Bland AI

Because Bland's BYO-SIP outbound path is unreliable, CenterLeap dials the customer from Telnyx (showing our number) and, on answer, bridges a SIP leg into Bland as an inbound call. Bland answers that inbound leg with the configured agent. So the agent config lives on the inbound number, not on a per-call outbound request.

Telnyx dials customer (caller ID = our number)
   └─ on answer → SIP bridge into Bland inbound number
        └─ Bland answers as the configured agent (voice + prompt)

From the Bland dashboard → API Keys. Send it as the authorization header on all requests (no Bearer prefix).

Set the agent on the Bland inbound number that the SIP bridge targets. The key fields: prompt (persona), voice, first_sentence (leave empty to have the agent listen first), max_duration, and record.

curl -s -X POST "https://api.bland.ai/v1/inbound/%2B1XXXXXXXXXX" \
  -H "authorization: $BLAND_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "You are Sara, a recruiting assistant ...",
    "first_sentence": "",
    "voice": "YOUR_VOICE_ID",
    "max_duration": 4,
    "record": true
  }'

The prompt is the agent's entire behavior. A production recruiting prompt should include: a listen-first rule (say nothing until a human speaks), a “process what they actually said” rule so the agent exits politely on disqualifiers instead of plowing through a script, the ordered questions, and an explicit voicemail rule.

This is Bland's weak spot when bridged inbound: it cannot reliably tell that it reached a voicemail. CenterLeap compensates at the carrier layer — Telnyx answering-machine detection fires a beep event, the agent leg is torn down, and a pre-recorded message is played into the mailbox. The Bland prompt is also told to say nothing to a recorded greeting and never claim it left a message.

If voicemail accuracy is the priority, see the Vapi integration, which provides native voicemail detection.

Read the current inbound-number config back to confirm it saved:

curl -s "https://api.bland.ai/v1/inbound/%2B1XXXXXXXXXX" \
  -H "authorization: $BLAND_API_KEY"

Then place a real bridged call through the Telnyx path and listen end to end — audio bugs (talking over a greeting, wrong voice) only surface on a live call.