TNN — Torbrook News Network
TNN Analysis · Tech / Business

Microsoft 365 outage takes down Outlook and Teams for hours as servers fail their own health checks, and Microsoft says its first fixes did not work

Nearly 130,000 users reported problems on Thursday as Exchange Online, Teams, OneDrive and Copilot faltered worldwide. Microsoft says traffic did not cause it: servers dropped out of service, the rest buckled under the load, and recent configuration changes are under review.

By the TNN Analysis Desk· November 12, 2026 · 7 min read
Microsoft 365 outage takes down Outlook and Teams for hours as servers fail their own health checks, and Microsoft says its first fixes did not work
Aerial view of a Microsoft Azure data center campus in central Washington state. The facility shown is not identified as the source of Thursday's outage. Photo: Tedder (CC BY 4.0), via Wikimedia Commons.

The most revealing sentence Microsoft published during Thursday's outage was not an apology or an estimate. It was a diagnosis: the strain on its systems, the company said, 'isn't driven by an overall increase in traffic, but by a large number of servers being unexpectedly unavailable, which is shifting load onto the remaining infrastructure'.

In other words, nobody overwhelmed Microsoft 365 from the outside. Servers inside the service stopped answering, failed the automated checks designed to keep the system healthy, and were taken out of rotation. The servers left behind then had to carry the full load of email and collaboration for a large share of the world's offices, and began to fail as well. For several hours on Thursday afternoon, the mechanism meant to protect the service was part of what was bringing it down.

What failed, and for whom

Problems began shortly before 1 p.m. Eastern time (18:00 GMT), according to CNBC. Shortly after 1 p.m., about 130,000 users had reported problems with Outlook on the outage-tracking site Downdetector. Microsoft said the core of the problem was Exchange Online, the cloud email service behind business Outlook accounts. It said the issue could affect users in any region.

The damage spread well beyond email. The Guardian reported Downdetector complaints covering Word, Excel, PowerPoint, Teams, SharePoint and OneDrive, as well as the Copilot AI assistant. Users described messages that would not send, inboxes that would not load and meetings that would not connect. OpenAI and Anthropic both reported that the connectors linking ChatGPT and Claude to Microsoft accounts were affected.

There was also some noise. Downdetector showed a rise in reports for Amazon Web Services and Google at the same time, which Amazon pushed back on directly. It said on X that its services were 'operating normally' and that an 'unrelated disruption elsewhere on the internet is prompting some inaccurate news reports'. When a service as widely used as Outlook goes down, users often blame whatever else they were trying to reach.

Reports fell from their peak to about 77,000 by just before 3 p.m. Eastern, according to Reuters figures cited by Al Jazeera. Microsoft later said it had restored Teams in North America, although some users continued to report connection problems.

Four hours of fixes

Microsoft's public updates, posted on its Microsoft 365 Status account and its service-health page, read as a record of an engineering team ruling things out in real time.

First, it attributed the connectivity errors to 'elevated load across multiple regions'. Next, it narrowed that down: a number of servers were failing automated health checks and had become unavailable, straining what remained. As a precaution it 'paused all deployments and configuration changes' while it reviewed recent changes for any link to the incident.

At about 4 p.m. Eastern (21:00 UTC), it reported progress. It had 'validated an alternative service health check method' on a subset of affected servers, which allowed them to return to service, and was rolling that change out more widely. It was also preparing automated systems to detect and restart servers that stopped responding. By 21:41 UTC, it said its review had found a small number of recent configuration updates that might be responsible, and that it was considering reversing them to speed recovery.

Then, at 22:11 UTC, came the admission that gave Thursday's story its edge. The actions taken so far, Microsoft said, 'have not provided the anticipated relief'. A core process on the affected servers was being overwhelmed and causing them to stop responding. The company said it was now fully isolating part of its infrastructure, restarting components and reintroducing traffic gradually, and was considering whether separating traffic region by region could allow parts of the world to recover independently.

