This year Anthropic, Microsoft and OpenAI each shipped a second mode beside the chat box. Anthropic and Microsoft both called it Cowork; OpenAI called it Work. If you have only ever used the chat side, this is a plain explanation of what the other side does, and why my initial view is that it matters more for a regulated firm than the next model upgrade will.
Chat answers. Cowork does.
Chat is what most people mean by "using AI". You type a question and get an answer. Moving that answer into the work, the spreadsheet or the client file, is your job. Chat can already reach connected tools in all three products, so this is not a hard line, but in chat the output is a reply, and you are the one who carries it into the work.
A work mode is built for handing over a task rather than asking a question. You point it at the files and systems the task needs and it goes and does the work there: reading the folder, drafting the document, opening the portal, then coming back with the finished item for you to check. Anthropic's instruction for it is "tell Claude what you need, not how". Microsoft describes Copilot Cowork as returning "a completed result, not just a draft or a recommendation". OpenAI says ChatGPT Work can "stay with a project for hours if needed, and turn a goal into finished work".
All three present the difference as a switch. In Claude, Cowork sits next to Chat. The Microsoft 365 Copilot app has a toggle between Chat and Cowork. In ChatGPT, Work is a mode you move into from the ordinary conversation. The difference is less about the model than about autonomy: how far the AI is allowed to reach into your systems on its own, and how long it is allowed to keep going before it checks in.
Four things change when you flip the switch
A short glossary first, because the vendors use overlapping words.
- Connector: a permissioned link to one system, such as your document store or your mailbox. OpenAI calls these apps.
- Skill: a written description of how your firm does a particular job.
- Plugin: a package that installs connectors or skills, or both, in one go.
- It reaches your systems. Connectors give it permissioned access to the document store, the mailbox, the calendar and the CRM. Chat mostly works from what you give it in the conversation. A work mode is built to go and get what it needs from the systems you have connected.
- It has a browser. Each of the three can now open web pages itself, so a task that runs through a portal with no connector can still get done. Where that browser runs and what it must ask you before clicking differ by product, which matters when the portal belongs to a client or a regulator.
- It runs long, and unattended. Claude Cowork tasks can be scheduled: "every Monday morning, build last week's marketing readout" is Anthropic's own example, and its product page puts it as "close your laptop, it keeps going". Copilot Cowork has scheduled prompts and event-driven tasks that fire when a matching email or Teams message arrives. ChatGPT Work has scheduled tasks too. All three cap how many you can run and how often.
- It carries your firm's method. This is the one most firms have not thought about yet. A skill for the year-end workpaper or the renewal comparison tells the AI how your firm does that job, so it has your method in front of it rather than a generic one. All three vendors use the word. In Claude and in Copilot Cowork a custom skill is a plain markdown file, and Anthropic and OpenAI both package skills into plugins.
The three products, one at a time
Everything below is taken from the vendors' own announcements and product pages as at 8 September 2026, and it will date.
Claude Cowork (Anthropic)
Research preview in January, general availability in April, on the paid plans. It runs on desktop, web and mobile. You choose which local folders it can see. Connectors cover Microsoft 365, Google Drive, Slack and a long list of third parties, and a built-in browser handles sites without one. Plugins bundle skills and connectors into a single install. Actions run under one of three approval modes, from approving each one by hand through to skipping approvals altogether, and Anthropic's own guidance is to stay on manual approval for high-stakes work. Enterprise admins control access and see usage from organisation settings.
Copilot Cowork (Microsoft)
Announced in March, general availability in June. Microsoft's words: "Working closely with Anthropic, we have brought the technology that powers Claude Cowork into Microsoft 365 Copilot." At launch it runs on Anthropic models, with a Microsoft-built model to follow, and it can pick the model to suit the task. It needs a Microsoft 365 Copilot licence and is then billed by usage, in units Microsoft calls Copilot Credits. It is off by default. Admins get spend limits for the whole firm, a group or a single user, plus audit logging. It grounds its work in your firm's own Microsoft 365 data (Microsoft's name for this is Work IQ), browses through a local Edge browser under your existing policies, and shipped with nine partner plugins. Custom skills are markdown files in a OneDrive folder, written by hand or drafted with Cowork through a guided flow, and can be shared with named people in the firm.
ChatGPT Work (OpenAI)
Launched on 9 July. An agent inside ChatGPT that "can take action across your apps and files" and turns a goal into documents, spreadsheets and presentations. Plugins connect it to Slack, Teams, Google Drive, SharePoint, email, calendars and CRMs, and on launch day OpenAI folded its app directory into a plugin directory. Workspace admins manage which plugins are installed and what each underlying app may access. It is available on every plan in the desktop app, and on the paid plans on web and mobile.
If you are staying in the Microsoft camp
Microsoft gives you several ways to build an agent: Agent Builder inside the Copilot app, Copilot Studio for low-code agents with their own workflows, Microsoft Foundry for developers on Azure, and Agent 365 to govern the lot. Each has its own licence and its own learning curve.
If your firm is staying with Microsoft, our recommendation is to head for Copilot Cowork and put the effort into skills. A Cowork skill is a markdown file describing how your firm does a job, one a partner can read and correct, and it runs with the permissions the person already has. A Copilot Studio agent is a canvas of boxes that needs its own access set up. Studio still makes sense for a customer-facing agent or a fixed process that runs with nobody watching. For work a person reviews anyway, Cowork is the direction of travel, and it is the product Microsoft chose to build on Anthropic's technology.
What this means for a regulated firm
The benefit is that the work gets done where it lives, by a method you wrote down. A year-end file or a renewal pack was never one question. It is a long run of steps across several systems, with a firm-specific way of doing most of them, and a chat thread is a poor place to hold that. A work mode with a skill can hold it, and the skill improves every time somebody at the firm fixes it. A skill is instructions, not a guarantee, so the output still gets checked; the difference is that the checking gets easier each time. That is the loop we keep going on about, now with a product to run it in.
The caution is that an AI which can act on your systems needs the same controls as a person who can. The controls differ by product and plan. Claude has approval modes and folder scoping, and also a setting that turns approvals off. Microsoft ships Cowork off by default, with spend limits and audit logging. OpenAI puts plugin permissions in the hands of workspace admins. Before you pick, compare the approval settings, what the agent can reach, what gets logged and what it can spend. Anthropic is also explicit that a document the agent reads can carry instructions it might follow, which is what prompt injection means in practice, and recommends manual approval and trusted sites for anything that matters.
When something goes wrong, the question I expect from a regulator, or from your own board, is not whether you used AI. It is what the AI did, how it was checked and who signed off. A work mode brings that question forward, because the AI is now expected to do rather than suggest.
Which one?
The one your firm already pays for is the likely default; that is what being bundled buys. We implement Claude for a living and write skills for it, so I am inclined towards Claude Cowork. The practical case is that it is the original and it runs outside any one vendor's office suite. Microsoft looked at the same problem and chose to bring Anthropic's technology into its own product rather than build its own, which is a fair signal of where the state of the art sat in March.
All three work without a skill written by your firm. A skill becomes worth writing when you want the job done your firm's way each time it runs, whoever runs it. The tools will help you draft one. The knowledge of how your firm actually does the job has to come from your firm.