When the Primary Voice Route Fails, the Caller Still Needs an Answer

A carrier may accept a call while the speech service is degraded. The voice application may be healthy while its CRM is timing out. A human queue may be online and already beyond capacity. Good routing asks whether the destination can complete this caller’s job now—not merely whether an endpoint answered a health check.

Define routes for inbound and outbound purpose, market, language, operating hours, caller or account state, disclosure and consent status, skill, queue capacity, provider health, and current incident. Make precedence explicit. A returning customer with an open case should not be sent to the same prospecting flow because the phone number also appears on a lead record.

Set limits by operation. A transient status read may be safe to retry; a calendar insert, outbound dial, or CRM update needs an idempotency strategy before retrying. Use exponential backoff where the provider recommends it, add jitter across many workers, and stop when the remaining call time makes recovery unrealistic.

Tell the caller when the system is having trouble and offer a next path. Silence followed by the same greeting is not recovery. Prevent loops with a route history, attempt count, and terminal states. Never transfer a caller back to the component that just failed without a changed condition.

Record provider call ID, application correlation ID, parent and child calls, route decision, health inputs, queue, tool operations, webhook events, transfer target, and CRM activity. Twilio exposes call status transitions and callbacks; OpenTelemetry provides traces, metrics, and logs that can be correlated across distributed components.

Validate webhook signatures, deduplicate delivery, and retain both event and observed timestamps. A delayed callback should not reverse a newer authoritative state. Redact sensitive content before exporting telemetry and avoid putting full transcripts, authentication answers, or phone numbers into general-purpose logs.

Disable the primary speech provider, slow the CRM, reject a webhook, exhaust a queue, remove an agent skill, break the calendar, and simulate a regional carrier problem. Confirm detection time, route time, caller message, context transfer, duplicate prevention, recovery, and alert ownership.

Measure route success by destination, time to first useful audio, transfer success, callback completion, repeat callers, loops, abandoned calls, duplicate side effects, and time to recover. Review the calls people experienced during an incident. Infrastructure uptime can be green while callers are still trapped in a technically available system.