# Microsoft 365 outage: your AI agent went down with it

> Outlook was down for over a day and OpenAI and Anthropic lost their Microsoft 365 connections. An agent is only as available as the weakest link in its chain.

- URL: https://www.iqsource.ai/en/blog/microsoft-365-outage-ai-agents-single-vendor/
- Published: 2026-11-22
- Author: Ricardo Argüello, IQ Source

Ask a vendor about uptime and you get one number. Ask how available your AI agent is and the honest answer is several numbers multiplied together, and the result is always smaller than the smallest one.

Ten days ago that math stopped being theoretical.

On Thursday, November 12, Microsoft began reporting Exchange Online mailboxes it couldn't reach in the Americas. By afternoon, incident MO1550821 covered Exchange Online, Teams, SharePoint Online, Microsoft 365 Copilot and Defender XDR in any region, <a href="https://www.bleepingcomputer.com/news/microsoft/massive-microsoft-365-outage-disrupts-outlook-teams-copilot-worldwide/" target="_blank" rel="noopener noreferrer">per BleepingComputer's running coverage</a>. <a href="https://www.cbsnews.com/news/microsoft-battling-major-outage/" target="_blank" rel="noopener noreferrer">CBS News put the disruption at more than 24 hours</a>, with over 130,000 Downdetector reports on Thursday alone.

The line in the CBS piece I keep rereading is the one about AI. OpenAI and Anthropic each reported problems connecting to Microsoft 365 tools. <a href="https://www.foxbusiness.com/technology/microsoft-apologizes-after-major-outlook-outage-worldwide-blames-technical-issue" target="_blank" rel="noopener noreferrer">Fox Business reported</a> those issues were still going Friday morning and only started easing that afternoon.

## Three vendors, one agent

Take an assistant that triages a sales inbox. It needs a model to read and decide. It needs the mailbox. It needs an identity, some account or token that lets it into that mailbox. Three different services, often from three different companies, and the agent works only while all three answer.

Run the numbers with a generous assumption. Say each of the three is up 99.9% of the time, which works out to under nine hours of downtime a year apiece. Chain them and the agent is up about 99.7%, roughly 26 hours a year. Same vendors, same SLAs, three times the dark hours, because any one of them can stop the agent.

On November 12 the model side was not the problem. The mailbox was. And an email agent without email is a very expensive idle process.

Linas Beliūnas framed the outage as irony. Microsoft cut staff while betting on AI, then had to wait on human engineers to bring Outlook back. Fair enough as a jab. For anyone running agents, it's a dependency story, and a pretty ordinary one.

## What Microsoft says broke

<a href="https://finance.yahoo.com/markets/article/microsoft-continues-to-restore-outlook-email-after-outage-rules-out-cyber-or-ai-agents-193042673.html" target="_blank" rel="noopener noreferrer">Microsoft told Yahoo Finance</a> the cause was an issue introduced during an earlier network capacity upgrade that surfaced after a recent network change. Not a cyberattack. Not AI agents. At one point, servers it had brought back went unresponsive again.

None of that was visible from the customer side, and none of it was fixable from there. What you controlled was how your own systems behaved for more than a day.

Microsoft, for its part, had a plan for its stuck mail. It stretched the email timeout window to 48 hours so delayed messages wouldn't bounce back to senders while queues drained. Small detail. It's also exactly the kind of decision most agent deployments never write down.

## The table we write before the prompt

When we design an agent at IQ Source, one of the first artifacts is half a page long and pretty dull. One row per dependency: the model, each data source, the identity it runs as, every system it writes into. For each row:

| Question | Why it matters |
|---|---|
| What does the agent do when this stops answering? | Pause, queue, or retry, and how many times |
| Who gets the alert? | A named person, not a shared inbox that is also down |
| How long can the work wait? | Past that point, a human picks it up by hand |

The second row bit a lot of companies during that outage. If the alert for "email agent failed" goes to email, it goes nowhere.

Multi-model routing helps with a different failure. I've argued that <a href="/en/blog/grok-bot-model-routing-buying-policy/">routing between models belongs in your AI contract</a>, and I still think so. It covers the day your model provider goes dark. It does nothing on the day your data source does. Same story as the <a href="/en/blog/litellm-attack-ai-trust-chain-broken/">LiteLLM attack</a> on the security side: every component you add to an agent's chain is one more place it can snap.

## Where to start this week

Pick the agent that touches the most customer-facing work. Write its dependency table. Then look at what it actually did on November 12 and 13. Did it stop cleanly? Did it hammer a dead endpoint? Did it keep answering from stale data?

If you can't tell, that's your first finding. Our <a href="/en/services/ai-operations/">AI operations</a> work starts right there, with what your agents did during the last outage, and builds the table from it.

## Frequently asked questions

### What happened during the November 2026 Microsoft 365 outage?

On November 12, 2026, a global Microsoft 365 outage disrupted Outlook, Teams and Copilot. Microsoft tracked it as MO1550821. CBS News reported it lasted more than 24 hours, and Microsoft blamed an issue from an earlier network capacity upgrade rather than a cyberattack.

### Did the Microsoft 365 outage affect ChatGPT and Claude?

Yes. During the November 2026 Microsoft 365 outage, OpenAI and Anthropic both reported problems connecting to Microsoft 365 tools. The connectors that let ChatGPT and Claude read Microsoft email, calendars and files stopped working until Exchange Online recovered on Friday afternoon.

### How should a company plan for an AI agent's vendor going down?

A company running AI agents should list every service each agent depends on, write down what the agent does when each one fails (pause, queue the work, or hand it to a person), and route an alert to someone. Without that, a vendor outage like Microsoft 365 stops work silently.

### Does using several AI models protect an agent from a Microsoft 365 outage?

Only partly. Routing across AI models protects an agent when a model provider goes down. It does nothing when the data source fails. If the agent works on Microsoft 365 email and Exchange Online is down, no backup model can read that mailbox.
