Voice-specific behaviourThe same AI. Tuned for spoken delivery.
Voice conversations have different constraints to SMS — shorter messages, no re-reading, spoken booking links. The platform injects a voice instructions block into the system prompt that overrides SMS defaults for every voice campaign.
40-word message limit
Voice responses are capped at approximately 40 words — calibrated for a natural spoken exchange. Contacts can't re-read a voice message, so the AI keeps responses concise and avoids packing in multiple questions.
How it works: Enforced via voice instructions block injected into the system prompt — applies to every voice campaign automatically.
Booking links sent by SMS
When a booking link needs to be delivered during a voice call, the AI sends it as a separate SMS to the contact's phone number rather than reading the URL aloud. URLs spoken over a phone are useless.
How it works: Voice instructions block instructs the AI to send booking links by SMS — no manual campaign configuration required.
Message sanitisation
Before any message is passed to a Say element or encoded into an answer URL, the platform strips markdown, bold markers, URLs, HTML spans, and other non-speech characters — so the AI never accidentally reads formatting syntax aloud.
How it works: Sanitisation runs on every voice message before TTS delivery — the contact only hears clean, natural speech.
Idempotent by CallSid
Every voice call is identified by its CallSid. If a gather webhook delivers a CallSid that already exists in the message table, the pipeline exits without creating a duplicate record — the same idempotency guarantee as SMS MessageSid.
How it works: Duplicate call handling checked before processing — a recovered worker cannot process the same call twice.