Skip to navigation

AI voicemail

When a rep reaches a voicemail, one click hands the call to the AI. It waits for the beep, leaves a message written for this prospect, and hangs up.

Leaving voicemails is where reps lose the most time: waiting through the greeting, the beep, the same message for the fortieth time. With AI voicemail the rep presses one button and moves straight to the next call. The AI stays on the line, waits for the beep, says the message in a natural voice, and hangs up.

AI voicemail is off until a workspace admin turns it on, in Sales calls → Settings → Let the AI leave voicemails (from a signed-in session, PATCH /v1/workspace/policies { "aiVoicemail": true }; API keys can’t change it). While it’s off, leaveVoicemail answers 403 with error: "ai_voicemail_off". AI voices are “artificial voices” under the TCPA (FCC declaratory ruling, Feb 2024). Get the person’s prior express written consent before an AI voice speaks to them, and pass consent so the call’s record shows where it came from (see Record the consent). Some states add their own rules. Check with your own counsel before you use it. This is not legal advice.

Leave one

From your server, with calls:write, while the call is live:

// A message written by the AI from the call's context:
await rydhm.repCalls.leaveVoicemail(callId);
// Or your own words, in a chosen voice, keeping the rep on the call to listen:
await rydhm.repCalls.leaveVoicemail(callId, {
message: "Hi Dana, it's Maya from Rydhm. I'll send the pricing we talked about. Call me back on this number.",
voice: "aura-2-thalia-en",
releaseRep: false,
});
FieldDefaultWhat it does
messageWritten by the AIWhat to say, up to 500 characters: your app’s own script for this prospect. Without it, the AI writes a short message from the call’s context: the prospect’s name, the goal of the call, your notes and playbook.
voiceThe rep’s voice, else the workspace’s, else BlakeAn id from GET /v1/voices, for this voicemail only. Any Cartesia voice id or Deepgram Aura or Flux voice (aura-2-thalia-en) works too; anything else answers 400 invalid_voice.
releaseReptrueTake the rep off the call at once, so they can dial the next one. false keeps them on the call to listen while the AI speaks: their microphone is taken off the call, so nothing they say lands in the prospect’s voicemail.
consentNone{ source, obtainedAt? }: where the prospect’s consent to an AI voice came from. Kept on the call and in the audit log. See Record the consent.

The call’s status becomes leaving_voicemail straight away. It answers 409 not_answered while the call still rings (ask once a voicemail picks up), 409 call_over when the call has already ended, 409 not_outbound on a callback (someone calling in), and 409 ai_not_in_call in the first moments of a call, before Rydhm’s listener has joined: try again a second later. Pair it with detection: when call.answered says by: "machine", offer the rep the button, or leave the voicemail automatically:

for await (const e of rydhm.repCalls.events(call.id)) {
if (e.type === "call.answered" && e.data.by === "machine") await rydhm.repCalls.leaveVoicemail(call.id);
}

Choose the voice

Your reps talk in their own voices. Rydhm’s voice speaks only its own lines: the AI voicemail, and the recording notice where everyone on a call must be told. A call uses one voice, chosen in this order:

  1. The voice you send with the voicemail (that voicemail only).
  2. The rep’s own voice: voicemailVoice on PUT /v1/reps/{externalId}.
  3. The workspace’s voice: an admin sets it in Sales calls → Settings → Voicemail voice.
  4. Rydhm’s default, Blake.

Leave a rep’s voicemailVoice as null and they follow the workspace’s voice. In the Rydhm app that reads “Same as the team”.

To let people pick a voice in your app, list them with GET /v1/voices (calls:read). Each voice has a name, a short description and sampleUrl: an MP3 of it reading a voicemail. It’s a static file, so playing it costs nothing and needs no key. default is the voice used when nothing else names one: the workspace’s (source: "workspace") or Rydhm’s (source: "rydhm").

const voices = await rydhm.voices.list();
// Show voices.items (name, description, ▶ sampleUrl), then save the one picked:
await rydhm.reps.upsert("usr_4821", { name: "Maya Patel", voicemailVoice: voices.items[3].id });
// null follows the workspace's voice again:
await rydhm.reps.upsert("usr_4821", { name: "Maya Patel", voicemailVoice: null });
{
"default": { "id": "a167e0f3-df7e-4d52-a9c3-f949145efdab", "name": "Blake", "source": "rydhm" },
"items": [
{
"id": "a167e0f3-df7e-4d52-a9c3-f949145efdab",
"name": "Blake",
"description": "Energetic · Male · US",
"language": "en",
"sampleUrl": "https://app.rydhm.ai/voicemail-voices/a167e0f3-df7e-4d52-a9c3-f949145efdab.mp3"
}
]
}

A voice that isn’t on the list (any Cartesia or Deepgram voice Rydhm can speak) still works; it has no sample, and default.name reads “Custom voice”.

