When the AI tool is down: a continuity plan for a small office
· By Bee Flow · 5 min read
On 18 November 2025 a Cloudflare outage made a large part of the web unreachable for hours, ChatGPT's login included. What stops when one link fails, and what to arrange beforehand: a second model, a small local model and a one-page plan.
![]()
On Tuesday 18 November 2025, at about 11:20 UTC (12:20 in the Netherlands), Cloudflare's network started failing to deliver traffic. A large share of the websites on the internet route their visitors through Cloudflare, so anyone opening one of those sites that afternoon was likely to get an error page. According to Cloudflare's own post-mortem, a change to the permissions of one of its database systems made a configuration file for bot detection twice its normal size, and the software that reads that file failed. Cloudflare states it was not an attack. Core traffic was largely flowing again by 14:30 UTC, and by 17:06 UTC all systems were back to normal.
AI services felt it too. OpenAI's status page logged login problems for ChatGPT and its API platform that day, caused by "one of our third-party service providers". Anyone who wanted a quote request summarised or a customer email rewritten at that moment had to do without for a while.
A few hours without AI help is not a disaster. It is a good moment to write down what stops at your office, and what you do when it does.
There are more links between your staff and the model than you think
When someone in the office asks an AI chat a question, that question passes along a chain. The browser or app on the laptop. Your own internet connection. The provider's website, often behind a network service such as Cloudflare. The login step, sometimes through your company's Microsoft or Google account. Only at the end of the chain sits the language model itself.
On 18 November the failure was in the middle of that chain. The model did not have to be broken to be out of reach. That is the first lesson for a continuity plan: a fallback only helps if it does not run through the same link. A second AI service sitting behind Cloudflare as well probably had the same problem that afternoon.
The same goes for integrations. An Outlook plug-in, an automation that summarises incoming mail, a script that reads invoices: these usually hang off one provider and go down together with it.
Start with a list of what depends on AI
Take half an hour and write down, task by task, where AI is used today. Next to each, note what happens if it is gone for an afternoon, and what happens if it is gone for a week. You will usually end up with three kinds of tasks.
- Tasks where AI is handy but the work goes fine by hand, such as polishing an email.
- Tasks that slow down without AI, such as reading through quote requests or writing up meeting notes.
- Tasks that stop, because a process was built on top of them, for example automatic sorting of the service mailbox.
A made-up example. At Example Installations Ltd, the planner has the new fault reports summarised every morning. If that drops out, he reads them himself. It costs an extra hour, and the work carries on. The colleague who feeds thirty quote requests a day into the order system through an AI integration is in a different position. That task is worth a fallback.
A second model as a fallback
For the tasks in the second and third group, arrange a second provider. Beforehand, not on the day. That means an account that works, a data processing agreement if personal data goes in, and colleagues who know how to log in.
Also look at how that second provider routes its traffic and where it runs. A European provider with its own data centre and a different network gives you more spread than two services behind the same network supplier. A provider's status page and documentation often tell you more than you would expect.
Test the fallback once with real work. A prompt that works well with one model sometimes gives a very different result with another. So keep your best prompts and instructions in an ordinary document, not only in the chat history of one tool. That way they come with you to any fallback.
A small local model for the basics
Part of the daily work does not need a large model. Shortening a text, rewriting an email, translating or summarising a page: a small open model running on your own computer or server can do that, with software such as Ollama or llama.cpp. Such a model needs no connection to a provider and keeps working when the internet stumbles.
Keep your expectations modest. On ordinary office hardware a small model is slower, and it is weaker at hard reasoning and long documents. For the basic tasks on your list it is often enough, and the text stays in the building. Let one colleague use it for a few weeks before the whole office relies on it.
![]()
What to write down beforehand
A continuity plan for AI fits on one sheet of paper. This belongs on it.
- Who decides to switch to the fallback, and after how long an outage.
- Links to your providers' status pages, so you can see quickly whether the problem is theirs or your own connection.
- Where the fallback's login details are kept. In the password manager, not in someone's head.
- Which data may go to the fallback. The privacy rules that apply to your usual tool apply there too.
- Where the saved prompts and instructions are.
- How the work is done by hand, for tasks that have no fallback.
- How colleagues hear that there is an outage and what they do then. A message in the team chat is enough.
After an outage, hold a short review of fifteen minutes. What went well, what took too long, what was missing from the sheet. Then update it. We would also walk through the sheet once a year without an outage, because providers and integrations change faster than you think.
How Bee Flow helps with this
In Bee Flow your team works with models from several providers in one workspace, and an administrator decides which ones are available. If one provider goes down, you move an assistant to a model from another provider, with the same instructions and the same knowledge base. That knowledge base of your own documents lives in Bee Flow itself and cites its source with every answer. Bee Flow can run in a European cloud or on your own server, so you choose the route your team's work takes. The page on hosting and deployment sets out the options. If you would like to know how that works out for your office, get in touch.