Back to all posts

cat posts/microsoft-365-two-day-outage-mo1550821-your-exchange-fallback-plan.md --category "Azure & Cloud" --views 25

Microsoft 365's two-day outage: your Exchange fallback plan

Exchange Online incident EX1550543 opened at 16:17 UTC on 12 November 2026 and Microsoft 365 incident MO1550821 followed, taking Teams calendars, Copilot, Defender for Office 365 and Intune down worldwide until the early hours of 14 November. Here is what Microsoft has said about the health-check cascade, how the 25/50/100% service credit formula works for your tenant, the fallback plan to build now, and the NIS2 and DORA reporting clocks that applied to Swedish organisations.

  • --author By Falak Mahmood
  • --date November 15, 2026
  • --read 11 min read
  • --views 25 views

At 16:17 UTC on Thursday 12 November 2026, Exchange Online stopped serving mailboxes for a growing share of tenants in the Americas. Microsoft logged the Exchange incident as EX1550543, then opened MO1550821 at 18:26 UTC when Teams calendars, SharePoint sharing, Microsoft 365 Copilot, Defender for Office 365, Intune and Windows 365 began failing on top of it. By the Swedish evening the impact was global. Microsoft did not post its final resolution notice until the early hours of Saturday 14 November UTC, roughly 35 hours after the first Exchange incident opened. SVT reported that Region Västmanland, Region Sörmland, the Västra Götaland health region and dozens of kommuner from Blekinge to Värmland lost mail and Teams for the better part of a working day. If your organisation runs on Microsoft 365, this one found the gaps in your continuity plan for you. The work now is to write down what you learned and claim what you are owed.

What Microsoft has said so far

Microsoft's status updates during the incident were unusually candid about the mechanism, even while the trigger stayed unknown. The first substantive explanation, posted on Friday, said the elevated load "isn't driven by an increase in traffic, but by a number of servers being unexpectedly unavailable". Later updates narrowed it down: a core process on the affected servers was "becoming overwhelmed, which causes the servers to stop responding", and the failed health checks "may originate from a specific component that's responsible for processing server health". Microsoft paused deployments, redistributed connection traffic across regions, restarted servers in stages and, when early restarts collapsed under live traffic, validated an alternative health-check method so recovered servers could stay in rotation.

The resolution notice, as logged by admins who tracked the Service Health feed, gave a preliminary root cause: "The incident was introduced by an issue during an earlier network capacity upgrade that surfaced following a recent network change." The process that was overwhelmed is the one that establishes secure connections, so the picture is a TLS-handshake bottleneck that fed on itself. Microsoft also stated plainly that "this was a technical issue not related to cyber or AI / agents", apologised, extended the Exchange Online message retry window to 48 hours so queued mail would not bounce, and committed to a full post-incident report. That report had not been published at the time of writing.

Why a health check can take down a service that is otherwise fine

Health checks exist to pull sick machines out of a load balancer pool. The failure mode here is the inverse: the checker, or the probe deadline it enforces, starts failing machines that are merely busy. The sequence is familiar to anyone who has run a large fleet.

  • A latent fault from the earlier capacity upgrade sits dormant until a later network change exposes it.
  • Some servers slow down. Their TLS-establishing process saturates and they miss probe deadlines, so the load balancer evicts them.
  • Clients rehash onto the survivors. Outlook, mobile mail and Graph callers reconnect, concentrating fresh handshakes on fewer machines.
  • The survivors miss their probes too. More evictions, more concentration, until the pool collapses. Restarting a server does not help if it is thrown straight back into the same storm.

This is the same shape as the Google zone failure that took UK banks offline in October: a control-plane mechanism designed for safety amplified a small fault into a regional outage. The engineering lesson for your own services on Azure is to treat health-check thresholds, connection draining and restart pacing as production code that needs load testing, not as configuration defaults. For Microsoft 365, where you own none of that, the lesson is about dependencies.

