What is a voice AI agent?
A voice AI agent is software that listens to spoken requests, works out what the person wants and responds or takes action, usually by speaking back in real time. It combines speech recognition, a language model that decides what to do, and speech synthesis for the reply.
How does a voice AI agent work?
Most voice AI agents run a loop with three stages:
- Speech to text. A speech recognition model turns the audio into text as the person talks, and detects when they've finished a turn.
- Reasoning and tools. A language model reads the text with the conversation so far, decides what the person wants and, if it's an agent rather than an assistant that only answers, calls tools: look up an order, book a slot, update a record.
- Text to speech. A voice model speaks the reply.
Newer speech-to-speech models collapse these stages into one model that hears and speaks directly, which cuts delay and keeps more of the tone of voice.
What makes voice hard isn't any single stage. It's the conversation mechanics around them:
- Latency. On a call, people expect a reply within a beat, so every stage has to stream.
- Turn-taking. The agent has to tell a thinking pause from a finished sentence.
- Interruptions. People talk over the agent to correct it. It has to stop, listen and adjust.
- Confirmation by ear. There's no screen to review, so the agent reads back the important details before acting.
Where are voice AI agents used?
Three settings account for most of them:
- Phone lines. Contact centers and front desks use voice agents to answer calls, book appointments, check order status and route callers to a person.
- Devices. Cars, smart speakers and wearables, where a screen is small, absent or unsafe to look at.
- Inside software products. People speak requests to the product they're already using, alongside typing and clicking.
The third is the newest, and it changes the design. On a phone line, voice is the whole interface. Inside a product, voice is one input among several, and the screen is still there to show results.
What changes when a voice agent works inside a product?
Inside a product, the user speaks the request and the screen shows the result. A field technician with gloves on says "log two hours on the Henderson job and mark the pump as replaced." The agent fills the timesheet, updates the job and shows a card with both changes. The technician glances, says "confirm," and moves on.
That combination fixes voice's biggest weakness. Long lists, tables and anything that needs checking appear on screen instead of being read aloud. Short requests and hands-busy moments go by voice.
It also raises the stakes on permissions, since a spoken request should be able to do only what that signed-in user could do by hand.
In Foldspace, a team can turn on dictation (Speech to Text) and two-way Voice Chat for the Product Agent, so users can type or speak, and a spoken request runs through the product's own APIs under the same signed-in user's permissions as a typed one.