ChatGPT logged 102 incidents in the 90 days to 26 October 2026, according to the outage tracker Pulsetic. That is roughly one a day. Eleven were rated major or critical, and the median incident took 1 hour 39 minutes to fix.
Claude had a rougher month. On Sunday 25 October, at 2:05 pm Hong Kong time, every current Claude model failed at the same moment for about thirty minutes. There was no smaller model to fall back to, because all of them were down together.
None of this means AI tools are unreliable toys. It means they behave like electricity, internet and card terminals: usually fine, occasionally not, and always at the worst possible moment. Businesses that rely on them need a simple plan for the bad hour. This guide explains what that plan is, what actually breaks, and how a Hong Kong SME can build one in an afternoon.
What is an AI outage plan?
An AI outage plan is a short, written set of decisions about what your business does when an AI tool stops working. It lists which tasks depend on AI, how long each can wait, what the backup is, and who switches to it. Think of it as a fire drill for ChatGPT, Claude, Gemini or any AI service you pay for.
It is not an IT project. For most small businesses it fits on one page and takes two or three hours to write.
The idea is borrowed from business continuity planning, which larger companies already use for power cuts, typhoons and system failures. The difference is that AI tools are newer, change faster and often sit quietly inside other apps, so many owners do not realise how much depends on them until the day they stop.
How often do AI tools actually go down?
Often enough to plan for. Pulsetic counted 102 ChatGPT incidents between 28 July and 26 October 2026. An IT consultancy tracking Claude counted 44 incidents between 12 August and 25 October. Network firm Ookla found AI app disruption days rose from 6 in the first quarter of 2025 to 51 in the first quarter of 2026.
Most incidents are small: one feature slows down, one model returns errors, or a connector stops syncing for an hour. A few are large.
--- 3 September 2026: ChatGPT, Claude and Grok were disrupted within the same window. Downdetector logged more than 37,000 ChatGPT reports, according to Shattered. OpenAI later said its routing system misread a configuration update.
--- 29 September 2026: claude.ai, the mobile apps, the API, Claude Code and Cowork failed together for about an hour, from 10 pm Hong Kong time. Sign-in failed, and some messages sent during that hour were not saved, according to Deployflow's incident log.
--- 21 October 2026: some ChatGPT users hit authentication errors in an incident that lasted 1 day and 5 hours, rated major on Pulsetic.
--- 25 October 2026: every current Claude model failed together for about thirty minutes.
Analysts expect more, not fewer. Forrester's 2027 predictions, published this month, include a global outage across several industries caused by a bug in AI-generated software, because AI-assisted coding is outpacing testing.
What actually breaks during an AI outage?
An AI outage is rarely just "the AI is down." In 2026 incidents, five different things failed: the model itself, the login system, billing, the connectors to other apps, and even the status page. Each needs a different backup, which is why a plan beats improvising.
1. The model stops answering. You see error messages or endless loading. Switching to a cheaper model from the same company sometimes helps, but on 25 October every Claude model failed at once, so it did not.
2. You cannot log in. On 29 September, single sign-on failed and Anthropic advised signed-in users not to sign out, because anyone who did could not get back in for the hour. A login failure usually blocks the app and the API together.
3. Billing locks you out. On 7 October, a billing fault wrongly paused some Claude organisations at their spending limits, refusing requests across every Claude product. Your account looked broke when it was not.
4. Connectors fail. On 22 October, Claude's connectors to other apps failed for about four hours. The chat worked, but anything that pulled from email, calendars or shared drives did not.
5. The status page itself fails. On 14 August, Anthropic's status page served an invalid security certificate for almost three hours. The one place you check for news was unreachable.
How does an AI outage hit a small Hong Kong business?
The damage depends on what you have handed to AI. A team that only uses ChatGPT for drafting loses an hour of convenience. A shop whose WhatsApp replies, quotations or bookings run through AI loses customers. Time zones matter too: a US morning outage lands in a Hong Kong evening, right when many customers message.
Here are four realistic Hong Kong scenarios.
A restaurant with an AI WhatsApp assistant. The assistant answers table bookings and menu questions. If it fails at 7 pm on a Friday, unanswered messages pile up during the busiest hour. The fix is a simple rule: if no reply within ten minutes, a staff member takes over the chat with three saved replies.
A trading company that drafts quotations with AI. Sales staff paste a client enquiry into ChatGPT to produce a price quote. If ChatGPT is down for two hours, quotes stop. The fix is a plain Word or Excel quotation template stored on the shared drive, plus a second AI tool on standby.
A marketing agency with a 6 pm content deadline. Captions and social posts are drafted in Claude. A thirty-minute model failure is survivable; a one-hour login failure at 5 pm is not. The fix is keeping the agency's key prompts and brand guide in a document outside the AI tool, ready to paste into another one.