What actually broke in a typical Swedish tenant

Exchange Online was the direct casualty, but the dependent-service list in MO1550821 is the part worth studying, because it shows how much of a modern tenant leans on mail and calendar infrastructure that most teams never think about.

  • Teams kept chat and calls running for most users, but calendar views and meeting join from the calendar failed, because those read from Exchange.
  • Microsoft 365 Copilot and Copilot Chat degraded or failed. Copilot grounds on mail and calendar through Graph, so any agent or Copilot Studio flow that reads a mailbox went dark with it.
  • Defender for Office 365 had nothing to scan while mail queued, then faced the backlog surge when queues drained on Friday. Watch for delayed-detection alerts dated 13 November.
  • Intune and Windows 365 were listed as impacted. If your emergency plan involved pushing a configuration or a Cloud PC fix through Intune during an incident, it had a circular dependency.
  • Purview and the Microsoft 365 Apps admin center were affected, which matters for anyone who needed to run an eDiscovery hold or check a tenant setting mid-incident.

Two things did not break for most tenants: Entra ID sign-in and Azure itself. That matters for the fallback design below. Your identity plane stayed up; your messaging plane did not.

The service credit math, and the deadline to claim it

The Microsoft Online Services SLA commits Exchange Online and the other Microsoft 365 services to 99.9% monthly uptime and pays service credits on a three-step scale. Credits are not automatic. An administrator must open a support request and file a claim, and the claim must be submitted by the end of the calendar month following the one in which the incident occurred. For this outage that means 31 December 2026.

Monthly uptimeService creditDowntime in a 30-day month (all users)
Below 99.9%25% of the monthly feeMore than 43 minutes
Below 99%50%More than 7.2 hours
Below 95%100%More than 36 hours

The catch is the formula. Microsoft does not measure wall-clock downtime; it measures user-minutes of downtime across your whole tenant:

Monthly Uptime % = (TM x TU - total downtime minutes across all users) / (TM x TU)

TM = minutes in the month (November 2026: 43,200)
TU = total number of users

Take a hypothetical 2,000-seat tenant. If every user lost mailbox access for the full 35 hours, downtime is 2,000 x 2,100 minutes, which gives 95.1% uptime and a 50% credit, just above the 100% threshold. If, as was more typical in Europe, roughly half the users were affected for about 20 hours, the figure is 98.6% and the credit is 25%. Either way the credit applies to the monthly fee for the affected service, and you must be able to show your working. Export the Service Health history for EX1550543 and MO1550821 now, record which user populations were affected and for how long from your own helpdesk tickets and sign-in logs, and attach the incident IDs to the claim. Admins who wait for the post-incident report before gathering evidence tend to find the admin center history has been trimmed by the time they look.

Your fallback plan: what to put in place before the next one

None of this requires leaving Microsoft 365. It requires accepting that a single vendor's messaging plane will fail for a day every year or two, and deciding in advance how your organisation behaves when it does.