AI voices are “artificial voices” under the TCPA (FCC declaratory ruling, Feb 2024). Get the person’s prior express written consent before an AI voice speaks to them, and pass consent so the call’s record shows where it came from.

FieldRequiredWhat it is
consent.sourceYesWhere the consent came from, in your words: a form, a signed agreement, a CRM record. 1 to 255 characters.
consent.obtainedAtNoWhen the person gave it (ISO 8601). It can’t be in the future.
await rydhm.repCalls.leaveVoicemail(callId, {
consent: { source: "Demo request form, acme.com/demo", obtainedAt: "2026-11-10T09:30:00Z" },
});
  • It’s kept on the call, with the time Rydhm recorded it: GET /v1/rep-calls/{id} returns it as voicemail.consent ({ source, obtainedAt, recordedAt }), next to voicemail.requestedAt, the time the AI was asked to leave the message.
  • It’s in the workspace’s audit log, with the voicemail request.
  • A consent without a source, with a source over 255 characters, or with an obtainedAt that isn’t a date or is in the future answers 400 invalid_consent.
  • consent is optional. Rydhm doesn’t check it: it keeps what your app sends. Getting the consent is up to you.

Who the message says is calling

The message names the rep (their name in PUT /v1/reps/{externalId}) and the company they call for: the call’s context.seller.company, or your Rydhm workspace’s name when the context has none. With The AI says it’s an AI on, a message the AI writes opens with “Hi Dana, this is an automated message on behalf of Maya at Acme”, and a message of yours gets that line in front unless it already says it’s automated. The call-back number is the number the call went out from.

What happens on the call

  1. The AI takes over the line, and the rep is released (unless releaseRep: false).
  2. It waits for the voicemail greeting to end and the beep (call.beep).
  3. It speaks the message, then hangs up.
  4. You get call.voicemail_left with the exact words and the voice, then call.ended with outcome: "voicemail_left". The words are also kept on the call as voicemail.text (with voicemail.by: "ai"; voicemailText and voicemailBy still work), and the call.voicemail_left webhook goes out. Webhooks carry the voice and the length, never the words.

When it doesn’t happen. If the AI can’t leave the message (the call wasn’t answered yet, it’s a callback, the line was gone, or the message couldn’t be spoken), you get call.voicemail_failed with a reason (not_answered, not_outbound, line_gone, no_audio), and the call goes on or ends as it would have. Show it to the rep: they can leave their own.

The rep’s own message. When a machine answered and the rep speaks their own message after the beep (eight words or more), that is a voicemail left too: call.voicemail_left with by: "rep" and the rep’s words as transcribed, voicemailBy: "rep" on the call, and the call ends as voicemail_left. It isn’t billed as an AI voicemail.

On a callback that nobody answered, the last step of the pool’s routing can be voicemail: then the caller leaves a message. It arrives the same way, with by: "caller" and their number in from on the event, and voicemailBy: "caller" on the call.

In the transcript, the AI’s lines have speaker: "ai".

Disclosure and recording

  • Transcribing a call records it in the legal sense, and some states require everyone on the call to be told; on a live call, your rep says it. The workspace’s disclosure setting (Sales calls → Settings → Recording, and telling people → The AI says it’s an AI, on by default; repCallDisclosure in the workspace policies) makes an AI voicemail say it’s an automated message on the rep’s behalf, and tells a caller leaving a message that it’s recorded and transcribed.
  • When the call is recorded, the AI voicemail is in the recording, on the rep’s side (the left channel).
  • Calling hours, your Do not call list and the US-only dialling rules apply to rep calls exactly as to AI calls. See US-only dialling.

What it costs

Provisional while rep calls are in preview: $0.02 for each AI voicemail, on top of the call’s minutes. Final prices will be on the Billing page before launch.