Copilot Cowork is now generally available in Microsoft 365, and it’s one of the AI developments I’ve been most excited about this year. It brings a way of working that’s already changed how I use AI into the environment where many of my clients run their businesses.

If you’ve mainly used AI to ask questions or help write things, Cowork is worth understanding. You give it a piece of work to complete, and it can work out the steps, use the tools available to it and produce the outputs. With the appropriate connections and permissions, it can work across your files and business systems, carrying a task forward without needing you to prompt every individual step.

This is what people mean by “agentic” AI. The model can take actions towards a goal, checking what happens and deciding what to do next.

The software that makes this possible is often called a “harness”. It surrounds the AI model with tools, instructions and access to information, along with controls on what it can do. Software engineers have been using this approach through coding agents. Cowork puts it into a form aimed at knowledge workers: people whose days revolve around documents, information and decisions.

When Anthropic launched Claude Cowork in January 2026, I was blown away by it. Enough to switch from ChatGPT to Claude for my own work.

One early example was preparing a fictional business for a training case study. I explained what I needed, and Cowork researched the context, invented a local brewery, created staff profiles and mocked up a website. It produced multiple assets from that overall brief without further direction. I could see how much further I could take this way of working.

Naturally, I wanted to bring it into my client work. But those clients worked in Microsoft 365, and sending their information into a separate Claude environment wasn’t an option for those engagements.

So I was happily surprised when I learned in March that Microsoft was bringing the technology behind Claude Cowork into Copilot. Given Microsoft’s close association with OpenAI, seeing Anthropic’s technology taking such a prominent role caught my attention.

Copilot Cowork initially came through Microsoft’s Frontier early-access programme, then became generally available on 16th June, using Anthropic models including Sonnet and Opus. The capability I’d been enjoying in my own work now had a route into my clients’ Microsoft environment, under their organisation’s access and governance controls.

The practical benefits can be substantial. In one client workflow I’ve worked on, preparing a document from meeting notes and supporting material used to take two to four hours. With Cowork, a draft roughly 90–95% of the way there came back in about five minutes, followed by another ten or fifteen minutes of human checking and editing.

We also saw improvements in quality through more thorough use of the supporting material. The workflow had been built around the firm’s standards and templates, so it knew what the finished document needed to contain.

Getting that result took some preparation. Cowork still needs a clear brief, access to the right information and someone with the expertise to check and sign off the result. General availability also doesn’t mean it’s automatically switched on: administrators need to enable it, and usage carries additional charges.

For a small firm, saving a skilled person a couple of hours on a recurring task creates room for more client work and attention to the decisions that really do need their expertise. For firms already using Microsoft 365, being able to do this within their existing systems makes it much more practical.

If you haven’t tried Cowork yet, I’d recommend giving it a go. Pick a real piece of work you know well, explain what a good result would look like, and see how far it gets. That’s when the possibilities start to become clear.

Microsoft has a couple of useful videos showing it in action:

  • Meet Copilot Cowork — an introduction to delegating work across Microsoft 365.
  • Copilot Cowork demonstration — Charles Lamanna demonstrates reviewing monthly one-to-ones, summarising customer meetings and producing a competitive analysis.