Microsoft 365 outage: your AI agent is only as good as its inbox

Claude was up. ChatGPT was up. And on Thursday evening, for a lot of people, neither of them could read a single email. The problem wasn't at OpenAI or Anthropic. It was at Microsoft. Ever had a backup plan that turned out to run on the same power cord?

What happened

At 1:26 PM ET on November 12, a Microsoft 365 incident began, according to BleepingComputer. Microsoft confirmed impact to Exchange Online, Teams, SharePoint Online, Microsoft 365 Copilot and Defender XDR. It blamed servers that went unexpectedly unavailable and strained the rest of the infrastructure, paused all deployments, and looked at whether reverting a few recent configuration changes would speed recovery. Per dpa, the outage was still running Friday morning, roughly ten hours in; Downdetector logged more than 130,000 reports in the US alone at the peak, and PC-Welt counts 74 affected countries.

The part almost nobody is telling: at 22:23 UTC, Anthropic said the mail, calendar and search tools in Claude's Microsoft 365 connector were failing "for most requests" and that Claude for Outlook was affected, all caused by Microsoft's incident. OpenAI posted elevated errors on its own Microsoft 365 connector the same day.

Multi-vendor doesn't save you if everyone drinks from the same well

A lot of teams comfort themselves with a rule: we're not locked into one AI vendor, we use Copilot, ChatGPT and Claude. Sounds like diversification. It isn't. The agent is swappable; the inbox isn't. If every workflow makes the AI read mail, calendar and SharePoint first, you have exactly one single point of failure, and its name isn't OpenAI.

To me, this is the logical consequence of something I argue a lot: treat AI as an operating system with access to your data and processes. But an operating system needs a plan for when the disk dies. So the real question isn't "which model?" It's "which data source has to be up for anything to work at all?"

The catch: connectors are the thinnest wire

This isn't just a Microsoft problem. Anthropic's status page lists two more connector outages from late October alone: the Gmail connector threw errors for about six and a half hours on October 27–28, the Google Calendar connector for a little over two days from October 26 to 28. The model itself wasn't the issue either time. In my view, the link between AI and your data is the most underrated risk in the stack right now, because it never shows up in a tool evaluation or a pitch deck.

For marketing, that's concrete: if your briefs, tone of voice and approvals live only in email threads and SharePoint folders, an outage doesn't just take your inbox. It takes the team's memory. My rule of thumb: your own AI project is the control room, and the control room shouldn't sit in the same building as the mailroom.

What I'd do now

• List every AI workflow that reads mail, calendar or files. Next to each: what happens if that source is gone for ten hours?
• Get brand context out of the inbox: positioning, tone, no-gos and live briefs as documents inside the AI project, not just as an email attachment.
• For the two or three critical flows (lead replies, campaign approvals), decide who takes over by hand, and subscribe to your vendors' status pages.

Bottom lineThree AI vendors aren't redundancy as long as all three have to walk through the same door.