Microsoft 365 Went Down. What Were Your AI Agents Doing?

On Thursday, 12 November, a lot of companies found out something they hadn't planned for.

It started with users in the Americas unable to reach their Exchange Online mailboxes. Within hours Microsoft was warning of disruption to Teams, SharePoint Online, Microsoft 365 Copilot, Defender XDR and "many other services". Outlook went down on desktop, web and mobile. "Microsoft outage" became the most searched technology topic of the week, with more than ten million searches in the US and over a million in India.

Microsoft said a process that establishes secure connections had become overwhelmed after a large number of servers became unavailable. It began rolling back a recent network configuration change that coincided with the start of the disruption, and said it had no evidence of malicious activity. At the time of writing it hasn't published a final root cause.

Then the knock-on effects started. Anthropic told users that problems with Claude's Microsoft 365 connector and Claude for Outlook were "caused by a Microsoft 365 service incident". The next morning, Claude had a separate sign-in outage of its own.

Two outages, one week, and a very practical question for anyone running AI agents in production: when the systems your agents depend on go down, what do your agents do?

The uncomfortable truth: agents inherit every dependency

A person whose email is down knows it. They pick up the phone, wait, or do something else.

An agent doesn't know anything unless you designed it to. If an agent reads orders from a shared mailbox, updates a SharePoint list and posts a summary to Teams, it now depends on all three. Add the model provider it calls, and maybe a second agent from another vendor, and you have a chain where any one link can fail.

Here are six hiccups an outage like this exposes, and what we'd fix first.

Hiccup 1: The agent keeps retrying, and keeps spending

What happens

The mailbox is unreachable, so the agent retries. Each retry calls a model to "think about" the error. Multiply that by every agent, every few seconds, for a day-long outage, and you get a bill for work that produced nothing.

The fix

Budgets that stop, not budgets that report. Every agent should have a spending limit that actually halts it, plus a sensible retry policy: back off, then pause and wait for a person. An outage should cost you time, not money.

Hiccup 2: Half-finished work nobody can find

What happens

The agent had approved three of five steps in a supplier claim when the connection dropped. Was the credit note raised? Was the distributor told? Nobody knows, because the agent's "memory" was a running conversation that's now gone.

The fix

Record every step as it happens: what the agent saw, what it decided, which step it reached. When systems come back, the work resumes from the last good step instead of starting over, or worse, running twice.

Hiccup 3: One model provider, one point of failure

What happens

Your agents call one model provider. That provider has a bad morning, like Claude's sign-in outage the day after Microsoft's, and every process that touches AI stops at once.

The fix

Don't hard-wire your operations to one model. Design agents so the model is a setting, not a foundation, and keep a tested second option for the work that can't wait. Not every task needs the same model anyway: classifying an email and drafting a legal reply are different jobs.

Hiccup 4: Nobody told a human

What happens

The most dangerous outage is a silent one. The agent fails quietly, the queue backs up, and someone notices on Monday when a large customer asks why nobody replied.

The fix

Failure should go to a person, fast. When an agent can't do its job, it should stop and tell its owner in plain words: what's stuck, since when and what's waiting. Agents and people work as one team, and teams tell each other when something's wrong.

Hiccup 5: Agents you didn't know depended on Microsoft 365

What happens

When something like this happens, IT teams start checking which business processes are affected, and they often find agents they didn't know existed: a sales-ops automation that reads Outlook, a no-code flow posting to Teams, a vendor tool that quietly added "agentic" features in its last release.

This is shadow AI. Agents make it riskier than shadow spreadsheets ever were, because agents act. They send, update and approve.

The fix

Keep an inventory of every agent, who owns it and which systems it reaches, whichever framework it was built on. You can't plan for an outage in a dependency you can't see.

Hiccup 6: "Explain what happened" takes a week

What happens

The outage ends. Leadership, a customer or the auditor asks: what did our agents do during those hours? Which decisions were made, which were skipped, and was anything sent twice?

If the answer is buried in scattered logs, someone loses a week reconstructing it. Regulators are moving in the same direction. India's AI Governance Guidelines put accountability at the centre, and the RBI's FREE-AI framework asks banks and NBFCs to include AI in business-continuity planning, incident reporting and audit.

The fix

A decision record from day one. For every run: inputs, decisions, the rule that applied, who approved it and where it stopped. After an incident, that record is your post-mortem, already written.

This isn't really about Microsoft

Microsoft runs some of the most reliable infrastructure on earth, and it still had a bad day. So will every provider your agents depend on, sooner or later.

Gartner has predicted that more than 40 percent of agentic AI projects will be cancelled by the end of 2027, pointing to rising costs, unclear value and weak risk controls. Outages like this week's are exactly where weak risk controls get exposed. An agent that keeps spending, loses its place, fails silently and can't explain itself is an agent leadership will switch off.

The good news is that none of the fixes are exotic. Budgets that stop, steps that are recorded, models you can swap, failures that reach a person, an inventory and a decision record. That's not advanced AI. It's operations discipline, applied to a team that now includes software.

A five-minute check for your team

Before the next outage, ask:

• Which of our agents depend on Microsoft 365, Google Workspace or a single model provider?
• What does each one do when a dependency is down: retry, pause or alert?
• Is there a spending limit that actually stops it?
• Can we see exactly where each piece of work stopped?
• Who gets told, and how fast?

If you can't answer these for every agent you run, that's the work to do this quarter. We're happy to go through it with you, one process at a time.