Insights From the Copilot and Claude Outages

Published on
November 18, 2026
Illustration of an AI assistant window showing a power-off symbol, with dotted paths branching out to three alternative tools
Contributors

Insights From the Copilot and Claude Outages

On the afternoon of Thursday, November 12, 2026, Outlook stopped loading for Microsoft 365 users in the Americas. Within two hours the failure had spread to Teams, SharePoint and Microsoft 365 Copilot, and roughly 130,000 users in dozens of countries reported Outlook problems to Downdetector. Microsoft traced the outage to routine network maintenance that exposed a latent bug, and recovery stretched into Friday morning in the UK and Europe. The disruption also reached tools built on Microsoft 365, including Claude's Microsoft 365 connector.

A separate problem followed. On Friday morning users could not log in to Claude, and a second round of errors that afternoon covered the start of the US working day. Anthropic has not linked either incident to Microsoft's outage. For parts of Friday on both sides of the Atlantic, two of the AI assistants enterprises rely on most were unreliable at the same time.

The new operating system for work

AI assistants have become something close to the operating system for daily work. Three years ago they were a chat window for pasting text into. Today Copilot, Claude and ChatGPT sit on the desktop, connect to email, files and ticketing systems, and draft, search and take actions on an employee's behalf. Most employees now begin their tasks in the assistant and reach their other applications through it. Microsoft says more than 90% of the Fortune 500 use Microsoft 365 Copilot. When that layer fails, employees lose the route by which much of their other work gets done.

Most organizations already treat email and identity as infrastructure that needs a continuity plan. AI assistants now belong in the same category, yet few organizations plan for their downtime. When everyday work runs through a single provider, one outage can stall drafting, analysis and research across the whole company at once. The deadlines remain, though, and when the approved tool is unavailable, employees find another one.

Employees responded with Shadow AI

We analyzed AI activity across a sample of Harmonic customers with large Copilot deployments, comparing the morning of November 13 with the same hours on preceding weekdays. Between 06:00 and 10:59 UTC, the number of regular Copilot users accessing free or guest AI tools rose by 120%. ChatGPT accounted for 84.4% of that free and guest activity, followed by Claude at 7.7%, Google Gemini at 7.5% and Perplexity at 0.4%.

Line chart showing regular Copilot users' hourly use of free AI tools on 12 and 13 November 2026, rising to 2.2 times the usual level as the European working day began during the Copilot outage

The organizations in the sample had selected an enterprise assistant, licensed it and directed employees to use it. When it became unavailable, employees with work to finish moved to alternatives, largely on accounts that, without a tool like Harmonic, would be outside their security teams' view. Most free and personal tiers of AI tools train on inputs by default, and pose both an outsize security and privacy risk without tool-agnostic guardrails in place.

The pattern is not confined to outages. Our AI Usage Index found that 64.5% of activity on personal AI accounts is business work. Employees will inevitably look beyond the approved set whenever the available official channels aren't meeting their needs.

Securing the workaround, and learning from it

The common response to shadow IT is to tighten restrictions on unapproved tools. In practice, blocking tends to move the same work into accounts with no oversight and heightened security risk.

Organizations that manage this dilemma best start from the assumption that workarounds to blocking will occur. They establish visibility across both corporate and personal accounts, and they apply controls based on the content being shared. Under that model, an employee analyzing a public article in a free ChatGPT account proceeds without interruption, while one pasting in a customer contract is blocked inline with an explanation.

Understanding shadow AI usage also provides a powerful signal of where to invest next to maximize productivity. During the Copilot outage, employees demonstrated which tools they choose when the decision is theirs. The products they selected, the teams that moved first and the work they carried with them are evidence that AI committees rarely have. Sustained demand for a tool on personal accounts indicates where an enterprise license may be justified, and moving that demand onto enterprise accounts returns the work to company controls.

Microsoft has committed to publishing a post-incident report, and both providers have restored service. Future outages will have different causes, and the evidence from November 13 suggests employees will respond in the same way. Organizations with visibility into that response are better placed both to contain the risk and to learn from it.

Harmonic's primer on safely enabling workforce AI covers this approach in more detail.

‍

FAQ

What happened during the November 2026 Microsoft 365 outage?

On November 12, 2026, a failure in Exchange Online stopped Outlook from loading and spread to Teams, SharePoint and Microsoft 365 Copilot. Microsoft traced it to routine network maintenance that exposed a latent bug, and recovery stretched into the next morning in the UK and Europe. A separate Claude login outage followed on November 13.

Did employees turn to shadow AI during the Copilot outage?

Yes. Across a sample of Harmonic customers with large Copilot deployments, the number of regular Copilot users accessing free or guest AI tools rose 120% between 06:00 and 10:59 UTC on November 13, compared with the same hours on preceding weekdays. ChatGPT accounted for 84.4% of that activity.

How can security teams reduce shadow AI risk without blocking AI tools?

Start from the assumption that workarounds will happen. Visibility across both corporate and personal AI accounts, combined with controls based on the content being shared, lets low-risk use proceed while sensitive data such as a customer contract is blocked inline with an explanation.

‍

Build Your AI Guardrails Now

Gain the visibility and control you need to guide AI use with confidence.

© 2026 Harmonic Security