Voice agents
Voice agents for calls, explained properly
A voice agent answers the phone, listens to a whole sentence, works out what the person wants, does something about it, and says the answer out loud. This page is what actually happens in between.
The short version
An old phone menu can only understand the options it just read out. A voice agent takes the sentence the caller would have said to a person, and answers it.
That difference is the whole product. Nobody has to learn your menu, remember which number was for billing, or repeat an order number to three different people. They say what they want, once.
Four things happen between their sentence and yours
Hears
Speech is transcribed as the caller talks, not after they stop. If they interrupt, the agent stops speaking and listens.
Understands
A language model reads the running transcript with your instructions and your knowledge sources, and decides what this turn needs.
Acts
If the answer lives in another system, the agent calls a tool for it — an order, a free slot, a record — while the line is still open.
Speaks
The reply is spoken back in the voice you chose. Because the four stages overlap, the pause between turns sounds like a conversation.
They overlap rather than queue. The agent is already thinking about the first half of a sentence while the caller finishes the second half, which is why the gap before a reply sounds ordinary instead of awkward.
One call, beat by beat
Pick a direction and walk through it. The boxed cards are the moments the agent reaches a connected system without putting the caller on hold — that is the difference between a voice agent and a recording.
Which way does the call go?
Simulated example
A customer rings the published number to ask where an order is.
Caller
Hi — where is my order? It is 48219.
Agent
Let me check 48219 for you — one moment.
Tool callorders.lookup
Reads order 48219 from the connected store, while the caller is still on the line.
Left the warehouse yesterday · out for delivery today before 7 pm
Agent
It is out for delivery today, before 7 pm. Shall I send the tracking link to your WhatsApp?
Caller
Yes, please.
Tool callmessages.send
Drops the tracking link into the WhatsApp chat on the same number.
Delivered to the chat
Agent
Sent. Anything else I can help with?
Caller
No, that was it. Thanks.
Left behind when the call ends
- Full transcript and recording on the contact record
- Summary written: order status shared, tracking link sent
- Disposition filed: resolved, no follow-up needed
A call has two directions, and they are not the same job
Someone calls you
- Answer every call, including the ones that arrive at 2am or while the line is already busy.
- Let the caller say what they want in a sentence instead of choosing from a menu they have to memorise.
- Identify who is calling and pull up their record before the first question.
- Hand over to a person when your rules say so, with the transcript attached.
You call someone
- Work through a list: confirmations, reminders, renewals, follow-ups on a form someone filled in.
- Detect when a machine answered instead of a person, and act differently.
- Ask the qualifying questions first, so the human call starts further along.
- Write what was said back onto the contact record, and trigger the next step.
An agent that can only talk is a recording
The useful part is what it can reach while the line is open. Connect a store and it can say where the order is. Connect a calendar and it can offer a slot that is genuinely free, then take it.
Connections are per workspace and you choose which ones an agent may use. If the system you need has no connector, the agent can call your own authenticated endpoint instead.
What is left when the caller hangs up
A transcript of every call
Both sides, turn by turn, searchable. Nobody reconstructs a conversation from memory or from a two-line note.
Notes and a summary, written for you
A short account of what happened and what was agreed, ready to read without playing the recording back.
The record updated
The contact, the deal and the timeline reflect the call, because the agent wrote to them while it was happening.
A signed webhook, if you want one
Call started, call completed, call failed, analysis finished — posted to an endpoint you own so your own systems can react.
Calls placed to Indian customers have their own considerations — which languages the agent speaks, and what it costs in rupees. Voice bots for business in India
The questions people ask first
What is a voice agent for calls?
Software that holds a phone conversation. Speech recognition turns the caller’s voice into text as they speak, a language model decides what to say and which tools to use, and speech synthesis says the reply out loud. The loop runs continuously, so the caller talks normally instead of pressing keys.
How is this different from an IVR?
An IVR is a decision tree: it can only understand the options it read out. A voice agent takes an open sentence, works out the intent, asks a clarifying question if it needs one, and can change course halfway through when the caller does.
Can the caller interrupt it?
Yes, and they should be able to. The agent stops talking the moment someone starts, listens to the correction, and carries on from there — the same way a person would.
Can it do things, or only talk?
It can act. Connected tools let it look up an order, read real calendar availability, make a booking, or write to your CRM in the middle of the call. What it can reach is whatever you connect on the integrations page.
What happens when it cannot help?
You write the escalation rules. When one fires, the call moves to a person and the conversation so far goes with it, so the customer is not asked to start again.
Does the caller know it is an AI?
That is your decision to make, and you set it in the agent’s instructions. Most deployments say so in the opening line; some are required to.
How do I try it?
Open an account — it comes with a one-time grant of API credits — build an agent, and call it. Nothing needs to be connected to a phone line before you can hear how it sounds.
Build one, then ring it
The fastest way to judge a voice agent is to talk to one. Open an account, write the instructions, and have a conversation with it before you decide anything else.