Microsoft 365 outage disrupts Claude and ChatGPT connectors Microsoft 365 outage disrupts Claude and ChatGPT connectors

The connector becomes the failure boundary

Anthropic first reported slow responses and errors at 17:52 UTC. Its notice covered requests to mail, SharePoint, Teams chat and calendar, plus problems establishing new connections. At 18:34 UTC, it identified an upstream-provider issue. The 20:28 UTC update said most mail, calendar and search requests were still failing. It also described failed mail and calendar actions in Claude for Outlook, with some users unable to open the add-in. The incident was assigned to Claude's web product and Cowork. Anthropic's status page separately listed the Claude API and Claude Code as operational. That distinction matters for applications that combine a working model with external business data. At 21:18 UTC, Anthropic said its upstream provider was restarting affected servers while these failures continued.

A second assistant reports the same dependency

OpenAI opened its own Microsoft connector incident, identifying ChatGPT's Connectors/Apps component. Its notice described broken new connections and tools behaving incorrectly. The report classified the problem as degraded performance and remained under investigation. OpenAI also explains that its availability figures aggregate tiers, models and error types, so individual customers can have different experiences. The two incident reports identify a shared Microsoft integration problem from separate assistant providers. A successful ordinary chat therefore answers a different operational question from whether an assistant can fetch a mailbox or complete a calendar action.

Office users see the underlying interruption

The disruption also reached people using office applications directly. UMass Amherst's November 12 service notice warned its campus community about Teams, Outlook, SharePoint and Office web apps. Old Dominion University described email and calendar access failures across desktop, browser and mobile clients. Its notice said outgoing messages could remain in Outbox or Drafts, while incoming mail was queued for later delivery. It advised using alternative communication channels for urgent matters and warned against suspicious support messages seeking login details. Those are concrete consequences outside the assistant interface: a workflow can finish composing a message while its delivery remains pending.

Service restoration is a separate checkpoint

The Associated Press reported that Microsoft traced the disruption to servers becoming unavailable after failing automated checks. Microsoft was testing a fix at the time of that report. The incident separates several states that an automation dashboard can otherwise collapse into one success indicator: generating content, reaching the connected service, and completing the requested action. The university notice's queued incoming mail describes a delivery mechanism; the assistant notices describe failed tool requests. Recovery of one does not by itself document completion of every interrupted action. For operators, the useful evidence is the action's actual result in the destination service.