A voice agent has about ten seconds to tell a person what it is, why it is calling and how to make it stop. If that moment feels like a recorded menu, the rest of the conversation is already uphill.
Consent is usually treated as a legal checkbox at the end of a flow. In voice, it is an interface problem before it is a legal one. The caller cannot scroll, read a footnote or click "I agree". They can only listen, interrupt or hang up. The opening has to earn the next sentence.
The first ten seconds
Our default opening does four things in one turn: it identifies the agent and the organisation, states the purpose, discloses recording, and offers a human. It also leaves room for the caller to interrupt. A disclosure that cannot be interrupted is a monologue, not a conversation.
“This is Aldaria calling for a client. This call is recorded. You can ask for a person at any time, or say stop and we will end the call. Is now a good time to talk about your appointment on Thursday?”
The opening is the consent moment. It is not a menu, and it is not a disclaimer hidden behind a keypress.
Four things the opening must do
- Identify. The agent names itself and the organisation it represents. "A virtual assistant" is not an identification.
- State the purpose. The caller should know why the call is happening before being asked for information.
- Disclose recording. If audio or the transcript is retained, the caller is told at the start, in plain language.
- Offer a human and an exit. The caller can ask for a person or end the call at any point. The agent does not make them repeat the request.
Designing the moment
Disclosure should be short enough to hold attention and complete enough to be meaningful. We write it as speech, not as legal text. That means short clauses, no nested conditions, and a deliberate pause after the organisation name so the caller can orient.
The agent must also be able to hear an interruption during the disclosure. If the caller says "who is this?" or "stop", the agent stops and responds. A system that finishes its compliance script while the caller is asking to be removed has misunderstood the point.
When the caller gives consent, we record the artifact: timestamp, channel, language, script version, the number dialled and the exact disclosure played. The record is not a checkbox in a CRM. It is evidence that can be reproduced.
Outbound is different
Inbound calls begin with the caller's intent. Outbound calls begin with ours, which raises the bar. Before a number is dialled, we check the lawful basis, the suppression list and the calling window. The caller ID must be real and traceable. The opt-out must take effect immediately, not after a nightly sync.
Voice clones add a separate consent. The person whose voice is used must give specific, written and revocable consent. The clone is tied to a scope: which scripts, which campaign, which period. A generic "I agree to the use of my voice" is not enough for a campaign it never contemplated.
What we prohibit
Some uses are out of scope regardless of consent. We do not build emergency services, political persuasion, unsolicited robocalls, impersonation of a real person, harassment or coercive debt collection. A client request that depends on deception is not a design problem we will solve.
The audit trail
The consent artifact is only useful if it is connected to the call and the version of the flow. We store the disclosure text, the model version and the timestamp alongside the call record, subject to the client's retention instructions. If the flow changes, the old artifact still points to the old text.
We also keep the suppression record separate from the marketing system. If a caller opts out, the next campaign must see the opt-out before the dialler runs, not after.
Not a one-time gate
Consent is not a gate at the start of a funnel. A caller can withdraw it in the middle of a call, and the agent must honour that in the next turn. A new purpose needs a new disclosure. A new voice clone needs new consent.
The interface is where this becomes real. If the only way to opt out is a hidden menu, the policy is fiction. If the agent can hear "stop" and actually stop, the policy is an interface.
This note describes our operating policy. It is not legal advice, and deployment rules vary by jurisdiction. Read the responsible voice commitments for the current position.