Voice agent (AI)

An AI voice agent is software that holds a spoken conversation, usually over a phone call, to understand what the other party wants and complete a task, either answering calls for a business or placing them for a person.

Two architectures. OpenAI’s voice agent guide describes the two common designs:

• Chained. Speech-to-text transcribes the caller, a text model decides what to do and calls tools, and text-to-speech speaks the reply. Each stage can be inspected, logged or replaced on its own, and the text between stages leaves a transcript, but every hand-off adds delay.
• Speech-to-speech. One model takes audio in, reasons, uses tools and produces audio out in a single session, which keeps latency low and lets the caller interrupt mid-sentence.

Either way the agent sits on ordinary telephony. Calls arrive through the phone network or SIP, and platforms that separate speech processing from call control can use protocols such as MRCPv2 to drive recognizers and synthesizers.

Both directions. Businesses use voice agents to answer support lines in place of IVR menus. Assistants use them to phone businesses for their users: Google’s Business Profile help describes automated calls it places to book appointments, check restaurant wait times and confirm product prices and availability. When both ends are voice agents, two programs that work with structured data talk through synthesized speech and speech recognition, and every order number has to survive that round trip.

Rules for outbound calls. In February 2024 the US Federal Communications Commission ruled unanimously that calls made with AI-generated voices are “artificial” under the Telephone Consumer Protection Act, so the Act’s rules on artificial or prerecorded voice calls apply to them.

The structured alternative. If the business publishes an agent endpoint, the calling agent can send a structured A2A request instead of dialing. A2A discovery starts from a domain, while a person’s agent often starts from a phone number, so a lookup from number to endpoint is needed first. Emissar is building Resolve, a registry for that mapping that an agent queries before dialing. Status: In development. When two voice agents meet on a phone call walks through the full picture.

Neighbouring terms. Speech-to-text and text-to-speech are the two ends of the chained design. STIR/SHAKEN authenticates the calling number on the phone network.