An accounting firm that uses an AI connector to read client emails. If the connector fails, the AI chat still works but cannot see the inbox. The fix is knowing which tasks need the connector and doing those manually for a morning.
How do you build an AI outage plan in five steps?
You can build a workable AI outage plan in one afternoon. List every task that uses AI, decide how long each can wait, choose a backup for the urgent ones, write down who switches, and test it once. The whole plan should fit on one page that any staff member can follow.
Step 1: List every task that touches AI. Include the obvious ones, such as drafting emails in ChatGPT, and the hidden ones, such as a booking system, chatbot or accounting app that uses AI behind the scenes. Most owners find three to eight tasks.
Step 2: Ask "what if this stops for four hours?" Sort each task into three groups. Can wait (internal drafts, research). Must continue manually (customer replies, quotations). Must have a second AI tool (high-volume work your team cannot do by hand).
Step 3: Prepare the backup before you need it. For manual tasks, save templates and standard replies outside the AI tool. For tasks that need AI, keep an account with a second provider and store your main prompts in a shared document. The 3 September outage barely touched Gemini, so a second provider on different infrastructure is a real safety net.
Step 4: Name a person and a trigger. Decide who declares "we are switching," and when. A simple trigger works: if the tool fails for 15 minutes during business hours, switch. Bookmark the official status pages, and also a third-party tracker, because status pages can lag or fail.
Step 5: Run a fifteen-minute drill. Pick a quiet afternoon, pretend your main AI tool is down, and ask staff to follow the plan. You will quickly find the missing template or the colleague who does not have a backup login.
Two habits help on the day itself. Tell staff not to sign out of AI tools during an incident, which is exactly what Anthropic advised during the 29 September login failure. And copy important AI output into your own files, because some messages sent during that outage were lost.
What do business owners get wrong about AI outages?
The most common mistake is assuming a paid plan or a big brand means the tool will always work. In 2026, outages hit free and paid users alike. Other myths include trusting the status page completely, assuming a smaller model is a backup, and believing a backup plan is expensive.
"I pay for the business plan, so I am protected." Paid plans buy more usage and better data terms, not immunity. The 29 September Claude outage hit every surface, paid or not. Check whether your plan includes any uptime commitment at all; many self-serve plans do not.
"If the status page is green, the problem is on my side." Not always. Incident logs show status posts sometimes appear late, and on 14 August the status page itself was unreachable. Check a third-party tracker and ask a colleague on a different network.
"I can just switch to the smaller model." Sometimes. But on 25 October every Claude model failed together, and login failures block every model at once. A real backup is a different provider or a manual process.
"A backup will double my AI bill." Usually it will not. Many AI tools have free tiers that are enough for a few emergency hours, and templates and saved prompts cost nothing. For a related cost trap, see our guide to AI usage limits, which can stop a tool even when the provider is fine.
Frequently asked questions about AI outages
These short answers cover what Hong Kong owners most often ask about AI downtime: how to check if a tool is really down, whether to pay for two AI subscriptions, how outages differ from usage limits, and what to tell customers when an AI assistant stops replying.
How do I check if ChatGPT or Claude is really down?
Open the official status page, then a third-party tracker such as Pulsetic or Downdetector, and ask a colleague to try on a different network or device. If all three show problems, it is an outage. If only you are affected, restart the app and check your plan's usage limit.
Should a small business pay for two AI subscriptions?
Only if a task cannot stop for a few hours and cannot be done by hand. For most small businesses, one paid plan plus a free account with a second provider is enough. Keep your key prompts in a shared document so anyone can move them across in minutes.
What is the difference between an AI outage and a usage limit?
An outage is a fault on the provider's side that affects many users at once. A usage limit is a cap on how much one account can use in a time window. Both stop work, but an outage needs a backup tool, while a usage limit needs a bigger plan or better spreading of work.
What should I tell customers if our AI assistant stops replying?
Be short and honest. A message such as "Our system is slow right now, a team member will reply within the hour" protects trust far better than silence. Prepare this text in advance in both Chinese and English so staff can send it immediately.
Why should every AI-using business have an outage plan?
Because AI tools now fail about as often as other cloud services, and the cost lands on whoever has no backup. A one-page plan, a few saved templates and one practice drill turn a lost afternoon into a fifteen-minute inconvenience. It is cheap insurance for tools your customers now depend on.
AI outages are not a reason to avoid AI. They are a reason to use it like a grown-up business: with a backup, a named owner and a habit of testing. If you want to go further, read our guides on AI model concentration risk and AI vendor lock-in, which explain how to keep the freedom to switch providers.
The businesses that win with AI are not the ones whose tools never fail. They are the ones that keep serving customers when they do. We understand AI. UD stands with you.
Reviewed by the UD AI team.
Not sure how much of your business already depends on AI, or where a backup would matter most? Take the free AI Ready Check and we will walk you through it step by step, from mapping the tasks your team hands to AI today to a practical plan for the hour your tools go quiet.