  • 1. Build one communication channel that is not Microsoft. On Thursday evening, Swedish IT teams were trying to tell staff about an Outlook outage using Outlook. Pick an SMS broadcast service, a status page on a non-Microsoft host, or a phone tree, and test it quarterly. Entra ID stayed up, so a status site behind Entra sign-in on Azure Static Web Apps would have worked; one hosted in SharePoint would not.
  • 2. Subscribe to Service Health properly. Configure email alerts to an external address and install the Microsoft 365 Admin mobile app. During MO1550821 the admin center itself was listed as affected, so the push notification path mattered more than the portal.
  • 3. Decide your inbound mail posture. Sending servers retry; the SMTP default is typically a day or more, and Microsoft extended its own retry window to 48 hours. For most organisations that means mail arrived late rather than never. If you process payment instructions, procurement bids or patient referrals by email with same-day deadlines, that is not good enough, and you should price a secondary MX with a mail continuity provider against the cost of one missed deadline. For everyone else, document that late delivery is the accepted outcome and tell the business.
  • 4. Stop users self-healing. The most expensive part of Friday for many helpdesks was undoing profile resets, Outlook reinstalls and cached-credential clearing that users tried on their own. Your first out-of-band message should say: do nothing, it is not your machine.
  • 5. Freeze change during vendor incidents. Pause your own Intune deployments, conditional access changes and DNS edits until the vendor incident closes. A change made during the outage is impossible to separate from the outage in a post-mortem.
  • 6. Map your Copilot and agent dependencies. Any Copilot Studio agent, Power Automate flow or custom Graph integration that reads mail or calendar failed silently or loudly on Thursday. Inventory them, add explicit failure handling for Graph 503 responses, and decide which ones need a degraded mode. An agent that approves invoices from a mailbox should fail closed and alert, not queue approvals for a retry you cannot see.
  • 7. Check what else rides on Exchange. Room booking panels, MFA via email, password-reset flows, on-call paging that emails a distribution list, ticketing systems that ingest mail. Each one is a hidden dependency. The outage gave you a free dependency scan; collect the helpdesk tickets and turn them into a list.
  • 8. Run the tabletop now. Walk through 12 November with your incident team while it is fresh. Who declared the incident internally, and when? Who decided whether it counted as reportable? Who talked to the business? Write the gaps into the runbook.

The Swedish and EU angle: reporting clocks that do not care whose fault it was

For Swedish organisations the outage landed in the first year of the new cybersäkerhetslagen (SFS 2025:1506), in force since 15 January 2026, which transposes NIS2. Article 23 of the directive requires essential and important entities to send an early warning within 24 hours of becoming aware of a significant incident, an incident notification within 72 hours, and a final report no later than one month after the notification. The directive's definition of a significant incident is about impact on the entity's own services, not about who caused it. A regional health authority or a kommun that could not reach staff or deliver services because its supplier's mail platform collapsed should at minimum have minuted the decision on whether the threshold was met. If you did not have that conversation on Thursday, that is the first runbook gap to fix. Our NIS2 engineer's checklist for Sweden covers the detection and classification steps that have to happen before any clock starts.

Financial entities under DORA face a tighter cascade. Commission Delegated Regulation (EU) 2025/301, Article 5, sets the initial notification at four hours from classifying an incident as major and no later than 24 hours from awareness, an intermediate report within 72 hours of the initial notification, and a final report within one month of the intermediate report. The same regulation also puts Microsoft in your register of information as an ICT third-party provider, and a day-long loss of corporate email is exactly the kind of event your resilience testing under Articles 24 to 26 is meant to have rehearsed. Finansinspektionen will want to see that the scenario existed on paper before it happened in production. The DORA lessons from Swedish fintechs post lists what supervisors actually asked for in the first year.

There is a procurement thread too. Several Swedish public bodies have standardised on Microsoft 365 and Copilot as the sanctioned AI and collaboration platform, in some cases while blocking alternatives on security grounds. That is a defensible choice. It also means a Microsoft 365 outage is now a whole-of-organisation outage, and the continuity annex in the next upphandling should say what the supplier owes beyond a 25% credit. Credits are a rebate, not a remedy. The remedy is the plan above, funded and tested, and the outage has just given you the budget argument for it.

What to do this week

  • Export Service Health history for EX1550543 and MO1550821 and file the SLA claim with user-minute evidence before 31 December 2026.
  • Record, in writing, whether the incident met your NIS2 or DORA significance threshold and why.
  • Stand up a non-Microsoft broadcast channel and test it with a real message to all staff.
  • Inventory every integration, agent and workflow that depends on Exchange Online or Graph mail and calendar endpoints.
  • Read Microsoft's post-incident report when it lands, and compare its timeline with your own detection time. The gap between the two is your monitoring debt.

subscribe # the AI news that matters, minus the noise

Book a Call

Sources

Tags

Related posts