How are you handling sudden changes to hosted AI services?

Start with an inventory, not a blanket plan to replace everything. Mark which workflows stop if one provider disappears, who depends on them, and what the manual fallback is. Then test that fallback. A backup that has never been used is mostly a hope.

I’d separate tasks by how much capability they actually need. A basic classifier or draft generator may work with a different, cheaper model, while a complicated reasoning task may not have a close substitute. Route simple work flexibly, and be honest about where a fallback means reduced quality rather than equivalent service.

For critical systems, put a layer between your application and the provider so you can switch endpoints without rewriting the whole product. Keep prompts, tool permissions, and output checks under your control, and regularly test another provider. That helps with outages, but it doesn’t guarantee the replacement will behave the same.

Local models can be useful for tasks that fit their capabilities, especially if connectivity or data handling is the main concern. They are not a universal escape hatch. Check the hardware, maintenance burden, and quality on your actual workload before counting one as a production fallback.

There’s also a governance question here, not just uptime. Before putting a service into a consequential workflow, review its data retention terms, where it can be used, and what happens if access is restricted. For some low-impact experiments, accepting the risk may be simpler than building elaborate redundancy.