THE WRITE-OFF · ISSUE 001

The day the inbox went quiet

Microsoft’s 36-hour Microsoft 365 outage left AI agents staring at empty inboxes. The outage was Microsoft’s fault. The exposure belonged to everyone who wired agents to email without a plan for going dark.

TODAY’S STORY

Robert J. Szczerba, Forbes

·

13 November 2026

On Thursday 12 November, Exchange Online started failing. By Friday, Outlook, Teams and Copilot were down or limping across more than 20 Microsoft 365 services, and the damage reached as far as Australia’s spy agency, its federal police and Telstra. Microsoft paused all deployments, extended its email timeout window to 48 hours so delayed mail wouldn’t be lost, and declared recovery on Saturday. It also went out of its way to say the cause was a botched network upgrade, with no AI or agents involved.

The AI angle came from downstream. Anthropic’s status page reported that Claude’s Microsoft 365 mail, calendar and search tools were failing for most requests. OpenAI reported errors on ChatGPT’s Microsoft 365 connection. Any business that had pointed an agent at a Microsoft inbox found its clever new digital worker staring at nothing.

THE VERDICT

The outage was Microsoft’s fault. The exposure was entirely self-inflicted.

Companies plugging agents into email built a process with a single point of failure, no way of telling “quiet” from “broken”, and no plan for the moment the backlog arrives. Swapping AI providers would have changed nothing, since every provider was reading the same dead mailbox.

A process with no plan for going dark

Silence looked like success

An agent polling an inbox can’t distinguish an empty inbox from a stuck one. Both return zero new messages. A human would notice that the phones were ringing while the inbox was oddly calm. The agent simply concluded it was a slow Thursday.

The backlog is the real hazard

Szczerba sketches a scenario (illustrative, not a reported incident) that should make any COO wince. A customer emails asking for a $2,000 espresso machine to go to her new apartment. The email sticks. She phones instead, her plans have changed, and a rep redirects the order to her parents. Next morning the delayed email lands, the agent treats it as fresh, and politely sends the machine back to an empty building. Nobody did anything wrong. The process did.

Nobody owned the off switch

Agents were deployed with authority to change orders, yet nobody was named to decide when they stop, who covers urgent work while they’re paused, or when they’re allowed back on. Maersk, to its credit, defaulted to the oldest resilience plan in business: it told customers to pick up the phone.

Diversification was an illusion

Running Claude and ChatGPT side by side feels like redundancy. When both depend on the same Microsoft tenant, it’s two drivers sharing one car.

Write the failure plan before go-live

Well-run businesses already know how to fail gracefully. The till takes cash when card payments go down, and there’s a paper form for when the system is offline. Everyone knows the drill before it’s needed. Agents deserve the same courtesy: a written plan, agreed before go-live, so anyone can predict what the agent will do when its inputs disappear and what it will do when they come back.

  • Make silence measurable. Log every failed inbox check as an error, never as “no new mail”. Send a heartbeat email every hour from an account outside Microsoft 365, and alert a named person if it arrives late or not at all.

  • Check state before acting. Before an agent changes anything, it compares the message timestamp against the record’s last modified time and the activity log. If a human has touched the record since, the agent escalates instead of acting.

  • Name an owner for the kill switch. One person decides when the agent pauses, who handles urgent requests in the meantime, and the conditions for resuming. Write it on a single page and test it.

  • Rehearse the flood. Delay two test orders, resolve one by phone, then release both emails. Pass only if both orders end up correct, with nothing done twice and nothing undone.

Deterministic by default, AI by advantage, resilience by design. An agent that reads email is fine. An agent that writes to your order system on the strength of email, with no reconciliation against the system of record, is a liability waiting for a bad Thursday.

THE QUESTION FOR MONDAY’S LEADERSHIP MEETING

“If our email went down for 36 hours, what would our agents do while it was down, and what would they do when it came back?”

If the room goes quiet, you have your answer, and possibly your next project.

NEWTHISTLE NUGGET

AI GOVERNANCE

An agent that cannot tell silence from failure will treat a broken system as a quiet day.

The Write-Off: The day the inbox went quiet