Voice AI Use Cases, Pricing in India and the Architecture of an AI Calling Agent

This is Part 3 of a three-part Voice AI guide. Part 1 covered what Voice AI is and the market. Part 2 introduced our AI voice agent. This part covers use cases, pricing and the technical design.

Voice AI use cases

Voice AI for real estate lead qualification

This is where our platform started. Property portals and ad campaigns bring in thousands of leads, and most never pick up or are not serious. The AI agent calls each lead, talks about the project, quotes prices and locations from project facts, and asks if the budget and location work for them. Interested buyers are tagged "Lead | Interested" or "Wants to visit" and go to the sales team's CRM. The rest are filtered out.

Lead follow-up for any sales team

Education, insurance, loans, automobiles and B2B services all have the same problem: too many leads, too few callers. The AI agent calls every lead within minutes and hands over only qualified ones. Leads that ask for details on WhatsApp are tagged "Lead | Send WhatsApp".

Reminders and confirmations

Payment due dates, EMI reminders, appointment confirmations and delivery checks. These calls are short and repetitive, and they are a good fit for automation.

Surveys and feedback

Post-purchase feedback, NPS calls and field surveys. Because every answer is transcribed and summarised, results come back as data, not as a pile of notes.

Citizen and public outreach

Large-scale outreach in Hindi, such as scheme awareness or service information. Campaign analytics can be grouped by telecom circle, so you can see reach by region.

Inbound call answering

Give customers a number to call. The AI agent answers 24x7, handles common questions from your instructions and facts, and records the call for your team.

Voice AI pricing

Our Voice AI uses simple, prepaid, per-minute pricing in INR.

• Pay per minute of connected call time. The standard rate is ₹10 per minute.
• Inbound and outbound calls can have separate rates. Volume customers get custom rates.
• Add money to a prepaid wallet. Call cost is deducted as you use it.
• Turn on auto top-up with a threshold and an amount, so campaigns never stop mid-way.
• The dashboard shows your balance, every transaction, and your average daily spend.

The per-minute rate includes the AI model, the voice, transcripts, summaries, lead tagging and CRM push. There are no separate model or TTS bills. Voicemail detection and automatic call ending also cut wasted minutes, so you pay for conversations, not for ringing and silence.

For a rough comparison: a human caller in India costs a monthly salary plus management and attrition, and can handle one call at a time. An AI calling agent runs many calls at the same time, and you pay only for the minutes used. Contact us for a quote based on your call volume.

Voice AI architecture

This section is for engineering teams who want to know how a production AI calling agent is built.

High-level flow

Customer phone
   │  PSTN
   ▼
Twilio / Indian DID provider ──► TwiML webhook ──► Voice AI backend
   │  bi-directional Media Stream (G.711 μ-law, 8 kHz)
   ▼
WebSocket call server (Node.js + TypeScript)
   ├─► OpenAI Realtime API (speech-to-speech)
   ├─► Optional custom TTS: ElevenLabs / Cartesia / PlayHT
   ├─► Project facts lookup (per-project, Hindi-aware)
   ├─► Transcript + recording storage (MongoDB, object storage)
   └─► Live events to the dashboard
After the call:
   summary + lead classification ──► CRM push (cron) ──► your CRM

Telephony and audio

When a call starts, the telephony provider asks the backend for instructions (TwiML) and opens a two-way audio stream over WebSocket. Phone audio is G.711 μ-law at 8 kHz. The backend sends this format straight to the Realtime model and back, with no resampling. It also removes any session settings that would switch the audio format, because PCM sent as μ-law turns into harsh static on the call.

The platform supports Twilio for global numbers, and an Indian DID provider for local caller IDs and bulk outbound calls. Call events (ringing, answered, hangup, failed) arrive by webhook and update the campaign in real time.

The conversation engine

• Model: OpenAI's GA Realtime API over WebSocket, speech-to-speech.
• Warm connection pool: a pre-configured model connection waits ready, which cuts the delay before the greeting.
• Barge-in: when the customer starts talking, the agent's audio stops. A short grace window avoids false stops from background noise.
• Voicemail detection: pattern matching on the live transcript for English, Hindi and Indian telecom messages. Detected voicemails end early.
• Tools: the model can call functions, such as looking up project facts or ending the call.
• Custom TTS: when you use a cloned voice, the model returns text and a streaming TTS provider speaks it.

Knowledge and instructions

Each project has its own instructions (greeting, tone, script) and its own list of facts. Instructions are versioned. A PDF upload is parsed and turned into draft instructions. Fact lookup uses fast keyword scoring (about 10 to 30 ms, no external API call) and keeps Devanagari vowel marks, so Hindi queries match correctly. The lookup function can later switch to vector search without changing the call flow.

After the call

• The full transcript is saved turn by turn, and the recording is stored and linked to the call.
• A small, fast LLM classifies the call into one of the 9 outcome categories. It uses only the customer's side of the call. An invalid answer triggers a repair pass, and a safe default is used if that also fails.
• A cron job pushes calls in the chosen lead categories to the CRM, using the field mapping and auth type (bearer token or API key) set for that CRM.

Campaigns and scheduling

A scheduler checks once a minute for campaigns whose start time has arrived. It converts the local working hours into UTC with correct daylight-saving handling. A campaign moves from "scheduled" to "active" with one atomic database update, so two server instances never start the same campaign twice.

Voice AI tech stack

Security

All API calls use JWT authentication, and passwords are hashed with bcrypt. Admin actions use a separate admin login. CRM credentials and field mappings are stored per integration, not in shared config.

Frequently asked questions

Try our Voice AI

If you run a sales, support or outreach team that makes a lot of calls, our AI voice agent can take over the repetitive ones. Email sumit.patel93@gmail.com to see a demo or get a quote.