I would not chain daily operations to one AI provider because Notion once showed how fast access can wobble. Users do not care about vendor negotiations when a client deliverable is due.
That is when your team discovers the feature they treated as infrastructure was actually a favor from someone else's contract.
Disruptions are normal now
Model access, rate limits, policy changes, regional blocks — all happen.
If your workflow assumes infinite API uptime, you will eat the outage in public.
Build fallbacks that humans can execute
Keep a secondary model ready, an offline template for the worst case, and a manual fallback doc for the tasks that cannot wait until engineering ships a patch.
Name one owner who can flip switches without a committee waiting on three approvals and a calendar invite.
Communicate early
If customer-facing features break, say so plainly.
Silence makes users assume your product is broken, not the upstream pipe.
Do not over-promise AI reliability in sales
Set expectations like any other dependency.
Notion restoring Anthropic access is a happy ending, not proof outages are rare.
Treat AI like electricity: mostly on, sometimes not, and your business needs a flashlight.
