The appealing part of an AI coworker is not having another conversation. It is coming back to a useful draft, an organized folder or a report that no longer exists exclusively as fourteen scattered notes.
Anthropic’s Claude Cowork and Microsoft’s Copilot Cowork are pursuing that idea in related but distinct products. The latest news is Microsoft’s March 30 opening of Copilot Cowork through its Frontier program. Claude Cowork itself has been around since January. This was not a simultaneous launch.
Claude Cowork Came First
Anthropic’s dated release notes record the initial research preview on January 12 for Max subscribers using its Mac desktop app. Pro access followed on January 16, and scheduled tasks arrived on February 25.
The underlying idea brings Claude Code’s ability to act on files into work beyond programming. Instead of repeatedly copying material into a chat and moving the answer back out, you give the system access to the relevant working material and ask it to produce something.
That changes the shape of a request. “Explain how to organize these notes” becomes “make a draft from these notes.” The result still needs review, but the handoff can involve an actual file rather than another block of text waiting to be pasted somewhere.
Think of the promise as less digital small talk and more help with the folder named final_final_really_final. Most of us have met that folder. Some of us have founded a small civilization inside it.
What Microsoft Announced in March
Microsoft had already introduced Copilot Cowork on March 9, describing a limited research preview that combines Anthropic’s approach to multi-step tasks with Microsoft 365. Microsoft’s identity, security and governance systems remain part of that product.
The March 30 announcement makes it available through Frontier. Microsoft describes work that progresses across tools and files, with a visible plan and opportunities for the user to steer. It cites Capital Group’s early experience with planning and preparing deliverables.
Those are useful examples of the intended work, not an independent measurement of how much time every customer will save. And this is access through a preview program, rather than a statement that every Office installation suddenly contains the same new coworker.
Researcher’s Critique Is a Separate Feature
Microsoft’s same March 30 post also introduces Critique in Researcher. One model prepares an initial response and another reviews it, using models from providers including OpenAI and Anthropic. Council offers a separate way to compare model responses.
That is not a description of every Copilot Cowork task running through a fixed GPT-writes, Claude-checks pipeline. It is a different feature within the wider Copilot offering.
Nor does a second model turn a report into verified truth. Agreement is helpful only when the underlying evidence supports the answer. Two confident colleagues can still share the same mistaken assumption; adding a third chair does not automatically fix the spreadsheet.
It is another angle on the competition between AI tools: choosing a model and designing a useful workflow are related decisions, but they are not identical.
Delegating Work Still Requires Boundaries
There is a practical difference between an incorrect answer and an incorrect action on a file. Start with copies of material you can afford to reorganize, describe the expected output, and be explicit about what must stay untouched.
Anthropic’s March 23 computer-use announcement acknowledges mistakes and prompt-injection risks, describes permission controls, and recommends avoiding sensitive data while testing that early capability. It also explains Dispatch, which lets a user assign work from a phone while the desktop application does the work.
For that desktop workflow, the machine must remain awake and the application running. A recurring task is not proof of an invisible cloud employee working after you shut the laptop.
The attraction is understandable: less shuffling between files, more finished work. The test is equally straightforward. Does the result hold up when you inspect it, and did the system stay within the task you assigned? A useful coworker earns trust through the work. A convincing job title is only the introduction.
