What to Do When Your AI Tools Go Down: An Outage Playbook

When an AI tool your business depends on goes down, switch to a fallback you chose in advance, tell the people affected, and keep working in a slower mode instead of waiting for it to come back. An AI outage used to be a small annoyance. Now that drafting, support replies and data processing run through a model, it can stop a whole part of the operation. The businesses that handle one calmly decided ahead of time what they would do instead. An outage playbook is that decision, written down before you need it.

First, know what you actually depend on

Before you can plan for an outage, list the AI and SaaS tools your operation can't run without for a day: the ones where "it's down" means real work stops. Separate the load-bearing from the merely convenient. You only need fallback plans for the first group. Most teams find it shorter than they feared, plus one or two tools they hadn't realized were critical until they wrote them down.

Count the AI features inside other products too. The Readiness Drill's playbook tells buyers to list load-bearing AI dependencies as systems in their own right: a support copilot, a product feature, or an ops workflow that routes through a single model or API. In the drill's shipped sample, a support copilot on a single LLM API scores 66, inside the GAPS band, and still reads WOULD NOT RECOVER because it has no fallback.

The outage playbook, tool by tool

1. Name the fallback. For each critical tool: if it's down for an hour, a day, a week, what do we do instead? The drill's runbook lists the usual options: a second provider, a cached last-known-good response, a queue-and-retry, or a hand-off to a person.
2. Decide the trigger. How long do you wait before switching? Ten minutes of a flaky tool isn't an outage; two hours down is. Pick a threshold so nobody agonizes in the moment.
3. Assign the calls. Who confirms the tool is down (and not your own internet), who switches to the fallback, and who tells the team and customers.
4. Keep serving customers. Have a plain message ready: what's delayed, what still works, and when you'll update next. Customers tend to forgive an outage handled openly far more readily than silence.

A second model is not a drop-in copy

Switching to another AI provider sounds like the clean fallback, and it is the one most likely to surprise you. A different model formats output differently, follows instructions more or less strictly, and refuses different things. Anything downstream that was tuned to the first model can break without an error. Our post on why AI model upgrades break production workflows walks through how that happens, and model-agnostic AI architecture covers how to keep a workflow portable enough to switch.

A manual fallback has the opposite problem. It works, but a support team that leans on an AI assistant for most replies may not keep up by hand for long. Sometimes the honest fallback is a pause, a queue and a message to customers. For some tools none of the options is good, and the plan is knowing how long you can wait. The drill's runbook adds one more check: compare the vendor's SLA with your recovery-time objective, because a system can't hit a two-hour recovery target if a critical provider doesn't commit to one.

Write it down and keep it reachable

A playbook stored on the system that's down is useless. CISA's #StopRansomware Guide makes the same point about incident response plans, telling organizations to ensure "a hard copy of the plan and an offline version is available." Keep yours as a printed copy, a separate note, or a file on your phone. Revisit it as your tool stack changes, since AI dependencies tend to pile up faster than anyone documents them. If one vendor already sits under most of your workflow, the signs of AI vendor lock-in are worth a read before the next outage.

Pairs well with

The AI Vendor Reliability & Spend-Justification Scorecard ($79) grades each AI or SaaS vendor on uptime, incident load and support alongside value, and a business-critical vendor that fails any reliability dimension reads DO NOT RENEW. The Shadow AI Discovery & Risk-Triage Kit ($69) surfaces the AI tools your team uses, including the ones nobody listed as dependencies. The AI Reliability Bundle ($329) includes the Prompt Regression Lab, which snapshots a baseline and flags regressions, a way to check a fallback model before you need it.