What took Microsoft 365 down for a day, and what to build before the next one

Starting around midday Eastern on Thursday 12 November 2026, people could not open their mailboxes. Exchange Online failed first in the Americas and then in other regions, and the failure spread to Outlook, Teams, SharePoint, Microsoft 365 Copilot and Defender XDR, SecurityWeek reported. Downdetector reports peaked at around 130,000 to 140,000. Microsoft tracked it as incidents EX1550543 and MO1550821, and on Saturday 14 November it said it had restored service to the remaining customers and regions, CNBC reported. Microsoft’s preliminary explanation is a network change.

The outage ran well past 12 hours. That figure came from early live coverage, when the disruption had crossed the 12-hour mark. Reports put it at well over 24 hours, and the all-clear for the last customers and regions came on Saturday 14 November.

The short answer to “what happened” is Microsoft’s own, and it is preliminary. The company said the incident “was introduced by an issue during an earlier network capacity upgrade” that “surfaced following a recent network change”, KUOW reported. In a post on X it added that “this was a technical issue not related to cyber or AI / agents”, Yahoo Finance reported. A full post-incident report was promised and had not been published when I wrote this, on 15 November. I found no Microsoft statement on layoffs.

I run delivery at Crimson Umbrella, so I spend my week on the unglamorous side of software: releases, change review, runbooks and the plans for when something breaks anyway. Here is what is known, what isn’t, and what I’d build.

What happened on 12 and 13 November

Microsoft’s explanation changed as the day went on, which is normal for a large incident and worth knowing when you read early updates. Its first status posts blamed elevated load across several regions. Later it said the load did not come from more traffic: servers had unexpectedly become unavailable, failed automated health checks and dropped out, and the rest absorbed their load until they struggled too, NBC Chicago reported. By late afternoon Eastern it said a core process that establishes secure connections was becoming overwhelmed.

The fixes followed the diagnosis. Microsoft redistributed traffic, paused deployments, restarted servers and isolated part of its infrastructure. At first it said no recent change lined up with the start, then it said a recent network configuration change did, and began rolling it back, SecurityWeek reported. Recovery came region by region over Friday. Microsoft apologised (“We are deeply sorry for the disruption this incident has caused our customers and partners”), and extended the window in which delayed email is retried to 48 hours, so messages queued rather than bounced.

The pattern here, one change made earlier, a later change that exposes it, and health checks that turn a partial failure into a total one, is a common way big services fall over. It is also the least exciting explanation, which is why the internet went looking for others.

Was it the layoffs?

There is no evidence that it was, and no evidence that it wasn’t. Microsoft did not address staffing in any statement I could find, and Charles Lamanna, the executive who gave the company’s statement, did not address it either, Yahoo Finance reported.

The facts people point to are real. Microsoft cut about 9,000 jobs in July 2025, NBC News reported, and about 4,800 more, 2.1% of its workforce, in July 2026, mostly in Xbox and commercial roles. Its chief people officer wrote at the time that “the roles eliminated today are not being replaced by AI”. On 26 October 2026 it offered voluntary buyouts to about 60,000 eligible employees, roughly 27% of its workforce, Newsweek reported, and stressed that the offer does not mean 60,000 people are leaving.

One detail cuts against the theory. The buyout departures were reported to take effect on 31 December, so on 12 November the people who might accept were still at their desks. A link to the July cuts is possible in principle, through lost knowledge of an old system, but nobody has shown it. Until the post-incident report names the change and who reviewed it, “the layoffs did it” is a guess.

Was it unmonitored AI?

Microsoft says no, in so many words: “not related to cyber or AI / agents”. That is the company’s claim, not an independent finding, and the final report may add detail. But the stated cause is a network capacity change, which is exactly the kind of failure that took services down long before AI wrote any code.

The question is still a fair one, because AI is now part of how software gets made. Satya Nadella said in April 2025 that software wrote up to 30% of the code in Microsoft’s repositories, as SiliconANGLE reported. Google’s DORA research for 2025 found 90% of software professionals use AI, and described it as a mirror and a multiplier of the team using it. A team with good change control gets faster; a team with weak change control gets weak change control at a faster pace.

So the useful question isn’t whether an AI broke Microsoft. It is whether your own changes, written by a person or a model, reach production faster than anyone verifies them. We wrote about that in what happens when agents open 40% of pull requests: the bottleneck moves from writing code to checking it.

A pattern across six big outages

Read what the companies published side by side and the same few failures repeat:

Outage What the company published
CrowdStrike, 19 July 2024 A content validator logic error and a missing bounds check let a bad content file ship with no staged rollout (root cause analysis)
Google Cloud, 12 June 2025 A policy change with blank fields hit a new code path that had no error handling and no feature flag (incident report)
AWS us-east-1, 19 and 20 October 2025 A latent race condition in DynamoDB’s DNS automation left an empty regional DNS record (AWS summary)
Azure Front Door, 29 October 2025 Valid configuration changes across two builds produced incompatible metadata that exposed a dormant bug (Microsoft review)
Cloudflare, 18 November 2025 A database permissions change doubled a feature file past a hard limit of 200 entries (Cloudflare’s post)
Microsoft 365, 22 and 23 January 2026 Microsoft said part of North American infrastructure was not processing traffic as expected, for almost ten hours, with no deeper cause published (The Register)

Almost none of these were attacks, and almost none were exotic. They were changes, usually to configuration, that passed every check the company had and failed in production because production was the first place the bad combination existed. Uptime Institute’s 2026 annual outage analysis reached the same place from the other direction: configuration and change management drive network outages, and failure to follow procedures leads the human-error ones.

