AI Assistants can now speak scripted utterances at the start of a webhook tool call and during long-running requests. Callers stay engaged while the assistant waits on an external API, database lookup, or third-party model, instead of sitting in silence.
Silence is where voice AI calls die. A webhook tool call that takes 8 to 12 seconds for a database lookup or third-party API leaves the caller staring at dead air, and on a phone call that reads as broken, not . Most voice platforms either go silent and lose the caller, or ask the LLM to improvise filler, which rambles, drifts off-brand, and can't be tuned.
This lets you script the wait-state UX the same way you script the conversation flow. "Let me look that up for you" at the start, "still working on this" at 5 seconds, "almost there" at 15 seconds. Deterministic, tiered, and per-tool. The same control you have over the conversation now extends to the gaps the conversation depends on.
Each filler message has a type
of request_start
(spoken immediately when the tool call begins) or request_response_delayed
(spoken after timing_ms
milliseconds if the response has not yet arrived). The timing_ms
field accepts 100 to 120000 ms and is only required for delayed messages. Once the response arrives, any remaining delayed messages are skipped.
{
"filler_messages": [
{ "type": "request_start", "content": "Let me look that up for you." },
{ "type": "request_response_delayed", "content": "Still working on this.", "timing_ms": 5000 },
{ "type": "request_response_delayed", "content": "Almost there.", "timing_ms": 15000 }
]
}
Learn more in the AI Assistants docs or the Telnyx docs.