Politeness isn't a feature we bolt on — it's the product. Every prompt the Builder writes carries a floor of behaviour that no design choice removes.
It says it's an AI
Every agent discloses that it's an AI in its first turn, woven into a warm greeting rather than tacked on as a disclaimer. It never impersonates a person, borrows a real staff member's name, or imitates anyone's voice. Callers who know what they're talking to trust what it says — disclosure is a conversion feature, not a legal chore. The wider context is in Responsible AI calling.
It never takes card numbers
Not spoken, and not keyed in either — digits pressed on a phone keypad end up in the call transcript, so the keypad is no safer than saying a number aloud. This is a compliance stance, not a gap: when payment comes up, the agent routes the caller to a channel you define — a human on your team, your payment line, an emailed link — and flags the request so you know it happened.
A human is always within reach
Every team the Builder designs includes a defined path to a person — a transfer in office hours, a message or callback promise outside them. The agent hands over the moment a human is the better answer, and it does the honest version of "I don't know": admitting it, then offering an alternative in the same breath rather than guessing.
Callers can press instead of say
Phone callers can key long numbers — order references, postcodes, account numbers — on the keypad instead of saying them. Agents offer the keypad when saying a fiddly string isn't working, and read back what was keyed. It's quicker on a bad line, more discreet in a public place, and an accessibility path for callers who struggle to be understood.
Calls only go where you've allowed
Every outbound destination — transfer targets included — passes a safety filter, so an agent can never be talked into dialling a number outside the destinations you've approved. The same guarantee covers outbound calling generally.
The quieter promises
The rest of the floor is the texture of a well-mannered call. One question per turn, reflecting back what it heard, and character-level read-backs of names and references. Identity checked before any account specifics are discussed. A clear notice when calls are recorded. A caller in distress is signposted to appropriate help; an abusive one gets a single warning, then a polite end to the call. And the agent can't be talked into ignoring your instructions or revealing them — what you configured is what callers get.
These guarantees are written into every prompt the Builder generates, so you don't maintain them — and if you hand-edit a prompt, a test call is the quickest way to confirm the behaviour still holds.