One incoming call, start to finish
An inbound voice AI agent, from ring to written brief
No menu, no hold music, no operator to become free. Below is everything that happens while the caller is still talking — the ring, the first sentence, the lookup, the booking, and the note left behind for whoever needs to ring back.
The board used to need someone standing in front of it.
The sequence, in order
Every stage below is a real part of the pipeline, not a diagram of one. The same pipeline runs a browser voice session, so an agent you tuned in the dashboard behaves identically on a phone call.
Watch one call all the way through
Simulated example
One ring — the call is already with the agent.
Caller: Hi — my order was supposed to arrive on Tuesday. Where has it got to?
Agent: Let me check that for you. What is the order number, or the phone number on the account?
Caller: It is 40817.
Your order endpoint → order 40817 — out for delivery, due before 6 pm today
Agent: It left the warehouse this morning and is out for delivery — due before six this evening.
Caller: Good. Can you send me the tracking link?
WhatsApp message → tracking link sent to this number
Agent: Done — the tracking link is on its way to this number on WhatsApp.
Caller: And if it still does not come, I want to talk to a person.
Agent: Understood — I have noted that, and a colleague will ring you back if it has not arrived.
Handoff filed
- Reason: caller wants a person if delivery fails
- Covered: order status checked, tracking link sent
- Key facts: order 40817, identity confirmed by order number
Agent: Anything else I can help with? … Thank you for calling.
Call ended — recording, transcript, summary and disposition saved.
- Ring
- First words
- Reply
- Lookup
- Action
- Handover
- Hang up
Ring
The call reaches the agent, not a queue
Your carrier delivers the call down a SIP trunk to a per-tenant inbound endpoint. A room is created and a voice agent is dispatched into it. Nobody is waiting for a free operator, and several callers can arrive at the same moment.
Keep your own carrier or rent a number from the console.
First words
The caller says what they want, in their own words
There is no menu to sit through. Speech recognition streams partial transcripts while they are still speaking and finalises on end of speech, with server-side voice activity detection closing the turn at roughly fifty-millisecond endpointing.
23 languages recognised, detected automatically, including a sentence that switches between Hindi and English halfway through.
Reply
The answer starts before the sentence is finished
The language model begins generating before the transcript is fully confirmed, pushes each finished sentence downstream, and the voice is synthesised chunk by chunk — so playback starts before the whole reply exists. That is why it sounds like a conversation rather than a recording.
Interrupt it mid-sentence and it stops talking and listens.
Lookup
It goes and checks, while they wait on the line
Mid-conversation the agent can call your own REST endpoints, an MCP server or a built-in tool: an order status, a free calendar slot, a contact record, a saved policy passage. If an integration is not connected the tool returns an error the agent speaks its way out of, and the call carries on.
An unknown tool name is skipped and logged, never taken as an outage.
Action
It finishes the job rather than promising one
A slot is booked into Cal.com or Google Calendar. A contact is updated. A tracking link, address or price goes out on WhatsApp while the caller is still on the line — on phone and WhatsApp calls those messaging tools are on by default.
It can send WhatsApp and email. It does not send SMS.
Handover
When it should not be the one answering
If the caller asks for a person, is upset, or wants something genuinely outside the agent’s scope, it says a colleague will ring back and files a handoff carrying the reason and a short brief: what they want, what has been covered, the order number, whether identity was checked.
It does not bridge a live call to your phone. Escalation is a callback, so write your prompt that way.
Hang up
The call leaves a record, not a memory
Recording, full transcript, summary, sentiment and disposition — on every call, not a sample. Handoffs queue in an inbox your team calls back from in one click, and signed webhooks push the same events into your own systems.
Stored on Microsoft Azure in India (Central India), with configurable retention.
Simulated example
The old way of getting the same answer
A keypad menu
0 key presses
Walk it. The errand is moving an existing appointment.
It says: For opening hours, press 1. For appointments, press 2. For anything else, press 3.
Or one spoken sentence
No levels. The caller just says it.
“I need to move my Thursday appointment to next week”
What it can reach for mid-sentence
A tool is something the agent may use during the call, chosen by name when you set the agent up. Web search is always on. Anything it cannot reach returns an error the agent explains out loud rather than a dropped call.
Look something up
- Your own REST endpoints
- An MCP server you run
- Saved knowledge and documents
- A web search
Book something
- Free slots in Cal.com
- Free slots in Google Calendar
- Create the event
Write something down
- Update a contact
- Record an opt-in
- Remember a fact for next time
Send something
- A WhatsApp message or template
- Quick-reply buttons or a list menu
- A location or a link button
- An email
Check an order
- Shopify order status
- Shopify product lookup
- WooCommerce order status
End it properly
- Request a call back
- File a handoff for a colleague
- Close the call
Switch a few on, read the agent back
Tick what this agent may use during a call and the sentence underneath is the truth it can act on.
On a call this one can check your own endpoints mid-call, find a free slot and book it and send a link or address on WhatsApp.
Simulated example — anything it cannot reach returns an error it says out loud, not a dropped call.
Connecting the systems behind those tools is a settings screen, not a project — see what connects out of the box, or build the agent itself without writing any code on the no-code builder.
Three things it will not do
- It cannot connect a live caller to a person — escalation is a written handoff and a callback.
- Inbox actions — notes, tags, ticket status — belong to the chat channels, not to a call.
- It does not send SMS. Follow-ups go out on WhatsApp or email.
Asked by people about to forward a line
- What is an inbound voice AI agent?
- It is software that answers your incoming phone calls in place of a menu or a queue. The caller speaks normally, the agent works out what they need, looks it up through your own systems while they are on the line, and either finishes the job or files a written handoff for a colleague to ring back. Every call leaves a recording, a transcript and a summary.
- How quickly does it answer and reply?
- It answers on the first ring because there is no queue to join. Within the call, speech recognition, the language model and the synthesised voice all stream at once, and the model starts composing before the transcript is fully confirmed, so the caller hears the beginning of an answer while the rest is still being generated.
- Can a caller interrupt it?
- Yes. Speak while the agent is talking and it stops immediately and listens. Turn-taking is handled by server-side voice activity detection with low-latency endpointing, which is what stops the agent talking over someone who has changed their mind mid-sentence.
- What happens when the agent cannot help?
- It tells the caller a colleague will ring them back, then records a handoff with a reason and a one or two sentence brief written for whoever picks it up — what the caller wants, what has already been covered, key facts such as an order number, and whether identity was checked. Your team returns the call from the handoff inbox. The agent does not connect a live caller to a person.
- Does it understand callers who switch language mid-sentence?
- Yes. Speech recognition covers 23 languages with automatic detection and handles code-mixed speech such as Hindi with English. The voice it replies in is chosen per agent, with Indian-language voices available in 11 languages.
- Do I have to move my phone number to use it?
- No. Connect a Twilio or Plivo account, or any SIP trunk, and import the numbers you already publish. Exotel and Tata Smartflo connections are set up with our team. If you are starting from nothing you can complete KYC and rent Indian numbers inside the console instead.
- Can I run it only when my team is busy?
- Yes. Forward your line on busy or no-answer so the agent only sees calls that would have rung out, or put a business-hours check in the call flow so it covers evenings and weekends only. The rate is the same either way because you pay for the minutes it talks.
Hear the sequence before a customer does
Call the agent from your browser, then read back your own transcript. Nothing is forwarded until you say so.
Next to this page: what a month of those calls costs across a whole queue, the voice and IVR product itself and the same agent dialling out.