"If you lose part of that pool, then all of the users are trying to access the remaining servers, and that overloads them. And then they go down. And so you end up with a cascading problem." — Mark Gregory, associate professor, School of Engineering, RMIT University, to the Guardian

Why a health check can make things worse

Large cloud services run on pools of interchangeable servers behind load balancers. Each server is checked constantly. If it fails to respond correctly, it is removed from the pool so users are not sent to a broken machine. Most of the time this is what keeps a service up: one bad server is replaced without anyone noticing.

The design assumes that failures are independent and rare. If something causes many servers to fail their checks at the same time, such as a configuration change applied to all of them or a check that is itself wrong, the system removes a large part of its own capacity in one go. The same traffic then goes to fewer machines, and they start to fail too. That is why Microsoft's first meaningful progress came from changing the health check itself, not from adding capacity, and why its next step was to isolate regions so a failure in one could not keep pulling down the others.

It has happened before

Configuration changes have been behind several of the biggest outages of recent years. On October 29, 2025, Microsoft's Azure Front Door network suffered a global outage that the company attributed to 'an inadvertent configuration change'. It knocked out sign-ins to Microsoft 365 and other services for several hours, until engineers rolled back to the last known good configuration.

Gregory compared Thursday's incident to the Optus outage in Australia in November 2023. In that case, routing information received after a routine software upgrade exceeded safety limits on key routers, which disconnected from the network's core and cut off about 10 million customers for most of a day. Australia's communications regulator later investigated. The pattern in both cases was the same: a small change to a system designed with safeguards, and safeguards that responded by shutting down parts of the network.

The largest example remains July 19, 2024, when a faulty configuration file in a CrowdStrike security update crashed Windows machines around the world. Microsoft later estimated that 8.5 million devices were affected, less than 1% of all Windows machines, yet airlines, hospitals and banks were disrupted for days. CrowdStrike said afterwards that the faulty data had passed its validation checks, and it added further review steps to its release process. That outage showed how small a percentage of affected systems can still cause worldwide disruption when the affected systems are the ones everything else depends on.

Gregory drew the same lesson from Thursday. He said companies need to invest more in testing configuration changes in separate environments before they reach the live network, because once a bad change is live the system's own redundancy can spread it. Microsoft's decision to freeze all deployments and configuration changes during the incident suggests it had reached the same conclusion while the outage was still running.

The case for perspective

There are limits to how much one afternoon proves. Downdetector counts user reports, not affected accounts, and its numbers reflect how many people go looking for an explanation as much as how many are affected. A peak of 130,000 reports is large, but it is a small fraction of Microsoft 365's user base, and many organizations appear to have been only partly affected. Exchange Online is also designed to retry delivery of queued mail for an extended period before returning it to the sender, so delayed messages are not necessarily lost. Microsoft had not reported any data loss as of Thursday evening.

Microsoft also did what operators are supposed to do: it froze changes, narrowed down the failure, published frequent updates and tested a recovery approach on part of the system before applying it everywhere. Gregory's advice that businesses keep a fallback, such as a separate email account, is reasonable, but in practice few organizations run a second email and collaboration system just in case.

The durable point is about concentration, not competence. One company's email and collaboration service is now the default working environment for a large share of businesses, schools and governments, and increasingly for the AI tools that connect to them. When a few configuration updates and a health check can take that environment offline for an afternoon, the risk is no longer one company's engineering problem. It is shared by everyone who built their working day on top of it.

This report is based on Microsoft's statements on its Microsoft 365 Status account and service-health page and to CNBC; reporting by CNBC, the Guardian and Al Jazeera, including Downdetector figures and Reuters data; comments by RMIT University's Mark Gregory to the Guardian; and published accounts of the October 2025 Azure Front Door outage and the November 2023 Optus outage. As of Thursday evening US time, Microsoft had not declared full recovery or confirmed a root cause; its references to recent configuration updates were preliminary. Times are as published by Microsoft and the outlets cited.