What AI calling is, and how a voice agent handles a real phone call

A real-estate sales office in Pune gets property enquiries from portals and a Meta lead form all evening. The telecallers are on other calls, or have gone home, and the first call back comes the next afternoon. Someone suggests an "AI that makes and answers calls", and the owner has no way to judge whether that means a clever menu, a chatbot with a speaker, or something that can really hold a conversation.

This post explains the moving parts of a voice agent on a call, in plain terms, so you can ask vendors sharper questions and decide where it fits next to the people you already have. A voice agent works alongside your telecallers; it does not stand in for them.

What a voice agent is

An AI voice agent is software that takes part in a phone call as one of the two speakers. It listens to the person, works out what to say, and says it aloud, in a loop, until the call ends. It is not a recording and it is not a keypad menu. It can answer questions you did not script, as long as the answer is in what you gave it.

What you give it is a prompt (its role, tone and rules), some business information (project names, locations, configurations, visit timings), and a small set of actions it may take, such as booking a site visit. That is the whole product from the business side. Everything else is plumbing.

The line: how phone audio reaches the AI

A phone call is audio travelling over a telephone network. For an AI to join, that audio has to be handed to software in real time, and the software's voice has to be sent back the same way. There are three common ways to do this.

• SIP trunk. A carrier delivers calls to your software over the internet. Many Indian and international providers offer this.
• WhatsApp voice calls. The agent is attached to a WhatsApp account and calls or answers there.
• A website call button. The caller talks through the browser; no phone number is involved.

Which one you pick decides what callers see (a normal number, a WhatsApp contact) and how much each minute costs. We compare them in choosing lines for AI calls.

Listening: speech recognition

The first stage turns the caller's speech into text. This is speech recognition, sometimes called speech-to-text. On a phone it is harder than it sounds: narrow-band audio, traffic noise, a speakerphone, a caller who switches between Hindi and English in one sentence.

Recognition also has to decide when the caller has stopped talking, which is its own problem. We cover it in barge-in and turn-taking. For now, remember that a mistake here is invisible later: if "Thursday" is heard as "Tuesday", the rest of the agent works perfectly on the wrong word.

Thinking: the language model

The text goes to a large language model, the same kind of technology behind chat assistants. It reads your prompt, the conversation so far and what the caller just said, then writes the next reply. It can also decide to use a tool instead of replying, which comes up below.

This stage is why the agent can handle "is the 2-BHK in Tower B ready for possession, and can I see it on a Sunday?" without a script for that exact question. It is also where wrong answers come from. A model that is not told the possession date or the price may guess. Good agents are instructed to say they will check, or to hand over to a person, rather than invent.

Speaking: the voice

The reply text is turned into audio by a text-to-speech voice. Quality here is what callers notice first: pace, pronunciation of names and Indian place names, how natural Hindi or Tamil sounds. Voices differ a lot by language, so always listen in the language your callers will use.

Chain or one realtime model

There are two common designs.

In a chain, three separate services do the three jobs above, one after another. You can pick the best recognition for Hindi, a language model you trust and a voice you like, and swap any one later. It is flexible and usually easy to inspect, because you can read the text at every step.

In a realtime model, a single model takes audio in and produces audio out. There are fewer hand-offs, which can make conversation feel more fluid, and you choose less. Cora Call supports both: a chain where you choose each stage's vendor, or a realtime model such as OpenAI Realtime, Gemini Live or ElevenLabs Agents. Neither is better in every case. We look at the timing difference in why the pause matters.

Tools: doing something during the call

Talking is not the point; outcomes are. During the call the agent can call a tool: check a calendar and book a site visit on Google Calendar, Cal.com or Calendly, update a CRM record, send details to a webhook, or transfer the call to a telecaller or sales executive on your team. The agent says something short while the tool runs, then reports the result.

What the business gets after the call

A phone call normally disappears. An AI call leaves a record: a transcript, an optional recording (a workspace setting you control), a short summary, extracted fields such as name, budget, preferred area and visit date, and the cost of that call. The telecaller who picks the lead up reads the summary instead of starting from zero. A manager can read ten transcripts in minutes and see what buyers really ask, where the agent stumbled and which questions the business has no answer for.

Who it suits, and where it does badly

It suits calls that are frequent, similar and low in emotional stakes: the first call to a new enquiry, qualifying questions (area, configuration, budget, timeline), site-visit booking and reminders, follow-ups after a visit, calling back old leads, and after-hours cover. It suits teams that lose enquiries because every telecaller is busy or the office is closed.

It does badly with:

• Distressed or angry callers who need to feel heard by a person.
• Price negotiation, booking-amount discussions, legal or loan advice, and anything where a wrong answer is costly.
• Very noisy lines, or callers with strong accents in a language the setup does not cover well. Test with real recordings.
• Businesses with no written answers. An agent cannot know what nobody has written down.

The sensible design is not "AI instead of people" but AI beside your telecallers: it takes the routine calls, and a clean handoff puts every hot lead in front of a person with a brief. We set out how to divide the calls in AI and your telecallers, and how the property version works from enquiry to site visit in AI calling for real estate.

Where to start

1. List your ten most common call reasons from the last month, inbound and outbound.
2. Pick the one with a clear right answer and a clear end, such as the first call to a new enquiry or booking a site visit.
3. Write down the facts the agent will need and the cases it must hand to a person.
4. Try the live demo, then run a small trial on real calls and read every transcript for a week.
5. Check the cost per call on the pricing page against what a missed call costs you.