Microsoft 365 Went Down for Half a Day. Your Continuity Plan Probably Assumed It Would Not
• 3 min read • Ahmed Daoudi
On the afternoon of November 12, 2026, a large part of the working world lost email, chat and its AI assistant at the same time. Microsoft 365 services including Outlook, Teams and Copilot started failing around 16:17 GMT, and parts of the incident were still open the next morning.
What we know so far
- Downdetector reports for Outlook peaked at roughly 129,000 shortly after 1 p.m. US Eastern time, and users in at least 74 countries were affected according to tracking by NeuralTrust.
- Microsoft said early on that this was not a cyberattack. It first described the cause as elevated load across several regions.
- It then clarified that servers were failing automated health checks and becoming unavailable, and the traffic they dropped overloaded the servers still running.
- Later statements pointed to settings applied to servers after an update, with a process that sets up secure connections becoming overwhelmed. A full root-cause report had not been published at the time of writing.
The failure pattern engineers should recognize
This looks like a classic cascading failure. A change makes some servers unhealthy. The load balancer removes them. Their traffic moves to the remaining servers, which now handle more than they were sized for, start failing health checks too, and get removed in turn. Retrying clients make it worse by sending more requests at the moment the system can least handle them.
None of this is unique to Microsoft. The same pattern takes down small APIs every week. The defenses are well known and often skipped:
- Roll out configuration in stages, the same way you roll out code, with automatic rollback when error rates rise.
- Use exponential backoff with jitter on clients, so retries spread out instead of arriving as a wave.
- Shed load deliberately. Rejecting some requests quickly is better than accepting all of them and failing slowly.
- Make health checks honest but not twitchy. A server that is slow under load is not the same as a dead one, and removing it may make things worse.
- Use circuit breakers between services so one failing dependency does not drag down everything that calls it.
The AI dependency nobody mapped
One detail stood out to me. Copilot for work stopped answering because it depends on the same Exchange infrastructure as email. Many companies have spent 2026 wiring AI assistants into daily workflows without asking what those assistants depend on. When the mail system fell over, the "AI layer" fell with it.
What to put in your continuity plan
- A backup communication channel that does not share infrastructure with your main one, agreed before you need it.
- A dependency map for critical workflows, including the AI tools built on top of them.
- A short list of who decides what during an outage, so people are not waiting for a Teams message that cannot arrive.
Outages at this scale are rare, but they are certain to happen again. The useful question is not whether a vendor this big can fail. It is which of your processes stop when it does.