The fixes the companies published rhyme too. CrowdStrike promised staged deployment of the content that had gone everywhere at once. Google committed to feature flags for that kind of change. AWS disabled the automation worldwide until it was safe. Cloudflare listed more global kill switches among its repair items.

What to build so one change can’t take you down

None of this needs Microsoft’s budget. These controls scale down to a team of four, and they are what we put in place when we build or take over a system.

  • Roll out in stages. Every change, code or configuration, goes to a small slice of traffic first, with an automatic stop if error rates or latency move. A bad change that reaches 1% of users is an incident report, not a headline.
  • Put risky behaviour behind a flag. New code paths ship switched off and turn on gradually, and every flag doubles as a kill switch someone can flip in seconds without a deploy.
  • Treat configuration as code. Config changes get review, validation against limits and tests, because most of the outages above began with a configuration change that nothing checked.
  • Make health checks fail safe. When servers that fail a check are removed, the survivors take their load. Cap how much capacity automation can remove at once, shed load before it cascades, and alert a person before the last healthy nodes go.
  • Alert on what users feel. Watch sign-in success, mail delivery and request latency as well as CPU and memory, so you hear about trouble from your dashboards before you hear about it from customers.
  • Write the runbook before the night you need it. Who decides, how to roll back, how to talk to customers, and practise it with a game day twice a year.
  • Hold AI changes to the same bar. A change from a model goes through the same review, tests and staged rollout as a change from a person, with no shortcut because the diff looked clean.

If your company runs on Microsoft 365

You can’t fix Microsoft’s network, and moving off it is rarely worth it. A 99.9% uptime commitment still allows about 43 minutes of downtime a month, Yahoo Finance noted, and the January and November incidents both ran far past that. What you can do is decide in advance what happens when email and Teams are gone for a day.

Pick a second channel for urgent communication that doesn’t depend on Microsoft 365, and make sure everyone knows it before they need it. Subscribe the people who need to know to Microsoft’s service health notifications. Keep offline copies of the few documents you’d need in a crisis, such as contacts, contracts and the runbook itself. If your own systems send alerts, invoices or password resets through Exchange, give them a fallback sender, because a payments failure caused by someone else’s outage is still your failure to your customers.

How we help

Most of our work is building software, so resilience is part of the job rather than an add-on. When we build a service, every release is tagged and rolls back with one command, and we write the runbook alongside the code. When we take over a system, the controls above are the first things we check, and faults in what we shipped are ours to fix under your contract. Rujen Shakya led the ISO 27001 and SOC 2 programme at a 35-person AI SaaS company and passed both on the first attempt. That’s his record and that company’s certification, not ours: we aren’t audited, but we bring someone who has run both programmes.

If you want a second pair of eyes on how your own changes reach production, an architecture review starts at $3,000 and ends with a written report. You can read how we work, or call or message either of our offices: between New York and Kathmandu, someone picks up around the clock. The quickest start is a twenty-minute call.

Questions people are asking

What caused the Microsoft 365 outage on 12 November 2026?

Microsoft said the outage “was introduced by an issue during an earlier network capacity upgrade” that “surfaced following a recent network change”, and that it was not caused by a cyberattack or by AI or agents. Exchange Online servers failed health checks and dropped out, overloading the rest. Microsoft promised a full post-incident report.

How long was the Microsoft 365 outage?

Longer than the 12 hours first reported. The outage began around midday Eastern on Thursday 12 November 2026, reports put it at well over 24 hours, and Microsoft said on Saturday 14 November that it had restored service to the remaining customers and regions.

Which Microsoft 365 services were affected?

Exchange Online and Outlook failed first, followed by Teams, SharePoint Online, Microsoft 365 Copilot and Defender XDR. The outage hit the Americas first and then other regions. Downdetector reports peaked at around 130,000 to 140,000.

Did Microsoft layoffs cause the outage?

There is no evidence either way. Microsoft has not addressed layoffs in its statements about the outage. It cut about 4,800 jobs in July 2026 and offered voluntary buyouts to about 60,000 employees in October, but the buyout departures were reported to take effect on 31 December, after the outage.

Did AI cause the Microsoft 365 outage?

Microsoft says it did not, describing the outage as “a technical issue not related to cyber or AI / agents”. Its stated cause is a network capacity upgrade and a later network change. That is the company’s account pending a full post-incident report, and no independent evidence points to AI.

What are incidents EX1550543 and MO1550821?

They are the identifiers Microsoft used in the Microsoft 365 admin centre for the November 2026 outage: EX1550543 for Exchange Online and MO1550821 for the wider Microsoft 365 incident. Administrators can look them up in service health history for Microsoft’s updates and, later, the post-incident report.

How do you keep working when Microsoft 365 goes down?

Agree a second communication channel that doesn’t depend on Microsoft 365, subscribe key people to service health alerts, keep offline copies of critical contacts and runbooks, and give any system that sends email through Exchange a fallback. Test the plan before you need it, because the November outage lasted more than a day.


Rujen Shakya is head of project management and a senior mobile engineer at Crimson Umbrella Technologies. You can find Rujen on LinkedIn.

Tell us what’s broken.

Twenty minutes, no deck. We’ll tell you straight whether this is work we should be doing, and if it isn’t, we usually know who should.

We come back within one business day.