Fallback & Degradation Handling

Defines what happens when the primary path fails: retry, route to an alternative model, fall back to a non-AI path, or fail explicitly. Determines whether a provider outage degrades the product or stops it.

Any production deployment, particularly where model calls sit on a user-facing path or inside a time-boxed process.

Batch work with generous windows, where retry alone suffices.

Falling back to a different model must be recorded against the output, because a result from the fallback model was produced under different conditions from one produced normally.

Silent degradation to a weaker model, so quality drops and nothing indicates why. Users experience it as the system getting worse for no reason.

A close-day-three outage where AI-assisted drafting degrades to template-based preparation rather than blocking the close entirely.