What is committed, how it is measured, and what happens when it is missed. Availability is defined on call connection and answer, because that is the failure with a direct revenue consequence. Response times are tiered by severity, and credits are automatic rather than requiring a claim.
Synthetic calls placed against every production number on a fixed interval, plus live traffic sampling. A call that fails to connect counts as unavailable. A call that connects but exceeds the latency threshold counts as degraded on a sliding scale, because a four-second pause is a real failure even though the call technically worked.
Customer-side failures — an expired CRM credential, a phone carrier issue upstream of us — and scheduled maintenance announced in advance. Upstream model provider degradation absorbed by the fallback chain is excluded; degradation that reaches the caller is not, which is the distinction that matters.