Diagnose an Answered but Silent AI Phone Call

An answered but silent AI call is not one single failure. The telephone connection may have been established while audio never reached the assistant, the agent may not have produced speech, or generated audio may not have reached the caller. The first task is to locate the missing step.

Avoid immediately rewriting the greeting. That may alter the conversation without addressing a routing or media problem. Start with one exact failed attempt and preserve its identifiers so every person investigates the same call.

Clarify who heard what

Write a plain description before opening logs. Could the caller hear anything? Did the assistant appear to receive the caller's words? Was silence present from the start or only after a transfer or pause? Did another participant hear the assistant?

These observations distinguish no audio from one-way audio and delayed speech. A hypothetical report saying “the system was silent” becomes much more useful as “the caller heard no greeting, but the saved transcript contains the caller's first question.” Even that transcript is only one piece of evidence; it does not prove the caller received audio.

Note the handset, speakerphone or headset where relevant. Do not attribute the problem to the user's device without a controlled comparison.

Separate signaling from media

SIP signaling establishes and controls the connection; audio travels through the media path. A connected status therefore narrows the investigation but does not establish audible conversation. LiveKit's SIP troubleshooting guide separates connection failures from media issues.

Ask the technical owner to check the negotiated media information and whether audio packets flowed in each direction. Twilio's trunk troubleshooting guidance identifies address and firewall problems among possible causes of one-way audio. Those are hypotheses to test, not a reason to change every network rule.

Do not paste raw packet captures or credentials into a broad support conversation. Share the narrow identifiers and observations needed by the authorised owner.

Establish whether the agent spoke

If incoming audio reached the session, inspect whether the assistant actually initiated a response. A missing agent, an admission failure or a provider error can leave a connected telephone participant without useful speech.

If speech was generated, find evidence that it was published to the intended session. A text answer in history is not identical to an audio response. Likewise, a provider reporting that it produced audio does not demonstrate successful playback at the far end.

This chain gives a practical decision tree: no incoming media suggests a route investigation; incoming speech but no response suggests an agent/provider investigation; generated output without caller audio suggests the outbound media and playback path.

Preserve the earliest symptom. A later retry that fails differently may be a second problem rather than confirmation of the first theory. Keep each attempt in its own row and connect it only to evidence carrying the same identifiers.

Compare with a useful control

A control should change one meaningful factor. Try the same supported assistant through a known working route, or compare a known working assistant on the suspect route. Keep the model, instructions and test request stable where possible.

A browser success can narrow the issue toward the telephone-specific path, but it does not prove the carrier is responsible. The browser uses different components. Burki's incoming-call testing guide helps organise route-specific evidence.

Do not run an uncontrolled series of paid calls while changing several settings. Define a bounded test, an expected observation and a stop condition before repeating it.

Record the demonstrated fix

Once a change appears to resolve the problem, repeat the relevant acceptance case and verify both directions of speech. Include the opening, an ordinary exchange and a clean ending. If the original failure happened after a specific transition, exercise that transition again.

A useful incident note states what was observed, what mechanism was demonstrated and what remains uncertain. Burki's SIP integration guide provides context for the route. Start with the missing audio step, and let the evidence determine which component needs attention.