When a call fails, identify the first broken boundary before blaming the telephone carrier or AI service. A caller hearing silence can result from transport, media, runtime, or downstream workflow problems.
Incident separation needs boundary evidence
Carrier incidents affect call setup, routing, signalling, or media transport. AI-provider incidents can affect recognition, generation, or speech output, depending on the architecture. Application and business-service failures form additional categories and should not be forced into either provider bucket.
For a proposed diagnostic view, correlate call ingress, media establishment, agent connection, dependency requests, and business operation results. Keep the supported provider and route metadata with each stage. Provider status pages can help, but the local evidence determines the affected workflow.
Do not infer two-way audio from a connected call event alone. Likewise, a successful model response does not prove the caller received speech through the telephone path.
Compare fictional outage symptoms
In one incident, inbound calls never reach the application, while direct agent tests work. The investigation starts with telephone routing and ingress. In another, calls establish media but speech generation requests repeatedly fail. That points to a different boundary.
A third case has normal conversation and a failed CRM write. Calling it an AI outage would send responders toward the wrong system. Preserve the distinctions even if callers describe all three as “the receptionist is broken.”
| Observation | Next boundary to inspect |
|---|---|
| No inbound application event | Number routing and carrier ingress |
| Connected call, no usable media | Audio transport and media configuration |
| Media present, recognition unavailable | Speech input dependency |
| Agent answers, speech not delivered | Speech output and telephone playback |
| Conversation succeeds, record absent | Tool and business integration |
Use controlled synthetic checks for separate boundaries where the deployment supports them. A single full-call probe can detect impact but may not localise the cause.
Avoid false certainty during diagnosis
Several dependencies can fail simultaneously, and one failure can cascade into another. Report observed symptoms and the current hypothesis separately. Record timestamps with a common time basis so the timeline is interpretable across services.
More probes increase coverage and cost. Prioritise checks that change the responder’s action and ensure they do not contact real customers or mutate live records unintentionally. A provider incident does not remove the need to handle stranded operations.
This diagnostic method is proposed. The voice platform case study describes work across telephony, agent workflows, and backend integrations. See outage runbook design for containment and recovery.
Provider incident FAQ
Does a green status page prove the provider is healthy?
No. The affected route, account, region, or feature may differ from the provider’s broad status report.
Can a successful direct agent test rule out carrier problems?
It helps isolate the agent path, but it does not exercise telephone ingress or media.
What should the first incident update say?
Describe confirmed caller impact, affected boundaries, current containment, and the evidence still being gathered.