AI Outage Plan: What to Do When ChatGPT or Claude Goes Down

"If ChatGPT or Claude goes down at ten in the morning, what is my team actually supposed to do?" Few owners had a crisp answer to that before September 3, 2026, when ChatGPT, Claude and Grok all confirmed outages within the same three-hour window. If your answer is still "wait and refresh the status page," your business does not have an AI outage plan. It has a hope.

An AI outage plan is a short, rehearsed procedure that tells your team which AI-dependent work stops, which work moves to an approved fallback, and which work reverts to a manual process when a primary AI tool fails. It names who calls the incident, what data may enter the fallback, and how work resumes afterward.

What happened on September 3, 2026?

Three of the most widely used AI platforms reported problems on the same morning, and none of them pointed to a shared cause. According to SoftwareReviews' analysis of the overlapping outages, Anthropic began investigating elevated errors at 9:26 a.m. Eastern, xAI's Grok had problems from 9:30 a.m., and OpenAI opened an incident for ChatGPT and Codex at 10:43 a.m. Third-party monitoring also showed a shorter spike in Gemini API failures, though Google did not post a matching incident.

A Cloud Security Alliance research note on the triple outage concluded the evidence points to overlapping, independently caused incidents, not one shared cloud or network failure. That is the uncomfortable part. Nothing exotic had to go wrong. Three normal bad mornings landed on the same day, and every team that had built a workflow around one of those tools stopped at once.

Your backup AI tool may not be a backup

Most teams that think about this at all assume the fix is simple: if ChatGPT is down, use Claude, and vice versa. September 3 showed why that instinct is weaker than it looks. When several providers are degraded at the same time, the tool you planned to switch to may be failing too.

There is a second problem. SoftwareReviews noted that the outages may have been unrelated, or that traffic shifting and automated retries may have pushed extra load onto alternative providers just as they became everyone's backup. When thousands of organizations fail over to the same second choice at the same moment, that second choice inherits a surge it was not sized for. A fallback that only works when nobody else needs it is not much of a fallback.

Even switching within the same vendor rarely helps. When Anthropic reported elevated errors on September 29 across Claude.ai, the apps, the API, Claude Code and Cowork, a SoftwareReviews vendor roundup pointed out that moving from the web app to the API or another Claude surface would not have restored service. A second issue also blocked many sign-ins, and some messages sent during the incident may not have been saved. The practical takeaway: a real fallback needs a different provider and a different sign-in path, and the work in progress needs to live somewhere other than inside the tool that failed.

Are AI outages becoming more common?

The data says disruption has risen as adoption has scaled. Ookla's analysis of 471 days of Downdetector data, reported by Mobile World Live, counted 3.7 million user-reported problems across ChatGPT, Claude, Gemini, Copilot, AWS and Azure between January 2025 and mid-April 2026. High-signal disruption days, defined as a day when a service logged more than ten times its own median report volume, rose from six in the first quarter of 2025 to 51 in the first quarter of 2026.

Claude accounted for 39 of those 51 days, while ChatGPT's underlying reliability trend actually improved over the period. The lesson is not that one vendor is bad. It is that reliability across the category moves with demand, and the platform that looks stable this quarter may be the one under strain next quarter. Your plan should not depend on predicting which provider fails next.

Some outages are not AI failures at all. The AWS post-event summary for the October 2025 US-EAST-1 disruption traced it to a latent race condition in DynamoDB's automated DNS management, with impact running from 11:48 p.m. Pacific on October 19 to 2:20 p.m. on October 20. Our breakdown of what that AWS outage meant for Ohio businesses shows how a cloud fault several layers down can take out tools that have nothing obvious in common.

The hidden cost: what employees do when the approved tool disappears

The biggest risk during an AI outage is often not lost time. It is improvisation. When the sanctioned tool stops responding mid-task, the fastest path for an employee under deadline is whatever free AI tool is open in another tab, with no review of what client data just went into it.

That behavior is already common on normal days. A PagerDuty survey conducted by Wakefield Research of 1,250 office professionals at large organizations found that two-thirds had used AI tools at work they believed were not permitted under company policy. An outage removes the one thing keeping people on the approved tool, which is that it works. We have written about why shadow AI is a board-level financial risk, and an unplanned outage is one of the most reliable ways to create it.

This is why an outage plan has to include data rules, not just tool names. If your fallback is a different AI provider, your team needs to know before the incident which data may go into it and which may not. Our never-paste list for general AI tools is a good starting point for drawing that line.

How to build an AI outage plan your team will actually use

You do not need a disaster recovery department. You need a one-page plan per critical workflow, written in plain language and practiced once. From an infrastructure standpoint, these are the elements that make the difference on the day:

1. Rank workflows by what breaks. Separate nice-to-have AI use, such as drafting internal emails, from work where an outage stops revenue or client service, such as intake, quoting, support triage or scheduling. Only the second group needs a full plan.
2. Define degraded mode for each one. Decide in advance whether the work queues until service returns, moves to an approved secondary provider, or reverts to a documented manual process. Write down which.
3. Pre-approve a fallback on different infrastructure. Pick a secondary tool from a different provider with a separate login, confirm it handles your real prompts, and keep the account active. State what data may go into it.
4. Keep work outside the tool. Store prompts, templates, drafts and task status in your own systems, so an outage costs a pause, not lost work.
5. Name an incident owner. One person watches the vendor status pages and your own checks, declares the outage, tells the team which mode they are in, and calls the all-clear.
6. Rehearse once and review after every real incident. Run the plan for an hour on a quiet day, fix what breaks, and update it after each actual outage.

Most of this is ordinary continuity planning applied to a new dependency. If you already follow the key steps to a business continuity plan that works, AI outages belong in the same document, and the disaster recovery testing strategies you use for servers and backups apply here too.

How long should an AI outage plan assume a tool will be down?

Plan for at least half a business day. The September 3 AI incidents each ran roughly two to three and a half hours, while the October 2025 AWS disruption stretched across more than fourteen hours from start to finish. A plan built around a 20-minute blip will fail during the outages that actually cost money, so size your manual fallback for a full morning.

The cost of not planning is easy to underestimate because AI outages feel like inconveniences. Multiply the people idled by the hours lost, then add the client commitments that slip, and the number climbs quickly. Our analysis of the real dollar cost of downtime for SMBs is a useful way to put that figure in front of leadership. An AI readiness assessment is another way to surface which workflows would stall first.

How Securafy helps when the AI tool goes dark

We treat AI platforms like any other critical dependency. That starts with identifying the three or four workflows where an outage would stop real work, then building the degraded-mode procedure for each: the secondary provider, the separate sign-in, the data rules for the fallback, and where in-progress work is stored. We set up independent health checks for the AI services you depend on, so your team hears about trouble from your own monitoring instead of a confused client.

Then we test it. A short tabletop or live switchover on a quiet day shows whether the fallback really works with your prompts and permissions, and whether staff know which mode they are in. If you also want to tighten the contract side of vendor dependence, our guide to building an exit plan for your AI vendor covers the clauses that support it.

Where To Go From Here

The next AI outage is not a question of if, and the teams that lose the least are the ones that decided in advance what happens when the tool goes dark. Pick your most AI-dependent workflow and write its one-page plan first.

If your team is moving faster with AI than your guardrails are, start with structured training rather than another tool. AI University gives your people role-based AI training with security built into the material, not bolted on afterward.

If you would rather talk through your specific environment first, book a strategy call with Securafy and we will walk your current AI usage, exposure, and the fastest path to safe adoption.