Terra: a voice concierge that runs group travel, powered by Vocal Bridge

A voice-first concierge that takes spoken updates for a 39-person Disney reunion and executes them across flights, hotels, and cars, with every action logged for a human to check.

The problem: group travel changes cascade, and they arrive as phone calls

Organizing travel for 39 travelers leaving from eight cities is a coordination problem where every change ripples outward. One traveler stays an extra day, and now their flight, hotel night, and rental car all move together. A maybe becomes a yes and needs slotting into the full trip.

These updates do not arrive as a tidy form. They arrive as a call, several at once, half-specified, in whatever order the coordinator thought of them. The person fielding those calls is the bottleneck, and any assistant that helps has to be trusted not to quietly get it wrong, because a booking that silently fails does not surface until someone is standing at a gate.

What Terra built

A coordinator talks to Terra the way they would talk to a human travel agent, handing over a stack of changes in one breath. Terra parses them into discrete tasks, executes the bookable ones against Sabre, researches the ones not ready to book, and writes every step to a dashboard. A live feed records what was requested, looked up, and resolved, so a person can verify the agent rather than take it on faith.

How they use Vocal Bridge

Terra is built on Vocal Bridge's Voice for your agent surface, where voice is layered onto an autonomous workflow. The split is clean. Terra owns the agent logic, the Sabre integration for flights, hotels, cars, and payments, the traveler profiles, and the dashboard. Vocal Bridge owns the real-time voice channel: the outbound call, the voice-to-voice model, barge-in, transcript streaming, and the conversational fillers that keep the line warm while a task runs.

import { defineTool } from "@vocalbridgeai/agent";

// execute a spoken change and record it for human review
export const resolveRequest = defineTool({
  name: "resolve_request",
  description: "Execute a traveler change and log it to the live feed",
  async run({ traveler, change }, ctx) {
    await ctx.say("Give me just a moment while I check the flights.");
    const result = await sabre.repairItinerary(traveler, change);
    await feed.log({ action: "resolve_request", traveler, change, result });
    return result;
  },
});

Illustrative. The actual Terra implementation handles richer traveler and itinerary data.

What's hard about voice here

The first hard part is a single turn carrying four separate jobs: a rental to research, a two-person trip extension with flights, room, and car, a VIP pass to confirm, and a traveler to add, all without pausing. Transcript streaming and barge-in let Terra hold the whole batch and read it back accurately before acting.

The second is dead air. Repairing an itinerary means calling Sabre, and those calls take real time. Silence on a phone line reads as a dropped call. Real-time voice-to-voice responses and conversational fillers let Terra say "give me just a moment while I check the flights" and stay present.

The third, and the one the team treats as the whole point, is grounding. Terra ties every spoken commitment to a logged action in the live feed, giving the team the right foundation for production: make successes, retries, and partial completions equally visible so a coordinator always knows what actually happened.

The quote

"This is the most important part, because this is a live verification of everything that Terra is doing to make sure that there's no hallucinations and she's not just making things up and not just saying, okay, I'll do it on the phone.", Gina Wang, Terra (Dunder Mifflin)