A stressed office worker faces cloud-service outages and cybersecurity alerts while using multiple devices.
When Microsoft 365 Copilot went dark last week, a lot of employees didn't wait for it to come back. They opened another AI tab and kept working. Two security vendors now have telemetry showing how often that happened, and it should make anyone who signs off on AI governance policy a little uneasy.

Under deadline pressure, a noticeable share of regular Copilot users moved to Claude, ChatGPT or Gemini. A lot of them did it on free accounts or without logging in at all, and some took sensitive business data with them. The outage didn't create shadow AI. It showed how much was already there.

A stressed office worker faces cloud-service outages and cybersecurity alerts while using multiple devices. The outage: what broke and why​

Some context first. Reporting on the incident says that on November 12, 2026, a Microsoft outage cut users off from Outlook and Exchange Online mailboxes, then from Teams features, SharePoint, Microsoft 365 Copilot and Defender XDR, in at least 74 countries. Benzinga reported that Microsoft says its Exchange Online incident began at 16:17 UTC. Microsoft tracked the incident under MO1550821 in the Microsoft 365 admin center. Admins writing their own post-incident notes should use that ID.

The reported length of the outage depends on who you ask and what they measured. An incident breakdown from eCorpIT gives Microsoft's confirmed incident window: 16:20 UTC on 12 November to 16:10 UTC on 13 November. CultureAI, the vendor behind the headline shadow AI numbers, says Copilot specifically was unusable for most people for about 19 hours, from Thursday evening into Friday lunchtime UTC.

Copilot didn't just stop answering. According to eCorpIT, Microsoft listed Copilot Chat failures and lost access to email, calendar and files, and said Graph requests for Exchange data could time out. The trouble also spread beyond Microsoft's own products: Anthropic and OpenAI both logged errors in their Microsoft 365 connectors.

On the cause, Microsoft was clear about what it wasn't. Windows Latest reported that the company described it as a technical issue not related to cyber or AI / agents. IBTimes UK reported that Microsoft's preliminary post-incident report traced the global outage to a previously undetected software defect in its networking gateway servers. That matters here. Shadow AI didn't cause the outage. It was how employees responded to it.

Section summary: a network-layer defect took down Exchange Online, and Copilot, Graph-based agents and third-party connectors went down with it for most of a working day.

What the telemetry shows​

CultureAI published its findings on November 17, based on telemetry from its own customer base. It compared November 12–13 with a normal week. Its key numbers:

  • About 7 in 10 regular Copilot users who opened the tool couldn't use it.
  • 1 in 11 regular Copilot users switched to another AI tool during the outage, roughly double the rate on a normal Friday.
  • 1 in 3 of those who switched entered personal or sensitive business data into the other tool.
  • 159% more people than normal used another AI tool on a free account or with no account.
  • 2.4 times as many people as normal put sensitive data into another AI tool, covering 11 different types of sensitive data.

Where they went:

Breakdown (CultureAI sample)Share
Claude57%
ChatGPT42%
Gemini4%
Company licence52%
Free account30%
No account10%
Individual paid plan6%

The tool percentages add up to more than 100% because some people used more than one tool. The account categories add up to 98%, and CultureAI doesn't explain the remaining 2%. The commonly quoted "40% on free or no accounts" combines the 30% and 10% figures.

There's good news in there too: just over half of the people who switched used a company licence. Many organisations already pay for a second AI tool, and employees found it. The worry is the other 40%. CultureAI describes that group as having no contract, no retention controls and no audit trail.

A second dataset, a different picture​

Harmonic Security ran its own analysis on a sample of its customers with large Copilot deployments. It compared the morning of November 13 with the same hours on earlier weekdays and found that between 06:00 and 10:59 UTC, the number of regular Copilot users accessing free or guest AI tools rose 120%.

The two datasets agree on the direction but disagree on the details. In Harmonic's sample, ChatGPT accounted for 84.4% of the free and guest activity, followed by Claude at 7.7%, Gemini at 7.5% and Perplexity at 0.4%. CultureAI's sample had Claude well ahead.

Why the difference? The methods aren't the same:

  • Harmonic counted only free and guest usage in a five-hour morning window.
  • CultureAI counted all switching, including to company licences, across the full outage.

Claude also had its own trouble that day, which could have pushed some users elsewhere. Harmonic notes that users couldn't log in to Claude on Friday morning, and that Anthropic hasn't linked that incident to Microsoft's outage.

