
The November Microsoft 365 outage disrupted Outlook and left some Copilot features without access to work data, showing how AI workflows can stall even when their models stay online.
During the Microsoft 365 outage on 12 and 13 November, the UK Supreme Court told people it could no longer receive email. Its online portal and contact forms were still working. Inside Microsoft 365, the pattern was similarly uneven. Some users could open a file in SharePoint while struggling to access Outlook. A previously scheduled Teams meeting could go ahead even when the calendar functions used to schedule it were impaired.
For someone trying to get through the workday, the immediate problem was finding another way to finish a task. For an AI assistant designed to read email, look up documents and act on what it finds, the same outage could leave it with only part of the record. A procurement agent might still see what the company ordered but lose access to the supplier’s reply. If it treats those two pieces of information as interchangeable, a connectivity fault can turn into a bad operating decision.
What happened during the Microsoft 365 outage
Microsoft’s preliminary post-incident report puts the main disruption at 23 hours and 50 minutes, from 16:20 UTC on 12 November to 16:10 UTC on 13 November. Exchange Online was hit hardest. Some users could not connect to their mailboxes, and email delivery was delayed. Microsoft continued monitoring the service and clearing the backlog after the active disruption ended. It issued closing updates on 14 November UTC under incident IDs EX1550543 and MO1550821. Microsoft said the cause was a technical fault unrelated to cyber activity or AI agents.
The fault traced back to a software change that started rolling out in July. It introduced a defect into the networking gateway that handles connections for several Microsoft 365 services. About 15 hours before the outage, an automated failover shifted Teams traffic onto the primary gateway, which also supported Exchange Online. Nearly 14 per cent of the gateway’s server capacity was already out of service because of a separate issue. The change in traffic and reduction in capacity together exposed the defect.
When servers stopped responding, Microsoft’s routing systems moved traffic to those still operating. That put more pressure on the remaining machines. At the low point, more than 80 per cent of the gateway servers were unresponsive. Waiting connections overwhelmed some servers as they returned to service, so engineers applied a fix and restored capacity in stages. Core Teams chats and meetings worked for much of the incident, while features tied to Exchange Online continued to have problems.
Why some Microsoft 365 Copilot features failed
Microsoft 365 Copilot uses Microsoft Graph to retrieve information the signed-in user has permission to access. Microsoft calls the process of giving the model that work context grounding. During the outage, the company warned that Graph requests for Exchange Online data, including email, calendars and contacts, could time out. It also reported that some users could have trouble opening Copilot Chat or getting a response, while some Copilot experiences could lose access to work data.
Take the procurement example. The purchase order in SharePoint lists Friday as the delivery date. The latest confirmation, if the supplier sent one, would be in a mailbox the assistant cannot reach. The assistant can still report the date on the purchase order and say it could not check the supplier’s reply. That gives the buyer enough information to decide whether to call the supplier before assigning labour to unload the shipment.
For teams building retrieval-augmented generation (RAG) applications, the engineering detail is easy to overlook. A search that ran successfully and found no message is a different outcome from a search that never reached the mailbox. If the application records both as an empty result, the model has no way to communicate the difference. The application needs to record which sources responded and make that information available to the user.
Where a backup model helps
With model routing across providers an application can switch to a backup model when its primary model becomes unavailable.. Its value depends on what else the workflow needs. If both models rely on the same connector, identity service or mailbox, the switch may restore the ability to generate text while leaving the business task blocked.
Start with the action the business expects the agent to complete. An assistant that drafts a purchase summary may only need read access to a document library. An agent that confirms a delivery date may need the latest email, authority to act and a way to record the decision in the purchasing system. Those workflows have different recovery requirements, even if they use the same model. A second provider is worth evaluating once the team knows which failure it is trying to cover.
How AI agents should handle retries and recovery
The return of a service can be as important to an agent’s design as the moment it becomes unavailable. A connector, an API client and a workflow orchestrator may each retry the same failed request. If each layer independently makes three attempts, a single job can produce up to 27 calls to the dependency. Google’s site reliability guidance recommends capping retries and avoiding retries at multiple layers. Microsoft Graph tells clients to honour the Retry-After header when throttled, and a circuit breaker can pause requests while a service is repeatedly failing.
Operators need to see failures by dependency. Track how long work has been waiting and whether attempts to retrieve a known record still succeed. If the email connector is failing while document search is healthy, pause the jobs that require email and keep the others moving. During recovery, let a small number of jobs through first. Releasing the entire backlog as soon as a request succeeds can make it harder to see whether the connector is still failing intermittently.
Write actions need particular care. A request to create a ticket can time out after the destination has already saved it. Resending the request blindly could create a duplicate. Where the destination supports it, use an operation identifier that lets it recognise a repeat request. Otherwise, check the destination before replaying actions whose outcome is uncertain. As the queue resumes, review jobs that depend on information which may have changed during the outage. A summary can be regenerated. An approval may need to go back to its owner.
Keeping work moving during an outage
Microsoft also reported problems with some Purview data-loss prevention and labelling functions, and with access to parts of the Defender portal. An outage plan should account for the possibility that the tools used to enforce or check data-handling rules are affected alongside the apps staff use to do their work. A fallback should specify which account people may use, what information they may move and where decisions will be recorded. That gives staff a way to meet a deadline without improvising with sensitive material in an unapproved tool.
For a policy assistant, an approved read-only copy of a small set of procedures may keep routine questions moving. The copy needs a refresh date, an expiry and the same access restrictions as the source. A workflow that approves payments or makes commitments to suppliers may need a person to take over. The team should know who can authorise the action, how to reach them if email is down and what record to reconcile against the source system once it returns.
Measure recovery at the workflow level
A continuity exercise can start by taking email and its connector out of a test workflow for a simulated business day while leaving SharePoint and the model available. Give the team a representative queue. See which items the assistant can still handle, which ones people must pick up and what happens when the queued actions start running again.
The measure that matters is how long it takes to resume the business task with enough evidence to act and a record of what happened in the meantime. A Copilot window that opens says little about whether the buyer can verify Friday’s delivery. The workflow is back when the buyer can confirm the shipment, make the staffing decision and trust that the system will record it once.
Discover more from Poniak Times
Subscribe to get the latest posts sent to your email.




