What is an AI voice agent, and how does it work?

An AI voice agent is software that holds a real phone conversation. It answers or places a call on an ordinary phone number, listens to what the caller says in their own words, works out what they mean, replies in a natural voice, and acts on the conversation: it answers questions from your documents, asks its own, books a site visit and writes the whole call into your CRM. Unlike an IVR, there is no "press 1". Unlike a recorded robocall, it listens.

For an Indian business the hard parts are specific: Indian phone lines, callers who mix Hindi, Gujarati and English in one sentence, prices in lakh and crore, and rules on who may be called and when. This guide explains how an AI voice agent works end to end on an Indian call, what it does well, and where it still falls short.

What an AI voice agent is, and what it is not

Three different products get called voice AI.

• An IVR plays recorded menus. The caller presses a key or says a fixed word, and the system routes the call. It cannot handle a sentence it was not built for. We compare the two in AI voice agent vs IVR.
• A robocall plays a message and does not listen at all.
• An AI voice agent hears free speech, answers from what it knows, asks follow-up questions and decides what to do next.

A voice agent is closer to a trained telecaller than to either of the other two. It is still not a person, and most of what follows is about what that means in practice.

How a call works, end to end

Every call passes through six parts. Each can fail on its own, and a demo rarely exercises all of them.

1. The phone line

The agent needs a real number and a connection to the phone network. In India that usually means a DID number from an Indian carrier, delivered over SIP trunking to the server that runs the agent. Inbound calls ring that number. Outbound calls leave through the same trunk. Calls are usually recorded on the carrier's side, and those recordings have to be collected and filed against the right lead.

This is the least interesting part to watch and, in our experience, the part where failures are quietest. A line can report itself healthy while refusing every call. The only proof that a phone line works is a real call that connects and carries audio both ways.

2. Hearing the caller

Audio arrives as a continuous stream. The agent has to decide when the caller has finished speaking, turn what they said into something it can reason about, and tell an interruption from agreement.

That last part matters more in India than a foreign demo suggests. "Haan", "ji", "hmm" and "achha" (yes, yes, hmm, okay) said while the agent is talking are how people listen here. They are not interruptions. An agent that stops dead at every one is unusable. At the same time, a short answer such as "haan" (yes) or "2 BHK" said when the agent is quiet must still count as an answer. Mobile lines in traffic and speakerphones in busy offices make both harder.

3. Deciding what to say

The agent reads the conversation so far and its brief: who it is, which company it is calling for, what it needs to find out, and what it must never say. Then it chooses the next reply.

Conversation is the easy part. Keeping one rule on every one of a thousand calls is the hard part. The clearest lesson from the calls we have reviewed: anything that must happen every time belongs in fixed rules, not only in the agent's brief. Honouring an opt-out, never asking a question twice and replying in the caller's language are all examples. A brief is advice. A rule is a control.

4. Knowing the answers

Prices, possession dates, carpet areas, amenities and payment plans should come from your own documents, looked up at the moment the caller asks. When the answer is not in them, the agent should say so and offer a callback. A guessed price or a guessed RERA detail on a recorded line is a misrepresentation, and a buyer will remember it.

In our experience, this is the part buyers test least. Ask a question you know the brochure does not answer, and listen to what happens.

5. Speaking

The reply is turned into speech, either as a separate step or as part of the same one. The voice and accent should be fixed, so the agent sounds like the same person on every call. Local place names should be said the way locals say them. Grammar has to match the voice, too: Hindi verbs agree with the speaker's gender, so an agent with a woman's voice must say "main bata rahi hoon" (I am telling you, in the feminine form), not "main bata raha hoon" (the masculine form).

6. Writing it down

The value of a call is what lands in the CRM afterwards. In BlackWolf's case that means the caller's name, budget, configuration, locality, timeline, sentiment and a summary, with the full transcript and the recording, and the site visit if one was agreed. A voice agent that talks well and writes nothing down is a demo, not a sales tool.

Every one of these steps adds delay. The gap between the caller finishing and the agent starting is the sum of them, and that gap, known as voice latency, decides whether the agent sounds like a person. We explain where it comes from in why latency decides whether a voice agent sounds human.

One call, from lead to CRM

Here is a sample call, to show the parts working together.

A buyer fills in a Meta lead form for a 3 BHK at 10:40 pm. The calling window is closed, so the lead waits in the queue. When the window opens the next morning, the agent calls. Its opening line, written with the developer when the agent was set up, names the company it is calling for. The buyer answers in Hindi, with English words mixed in:

"Haan, 3 BHK dekh raha tha. Budget around 90 lakh hai, but possession kab tak milega?"

