When a model goes dark on no notice, redundancy lets you switch providers. It does not restore the decision-maker you quietly outsourced. That is the line between using a model and depending on one.
A frontier model goes dark on no notice. No migration window. No transition plan. No SLA covers the reason it happened. If your operation goes blind when it does, the lesson is not the one most people draw.
The common read is: diversify, build redundancy, never depend on one model. That is correct, and it is the easy half. It is a procurement fix. The harder half is governance.
Redundancy lets you switch the model. It does not restore what you actually lost. Because a sudden outage answers one question for you: when the model stopped, did you lose a tool, or did you lose the decision?
If a human still owned the decision the model was making, you lost a tool. You switch providers, you slow down, you recover. If no one did, you did not lose a vendor. You lost the decision-maker you had quietly outsourced, and no failover brings that back. You cannot fail over to judgment that is no longer in the building.
That is the line between using a model and depending on one. Using it means a person still owns the call and could make it without the machine. Depending on it means the machine became the owner and no one signed for the transfer.
Switching models is resilience you can buy. Owning the decision is the part you cannot.
AI operates. You own the decision. An outage only reminds everyone which half they skipped.