Why Plan for It
Every major assistant had multi-hour outages in 2026. Microsoft 365 Copilot was caught in an outage that ran for more than a day from November 12. Claude had two major incidents the next morning. ChatGPT's longest major outage of the year lasted about 6 hours 40 minutes in July. A team that has moved its drafting, research or coding onto one assistant loses that work for as long as the outage lasts, unless it has decided in advance what to do.
Step 1: List What You Actually Use the Assistant For
Write down the five or six jobs that would stop. For each, note three things.
| Job | Can it wait a day? | What it depends on |
|---|---|---|
| Drafting customer replies | No | The chat app and saved instructions |
| Summarising meetings | Yes | The meeting integration |
| Code assistance | Partly | The coding tool and its login |
| Scheduled reports | Depends on the deadline | Automations and connectors |
This is an example. The point is to find the jobs that cannot wait, because those are the only ones that need a fallback.
Step 2: Understand How These Outages Fail
- Login fails while the model works. Authentication was the failing part in outages at Anthropic, OpenAI and Microsoft this year. People who are signed out are locked out.
- The app and the API fail separately. On March 11, 2026 the Claude app was down and the API was not.
- Integrations fail with their host. On November 12, a Microsoft outage broke the Microsoft 365 connectors inside both Claude and ChatGPT while the assistants themselves kept working.
- Scheduled work slips. During Claude's November 13 incident, Anthropic said some scheduled tasks ran more than two hours late or were skipped.
Step 3: Set Up a Second Assistant Before You Need It
A fallback that nobody has an account for is not a fallback.
- Pick one from a different company. If you rely on Copilot, the alternative should not depend on Microsoft's sign-in. If you rely on Claude, pick ChatGPT or Gemini, and the reverse.
- Create the accounts now and have each person sign in once. A free or low-cost tier is enough for an emergency.
- Check your data rules. Decide in advance what may be pasted into the second assistant. An outage is a bad time to make that decision.
Step 4: Keep Your Instructions Somewhere Else
The work a team puts into an assistant, such as system prompts, style guides, templates and reusable instructions, usually lives inside that assistant. Keep a copy in a shared document. During an outage you can paste it into the second assistant and get most of the same behaviour. This also makes it easier to switch for other reasons.
Step 5: Decide Who Calls It
Name one person who checks the status page and tells the team to switch. Without that, each person spends twenty minutes working out whether the problem is theirs. Use the official pages listed in how to check if an AI assistant is down, and route their alerts to a shared channel.
Step 6: The First Hour
- Confirm: check the vendor's status page and the specific product line you use.
- Announce: one message to the team saying what is down and what to use instead.
- Switch the urgent jobs only: move the work that cannot wait to the second assistant. Leave the rest.
- Pause automations: scheduled jobs that fail and retry can pile up or run twice when service returns.
- Do not log out. If you are still signed in, stay signed in. Login is often the last thing to recover.
Step 7: After It Is Over
- Check scheduled work: find what was skipped or ran late and rerun it by hand.
- Check for duplicates: anything retried during the outage may have run twice.
- Note what was painful: the job that hurt most is the one to prepare better next time.
- Read the vendor's explanation, if there is one. OpenAI published 24 incident write-ups in 2026. Anthropic published eight, none after March.
For Teams Building on the API
If your product calls a model directly, the same idea applies in code: a second provider, pinned model versions, and timeouts with backoff. See LLM API uptime and reliability.
How Much Downtime to Expect
At 99.5% uptime, a service is unavailable for about 3.6 hours in a 30-day month. Anthropic's status page showed 99.54% for claude.ai over the 90 days to November 13, 2026, and OpenAI's showed 99.26% for its ChatGPT group. Those hours tend to come in a few long incidents during working hours, not spread evenly. Full figures are in the AI assistant outage tracker.
Sources
Outage facts come from status.claude.com, status.openai.com and Microsoft's admin centre updates, read November 13, 2026. The downtime figure for 99.5% is arithmetic. The plan itself is our recommendation.