How an AI voice agent works on a phone call
An AI calling system joins three pieces in real time. Speech recognition turns what the caller says into text. A language model decides what to say next, using your instructions and your knowledge base. Speech synthesis turns that reply into a voice. Good systems do this fast enough that the caller can interrupt, and the agent stops and listens.
Around that loop sit the parts that make it useful for sales: a phone number, a dialler that respects calling hours and retry rules, a transcript and summary for every call, and a way to transfer to a human when the AI is stuck or the lead is ready to buy.
- Speech to text: understands the caller, ideally including mixed Hindi and English.
- Reasoning: follows your script, answers from your content, asks qualifying questions.
- Text to speech: replies in a chosen voice and language.
- Telephony: the number, call routing, recording and transfer.
- After the call: transcript, summary, tags, lead score and the next follow-up.
Inbound and outbound calling
Inbound AI answers calls to your number. Typical uses are after-hours enquiries, missed calls, appointment booking and first-level support. The value is simple: no call goes unanswered, and every caller gets a transcript and a next step.
Outbound AI places calls to a list. Typical uses are calling webinar registrants who did not attend, qualifying new leads within minutes of sign-up, renewal reminders and re-engaging old enquiries. Outbound carries more regulatory weight in India because it is commercial communication under TRAI rules, so consent, calling hours and caller ID matter more here.
| Inbound AI | Outbound AI | |
|---|---|---|
| Who starts the call | The lead or customer | Your business |
| Common uses | After-hours answering, missed calls, bookings, FAQs | Lead qualification, no-show follow-up, reminders, re-engagement |
| Main risk | Wrong or incomplete answers | Unwanted calls, consent and TRAI compliance |
| What to measure | Calls answered, resolved, transferred | Connect rate, qualified leads, bookings, complaints |
Indian languages and accents
Most Indian buyers switch between English and a regional language within one sentence. Test any vendor with real recordings of your callers, not a demo script. Listen for three things: whether the agent understands code-mixed speech, whether names, numbers and prices come out correctly, and whether the voice sounds natural in the regional language.
WebinarRun AI voice supports 97 languages with auto-detect, including 14 Indian languages: Hindi, Tamil, Telugu, Kannada, Malayalam, Marathi, Bengali, Gujarati, Punjabi, Odia, Assamese, Urdu, Nepali and Sindhi. Whatever tool you pick, run a test call in each language you plan to use before you send a campaign.
What AI calling costs in India
Vendors price voice AI in two main ways. Per-minute pricing charges for each connected minute, sometimes with a monthly platform fee on top. Credit or wallet pricing sells a balance in rupees or dollars that is drawn down by calls and other AI usage. Both can work. What matters is the all-in cost per useful conversation.
Public price lists from Indian voice AI vendors in 2026 mostly sit between ₹2 and ₹12 per minute, and one Indian market review puts ₹3 to ₹6 as the common mid-market band. Global platforms advertise roughly $0.05 to $0.30 per minute, and several of the lowest headline prices exclude speech, model or telephony charges that are billed separately.
Add the costs that are easy to miss: phone number rental, unanswered and very short calls, retries, transfers to humans, call recording storage, and setup or integration work. A quote that looks like ₹3 per minute can land at two to three times that once these are counted. Use our calculator to compare with a telecalling team using your own numbers.
| Cost item | What to ask |
|---|---|
| Per-minute or credit rate | Is it per connected minute? Does it include speech, AI and telephony? |
| Phone numbers | Monthly rent per number, and KYC needed? |
| Platform or seat fee | Fixed monthly fee, minimum commitment or annual contract? |
| Failed and short calls | Are unanswered or sub-10-second calls billed? |
| Extras | Recording, transcripts, transfers, CRM sync, setup |
Ranges are from public vendor price lists and market reviews listed in Sources; check current rates before buying.
Phone numbers, KYC and caller ID
In India, business phone numbers come with KYC: expect to submit company documents before a number is activated. Ask the vendor whose name the number is registered in, whether you can keep it if you leave, and how caller ID will show to the person you call.
TRAI now treats calls placed by software, including artificial voice calls, as Application-to-Person (A2P) calls. Its Third Amendment to the commercial communication regulations, notified on 18 September 2026, requires senders to declare A2P calling and the numbers used to their telecom provider in advance, and undeclared A2P calls are treated as unsolicited commercial communication. Ask any vendor how this declaration is handled for the numbers you will use. Our guide on AI calling law in India covers this in more detail.
With WebinarRun you can buy Indian numbers inside the app at ₹499 per month each, after business KYC. Numbers in other countries are not available yet.
What to check before you choose
Run a two-week pilot on one use case, such as calling registrants who missed a webinar, and judge the tool on outcomes rather than the demo.
- Language quality on your own callers, including code-mixed speech.
- Latency: does the agent reply quickly, and can callers interrupt it?
- Knowledge: can it answer from your documents and live systems, and does it show what it could not answer?
- Compliance controls: calling hours in the contact’s local time, daily caps per person, do-not-contact lists, and recording notices.
- Handover: warm transfer to a person, and a clear record of why it transferred.
- Follow-up: does an unanswered call trigger a WhatsApp message, and does the outcome update the lead score?
- Data: who owns recordings and transcripts, where they are stored, and how you delete them.
- Pricing: the all-in cost per connected minute and per qualified lead.
Where WebinarRun fits
WebinarRun AI calling runs outbound campaigns (audience by tag, chosen contacts or CSV; calls per minute, concurrency, retries, calling hours in each contact’s local time and a daily cap per person) and answers inbound calls always or after business hours. Calls are recorded and transcribed, summaries go into contact notes, and leads are re-scored after a call. Unanswered calls can trigger a WhatsApp follow-up.
The difference from a standalone calling tool is the loop: the same AI Brain answers in webinars, WhatsApp and calls, so a lead who asked a question in your webinar chat gets a call that already knows the context. Outbound calls on WebinarRun are made by the AI only; there is no dialler for human agents.