The practical takeaway is that neither dataset is a census of all Copilot users. Both are vendor telemetry from paying customers. And both companies sell tools that address the problem they're describing.

Why this matters​

You don't need to trust a vendor's dashboard to see what's going on. Muhammad Yahya Patel, EMEA vCISO and cybersecurity advisor at Huntress, told IT Pro that outages create "the exact conditions under which people make poor security decisions." He listed them as time pressure, frustration, and a workaround that feels harmless.

Emanuela Zaccone, principal AI security strategist at Sysdig, put it differently. She told IT Pro the problem is continuity, not bad intent, because AI has become "load-bearing" for many people's daily work. If something holds weight, it can't be a single point of failure. It sounds obvious, yet very few governance plans account for it.

The UK National Cyber Security Centre made a similar point in its September 2026 guidance on shadow AI, before this outage happened. It warned that data handed to consumer AI services may be stored, retained or used to improve the service unless specific privacy controls apply. It also said the realistic goal is to reduce risk, not to eliminate shadow AI.

A caveat on the evidence: none of this research shows that any specific prompt was retained or used for training. That depends on each service's settings and terms, which the telemetry can't see. The verified risk is loss of visibility and control.

IT Pro also cited a Gartner projection from November 2025 that around 40% of organisations will experience shadow AI-related breaches by 2030. That figure hasn't been checked against Gartner's original publication, so treat it as a projection.

Section summary: employees weren't trying to leak data. Their tool disappeared and they reached for the nearest alternative. Governance plans that assume one sanctioned tool is always available failed that test.

A fallback plan in six steps​

These steps are based on the interviewees' advice, the NCSC guidance and general incident-response practice. There's no Microsoft-endorsed playbook for this, so the steps describe policy, not product settings.

  1. Find out what's actually in use. Zaccone's starting point is an inventory of what's running on endpoints and who uses what. You can't plan a fallback around tools you don't know about.
  2. Name the approved fallback before you need it. Patel recommends pre-approving a small set of alternative AI tools with suitable data-handling agreements, telling staff about them clearly, and making them easy to reach. If you already pay for a second enterprise AI licence, as many CultureAI customers apparently do, make it the documented backup.
  3. Define what work can continue. Decide which tasks can move to the fallback and which must wait. Public research and drafting are probably fine. Customer contracts, personal data, credentials and source code probably aren't.
  4. Control the data, not just the domains. Harmonic's Michael Marriott suggests assuming workarounds will happen, getting visibility across corporate and personal accounts, and applying controls based on what's being shared. In his example, analysing a public article in free ChatGPT goes through, while pasting in a customer contract is blocked with an explanation. Not every organisation can enforce inline content controls on every service, so be realistic about your current tooling.
  5. Add it to your incident communications. Zaccone recommends that staff hear the same guidance on a normal day and on the day something breaks. When the Microsoft 365 admin center posts an incident like MO1550821, the outage notice to staff should include the AI fallback instructions.
  6. Afterwards, review what people did. Harmonic argues that where staff go during an outage tells you which tools they actually want. Sustained demand on personal accounts may justify an enterprise licence, which brings that work back under company controls.

What about blocking everything? It's tempting, but CultureAI argues it would just push the same deadline-driven work onto personal phones, where IT can see nothing. That's the vendor's view, not something it measured, but it fits the NCSC's position that shadow AI won't go away completely.

The bottom line​

Microsoft's outage was an infrastructure failure, and Microsoft has said AI didn't cause it. It still worked as an unplanned test of every organisation's AI governance policy, and the vendor data suggests many failed it.

Whichever AI tool you depend on will go down again at some point. The question is whether your staff know what to use instead before it happens.

How did your organisation handle the November 12–13 disruption? Did you have an approved fallback, or are you still trying to work out where the data went?

 

References

  1. Looking for workarounds: Why some employees turn to shadow AI when outages hit – and how IT leaders can stop it - IT Pro IT Pro 2026-11-20T10:49:25+00:00
  2. The hidden risks of shadow AI | National Cyber Security Centre ncsc.gov.uk
  3. Microsoft Outage Triggers Chaos Across Outlook, Teams and Copilot — Spills Over to OpenAI and Anthropic ( - Benzinga benzinga.com