(Yes, I was looking at a 3 BHK. The budget is around 90 lakh, but when will I get possession?)

In one breath the buyer has answered two qualifying questions, budget and configuration, while asking one of their own. A good agent answers the possession question from the brochure, notes the budget and configuration without asking for them again, and moves on to what it still needs: locality, purpose and a day for a visit.

When the call ends, the lead in the dashboard shows ₹90 lakh, 3 BHK, the Saturday visit if one was agreed, a summary that includes the possession question and the transcript, with the recording once the carrier supplies it. If the buyer had instead said "mujhe dobara call mat karna" (do not call me again), the number would have gone on the suppression list the moment it was said.

What voice agents are good at

• Calling every lead, fast. Speed to lead is where people fall behind, especially at night and at weekends. An agent calls within minutes, during permitted hours, and answers inbound calls at 2 am.
• Consistency. The four hundredth call of the day gets the same questions, the same facts and the same care as the first.
• Remembering. A well-built agent keeps a running note of everything the caller has said, so a budget mentioned in passing is a budget answered. On a follow-up call it starts from what the CRM already knows. We explain how in how a voice agent remembers and follows a language switch.
• Language. BlackWolf's agents speak the caller's language, whichever it is, and switch when the caller does, even mid-sentence. What that takes is in what it takes for voice AI to sound local in Hindi and Gujarati.
• Writing everything down. Every call is on record, which makes lead qualification something you can audit instead of something you take on trust.
• Persistence without nagging. No answer? It tries again at a different time of day, and it stops the moment someone says they are not interested.

Where they fall short

• Closing. Negotiating a price, reassuring a family that is unsure, building trust over several visits: that is your team's work.
• Anything outside the documents. A good agent refuses rather than inventing. A vendor whose agent always has an answer should worry you.
• Difficult lines. Heavy background noise, a speakerphone in a crowded room or two people talking at once still cause misunderstandings.
• Callers nobody planned for. A child picks up the phone. A broker answers instead of a buyer. A wrong number. Each needs deliberate handling, and a new agent will meet some of them for the first time on a live call.
• Your words, in every language. Speaking a language fluently is not the same as getting your project names, localities and price terms right in it. That is why each business's agent is tried on the business's own questions, in its callers' languages, before a caller hears it.
• Being right every time. Any agent will occasionally say something wrong. Transcripts and recordings are how you catch it, which is why every call should be kept.

Compliance is shared

An AI voice agent is a way of placing commercial calls, and India's rules on commercial calls apply to it. They sit under TRAI's TCCCPR framework. As of July 2026, TRAI states (PDF), opens in a new tab that promotional calls by entities of any sector must come from the 140 number series, and that entities wanting one must register with a telecom operator under that framework. TRAI's September 2026 amendment (PDF), opens in a new tab goes further: it names calls placed by software without direct human dialling, including those using artificial voices, as application-to-person (A2P) calls, and requires the sender to declare them to its operator in advance, with the numbers they will come from. That requirement takes effect sixty days after the amendment's publication in the Official Gazette.

Most of those obligations belong to the business making the calls, not to the software: DLT registration as a principal entity, a registered telemarketer, the right number series, the A2P declaration and scrubbing lists against the national do-not-call register. What software can do is enforce its own share on every call. BlackWolf checks its own suppression list before every dial, dials only inside the calling window, writes an opt-out to that list the moment it is said, stops follow-up calls when a lead says no, and keeps every call on record. Naming your company at the start of the call is part of each agent's scripted opening, agreed with you; it is not something the platform enforces. The split is set out on our compliance page and in our guide to TRAI's rules for AI calling. This is a summary, not legal advice.

The rest of this series

• What does AI calling cost in India? How per-minute pricing works, why rounding matters, and a worked example.
• What it takes for voice AI to sound local in Hindi and Gujarati. Script, Hindi mixed with English, lakh and crore, names and pronunciation.
• How to evaluate an AI voice agent vendor. The questions to ask before you sign, and the test calls to insist on.
• Why latency decides whether a voice agent sounds human. Where the pause comes from, and what good looks like.
• How a voice agent remembers and follows a language switch. Never re-asking, and following a caller from Gujarati into Hindi.

What to do next

If you sell property, see how this works for real estate. Whoever you are considering, put their agent on a real phone call in your buyers' language before you trust a demo.