

Your first Cowork task

Hand over one real task from your own week, without personal data, like the handbook check in example 1. If AI tools are new to you, my free AI course for HR takes about 60 minutes. In the Cowork Bootcamp, a small group of HR people works on their own tasks over three live 60-minute sessions and saves the workflows as skills, for $590 excluding VAT plus a Claude Pro subscription. Pro is a personal plan, so you practice on material without personal data. Ask me about the next cohort through the form below.

Frequently asked questions

Is Claude Cowork free?

No. Cowork is included in every paid Claude plan but not in the free plan. As of November 2026, Pro costs $17 a month and a Team seat $20 a month, both billed annually and before tax. Enterprise costs $20 per seat plus usage at API rates.

Where can I use Claude Cowork?

In the Claude desktop app for macOS and Windows, with Linux in beta, and on claude.ai, in the mobile apps and in the Chrome side panel on paid plans. Enterprise owners decide whether web and mobile are on. To work with files on your own computer you need the desktop app.

What is the difference between Claude Cowork and Claude chat?

In chat Claude handles one request at a time, including single files. In Cowork you describe an outcome, and Claude plans the steps, works with your files and apps and returns finished files. In accounts with the new layout the two are merged, and Claude uses Cowork’s tools when a task needs them.

Is Claude Cowork an AI agent?

Yes. It carries out steps toward a goal with the access you give it. Unlike agents built into business systems, you set up each task yourself, see the plan and choose whether Claude asks before each action.

Is Claude Cowork safe for company data?

It can be, on Team or Enterprise, where content isn’t used for model training by default and Anthropic’s Data Processing Addendum is part of the terms. Data is stored in the US. Use manual approval, one working folder per task and a written list of data that never goes in.

What is the difference between Claude Cowork and Claude Code?

Both run on the same agent technology for multi-step tasks. Claude Code is built for software engineering. Cowork is built for other knowledge work, such as documents, spreadsheets and research, and needs no terminal.

Sources

• Anthropic: Get started with Claude Cowork
• Anthropic: Cowork on Team and Enterprise plans
• Anthropic: Use Claude Cowork safely
• Anthropic: Claude Cowork and chat are one Claude
• Anthropic: Schedule recurring tasks in Cowork
• Anthropic: Claude Cowork product page
• Anthropic: Plans and pricing
• Anthropic: Usage credits and spend limits on Team and Enterprise
• Anthropic: Cowork monitoring
• Anthropic: Data retention on consumer plans
• Anthropic: Change your model training setting
• Anthropic: Where data is stored
• Anthropic: Data Processing Addendum
• Anthropic: Claude Desktop on third-party platforms
• Anthropic: Features on third-party platforms
• Anthropic: MCP and the Agentic AI Foundation
• Anthropic: Get started with connectors
• Anthropic: Microsoft 365 connector
• Anthropic: Custom connectors with remote MCP
• Anthropic: What are skills?
• Anthropic: Use skills in Claude
• Anthropic: How to create custom skills
• Anthropic: Connectors, skills and plugins
• Anthropic on GitHub: HR plugin
• Microsoft Learn: Get started with Copilot Cowork
• Microsoft Learn: Anthropic models in Microsoft Online Services
• EUR-Lex: AI omnibus, Regulation (EU) 2026/1744
• NYC Department of Consumer and Worker Protection: Automated employment decision tools
• State of AI in HR (AI & HR Barometern 2026, humAIn Academy)