Politeness is the product, and most of it is built in: every agent the Builder creates carries a floor of behaviour you can't accidentally remove. This page explains those defaults and the small part that's yours to do.
This page describes product behaviour, not legal advice. Rules on AI disclosure, call recording and outbound calling vary by where and how you operate — check your own obligations.
Callers always know it's an AI
An agent's first words disclose that it's an AI, woven into a warm greeting rather than bolted on. Agents never impersonate a person, borrow a real staff member's name, or clone a voice. Callers who know what they're talking to trust what it tells them — the full set of guarantees is in What every agent promises your callers.
Card numbers: never spoken, never keyed
Agents will not take payment card details on a call — not by voice, and not by keypad either, because keyed digits land in the call record like everything else. This is a compliance stance, not a gap: the agent routes payment to a secure channel you define — a human, your payment line, or an emailed link — and if a caller tries to hand over card details anyway, the attempt is flagged so you can follow up properly.
A human is one ask away
Every team the Builder designs includes a defined path to a human — see Teams, transfers & subagents. When the agent doesn't know something, it says so honestly and offers an alternative in the same breath. A caller in distress is signposted to the right help; an abusive caller gets one warning, then a polite end to the call.
Recording, with callers in the know
Recording is optional and per-agent, with a per-number override — when it's on, recordings are stored encrypted and you can review any call afterwards. The agent's generated instructions include a recording notice so callers hear about it up front. Keep that notice: telling callers a call is recorded is your responsibility as the operator, and the notice does the telling for you.
Outbound calls people actually want
polite.ai is inbound-first by design. There is no dialler and no scheduler in the platform — an outbound call happens only when your own system requests one through the API, and every destination passes an outbound filter before it's dialled, so an agent can never be talked into ringing a number you didn't sanction.
Use outbound where the recipient expects the call and the agent arrives with context: reminders for tomorrow's appointments, callbacks promised on an inbound call, order-ready notifications. Cold-calling strangers with an AI is neither what the product is built for nor a good idea. One honest limitation to design for: agents can't reliably detect answering machines, so tell your agent what to do if the call sounds like voicemail — leave a short message and hang up.