A natural voice at every turn

GPT-Live changes the unit of a voice conversation. Giga agents can now process what the customer says while speaking, decide many times a second whether to talk, listen, or hand off a task, and carry on the conversation while that task completes in the background. Interruptions, revisions, and language switches become part of the flow.

Bloom, a flower shop that takes orders through Giga’s agent, is a small demonstration of the difference between earlier LLM-powered voice models and the new GPT-Live API.

Gio wants to send birthday flowers to her mom, who is staying at a hotel in Madrid. She never fills out a form or repeats herself. The conversation is faster, more natural, and absorbs mistakes and changes of mind as they happen.

The shape of the interaction

The agent is built around OpenAI’s new GPT-Live voice API. It keeps listening while it speaks, and it can hand deeper work to a background process without dropping the conversation.

The task is ordinary. The interaction is not.

Most digital orders divide a simple request into a sequence of screens: choose a product, enter a recipient, find an address, select a delivery window, write a card, review the details, and submit.

That structure works when the user already knows exactly what to enter. This experience doesn’t translate well to a conversational model. Gio starts with an intention — “It’s my mom’s birthday tomorrow” — and fills in the details as she thinks of them. A good assistant needs to hold the thread, not force her to turn the thread into a form.

The agent is designed around that premise: the conversation is the interface, while the order takes shape in the background.

One conversation, six pieces of work

Gio tells the agent her budget, the flowers her mom likes, her mom’s name, and the hotel she is staying at. As she speaks, those details appear automatically, in the order they surface.

The agent also delegates work that shouldn’t interrupt the conversation. It resolves finding Hotel Oliva in Madrid, checks tulip availability and pricing, and locks in a desired delivery window. Gio can keep talking while those tool calls happen.

• Voice-only completion: no typing, form filling, or manual recipient entry.
• Background delegation: hotel, inventory, pricing, and delivery checks continue while the conversation stays alive.
• Progressive order building: each confirmed detail lands in the order as it is spoken.
• Full-duplex interaction: Gio can interrupt, and the agent yields immediately instead of finishing its sentence.
• Multilingual expression: Gio moves from English to Spanish naturally when she writes the card.
• Explicit approval: The agent asks before placing the order and shows confirmation only after submission succeeds.

Listening is an active capability

When Gio says, “Actually, she won’t get there until lunchtime. Can you do after two?”, the agent stops speaking. She doesn’t wait for the morning delivery sentence to finish. It does not ask her to restate the hotel, the recipient, or the bouquet.

That moment is small, but it is a major difference between a voice interface that is sequential and inflexible and one that can participate in a natural conversation. The assistant must recognize a change in intent, preserve context, and make modifications from new constraints discovered on the fly.

Gio then adds two more instructions: leave the flowers at reception, and do not call her mom because it is a surprise. The agent acknowledges both and carries them into the final order.

A bilingual card, without a language switch

Gio composes the message in Spanish:

“Mamá… Un abrazo desde lejos. Te quiero, tu hija favorita.”“Mom… A hug from far away. I love you, your favorite daughter.”

She pauses. The agent gives her room. Then Gio returns to English with a joke: “I’m her only daughter.” The agent catches it — “Well, good odds.” — and keeps the task moving.

The exchange is not a language demo separated from the work. It is one interaction. The language changes because Gio changes it, and the order remains coherent throughout.

The surface should show the work

Bloom’s visual layer is intentionally quiet. No dashboard, no stepper, and no collection of controls competes with the conversation. The page makes only the useful state visible:

• the recipient and hotel as they are understood,
• the card message as it is spoken,
• the bouquet and delivery total when the lookup returns,
• the updated afternoon window and reception instructions, and
• the final confirmation after the order is placed.

The UI is not a transcript pasted beside a mock order. It is the order being composed by the conversation. The page gives Gio confidence that the agent heard the important parts without asking her to manage the system herself.

From mere comprehension to agentic action

While in this case, Bloom is a demonstration, the design question is real. A voice agent can move from understanding a request to carrying it through across the systems that power a real business — without making Gio wait.

That is where voice agents become more than a friendlier front door. They need to listen while they work, delegate without losing context, turn unstructured speech into a structured task, and ask for approval only when it matters.

For Gio, the space between idea and completion is incredibly short. Tulips for María González, delivered to Hotel Oliva in Madrid tomorrow between two and five p.m., left at reception, with no call to the recipient. All in seconds.

See how Giga and OpenAI’s GPT-Live can transform your customer experiences. Request a demo today.