For most of the past year, using AI for real work meant a lot of copy-paste. You asked the assistant to draft the follow-up email, it wrote a good one, and then you pasted it into Outlook yourself. Two updates Anthropic shipped in early July close that gap: Claude can now write directly inside your Microsoft 365, and its work sessions follow you from laptop to phone. The practical effect is that the assistant starts finishing tasks instead of just handing them back.
What actually changed
The Microsoft 365 connector used to only read and search your data. As of July, with write tools switched on, Claude can draft and send email, organize your inbox, add and change calendar events, and create or update files in OneDrive and SharePoint. Separately, Claude Cowork — the mode where it runs longer, multi-step jobs — expanded from desktop-only to web and mobile. Sessions now save to your account and run remotely, so a task can keep working after you close your laptop, and you can check it from your phone.
Where this earns its keep
The wins are the boring, repeatable handoffs. “Draft replies to these six client emails and save them to my drafts.” “Pull the numbers from this SharePoint folder into a summary doc.” “Block two hours Thursday and send the three subcontractors a scheduling note.” These are tasks where the thinking was never the hard part — the clicking was. Letting the AI do the clicking inside the tools you already run is where an hour a day quietly comes back.
The permission step you should not skip
Here is the part that matters and the part vendors gloss over: write access does not turn on by itself. A Microsoft admin has to consent to the new permissions and enable them for your organization. That is a feature, not a hassle. An assistant that can send email and change files on your behalf is powerful precisely because it can send email and change files on your behalf. Turn it on deliberately, decide who gets it, and keep a human eye on anything that goes to a client or moves money-adjacent data.
The honest caveat
“Claude can send email” does not mean “let Claude send email unsupervised.” For a first month, keep it drafting into your outbox rather than sending directly, and read what it produces before it goes out. The technology is reliable enough to save real time and not so reliable that you want it firing off client messages while you are not looking.
The takeaway
If you live in Microsoft 365, this is worth a look this week: have your admin review the write-tool permissions, enable them for a small group first, and pick two recurring tasks to hand over. Start with drafts, confirm the quality, then widen the leash. The point is not that AI can now touch your tools — it is that you get to decide exactly how far.