Realtime voiceTutorial7 min

Build a Realtime voice agent intake flow

How to design a low-latency voice workflow that captures the right facts and hands off cleanly.

Real example

Example: voice intake for a service appointment

A home-services company wants a voice agent to capture appointment requests after hours. The agent needs name, address, issue type, urgency, availability, and escalation triggers.

Start a Realtime session with a narrow call objective. During the call, collect required fields, confirm spelling for critical data, and summarize the appointment request for dispatch review.

The voice agent becomes an intake worker with a bounded job instead of an open-ended conversational demo.

Server-created Realtime session secret

export async function POST() {
  const session = await openai.realtime.sessions.create({
    model: "gpt-realtime",
    audio: { output: { voice: "alloy" } },
  });

  return Response.json({
    client_secret: session.client_secret.value,
    expires_at: session.client_secret.expires_at,
  });
}

Tutorial path

How to implement it

Step 01

Define the call goal, required fields, disallowed decisions, and escalation triggers.

Step 02

Create a browser or phone session with a server-issued short-lived credential.

Step 03

Stream conversation events to the UI and backend transcript store.

Step 04

Extract call outcomes into structured fields after the call or at controlled checkpoints.

Step 05

Show the operator a call summary, missing fields, and recommended next action.

Checklist

Ready when these are true

Call objective documented

Short-lived client credential

Transcript stored securely

Escalation triggers tested

Post-call summary reviewed

Field notes

What matters in practice

Voice agents need a clear call objective before they need a persona.

Interruption, uncertainty, and handoff are core product states.

The transcript should become structured intake data, not disappear after the call.

Avoid these mistakes

Common failure modes

Do not let the voice agent make commitments the business cannot keep.

Do not skip interruption and noisy-audio testing.

Do not treat the transcript as the final structured record.

Practical tip

Voice UX improves when the agent confirms only high-risk fields, not every detail.

Apply this to a build