Agent, not bot: what the word means here
A bot follows a flowchart. An agent has a goal, a brief and a set of tools, and it works out the path during the call. If a caller answers the third question before you ask the first, an agent notes the answer and moves on; a flowchart asks again.
That difference is why calls with an AI calling agent feel like conversations rather than an IVR. It is also why the agent needs boundaries — what it may promise, what it must never say, when to end the call — written into its brief.
The four things an AI calling agent does on every turn
| Stage | What happens | Why it matters on Indian calls |
|---|---|---|
| Listen | Speech is transcribed as the caller talks, including mixed-language speech | Callers say "Sunday ko visit possible hai?" — one sentence, two languages |
| Decide | A language model reads the transcript, the brief and the facts it was given, then chooses a reply or an action | Handles off-script questions such as parking, EMI options or the nearest branch |
| Speak | The Swara voice engine speaks the reply in the chosen voice | A voice that sounds local keeps people on the line |
| Act | It books a calendar slot, writes to a spreadsheet, calls an API or sends a WhatsApp message | The call produces an outcome, not just a recording |
On Outpero, the time from the caller finishing a sentence to the agent starting its reply is 0.8 to 1.3 seconds on average. The agent also handles barge-in: if the caller interrupts, it stops and listens instead of finishing its line.
A call, turn by turn
Here is how an Outpero agent call for a two-wheeler dealership in Chennai might run after a Google Ads lead form:
- 01Within 30 seconds of the form, the agent calls and opens in Tamil: "வணக்கம், நான் Ravi Motors-லிருந்து பேசுறேன். நீங்க இப்போ scooter test ride-க்கு enquiry பண்ணீங்க, ரெண்டு நிமிஷம் பேசலாமா?" (Hello, I'm calling from Ravi Motors. You just enquired about a scooter test ride — can we talk for two minutes?)
- 02The caller replies in Tanglish that they want an electric scooter and asks about the on-road price. The agent answers from the price list in its brief.
- 03The caller asks about exchanging an old bike. The agent records the model and year as captured fields.
- 04It offers Saturday 11 am or Sunday 4 pm, books the chosen slot in the showroom calendar, and sends the location on WhatsApp.
- 05The call ends; the sales team receives a summary, sentiment and the fields — model, exchange vehicle, test-ride slot — at their webhook.
What to give an AI calling agent before it goes live
- —A goal: "book a test ride", "confirm the COD order", "get a promise-to-pay date".
- —The facts it is allowed to use: prices, timings, branch addresses, policies.
- —Fields to capture: the exact data your team needs after the call.
- —Hard rules: no discounts beyond a set limit, no medical advice, end politely if the person asks not to be called.
- —Tools: which calendar, which spreadsheet, which API, which WhatsApp message.
Hear it on a real call before you read further?
Where AI calling agents get used in India
The agent shines where the conversation is short and structured but the replies vary: real estate site-visit booking, hospital and clinic appointment confirmation, admission enquiry follow-up for colleges, loan application follow-up for NBFCs, insurance renewal calls, candidate screening for recruiters, and COD confirmation for D2C brands. The same agent can make those calls and also pick up when customers ring your business number, at any hour.
Languages it can hold a call in
Hindi, Telugu, Tamil, Marathi, Kannada, Malayalam and Indian English, each with natural switching to English mid-sentence in the same voice. Bengali, Gujarati, Punjabi, Odia, Urdu and Assamese are not available, and there is no separate per-dialect training, so test with callers from your own region before going live.
Cost per agent conversation
Outpero has no setup fee or monthly base. You pay per connected minute from prepaid credits: ₹5 on Standard, ₹6.8 on Premium native-speaker voices with ultra-low latency. A three-minute test-ride booking call costs ₹15 or about ₹20 depending on the tier; unanswered dials cost nothing. A dedicated number is ₹649/month. Against a human caller at roughly ₹18,000 a month who can take one call at a time, the agent's cost scales with the work done, not with headcount.
Compliance for AI calling agents
An AI calling agent must follow TRAI's TCCCPR 2018 like any caller: no promotional calls to DND / NCPR-registered numbers, DLT registration for commercial senders, and calls within daytime hours. Outpero's agents call from a DLT-registered line and bulk lists are DND-screened. You decide what the agent discloses; having it say up front that it is an AI assistant is good practice.
How to evaluate an AI calling agent
Run three tests before you buy: interrupt it mid-sentence; ask it something not in the brief and see whether it admits it does not know instead of inventing an answer; and switch languages halfway through. Then check what lands in your CRM afterwards. An agent that talks well but leaves you without structured data is half a product.
Limits worth knowing
An agent is only as good as its brief; vague instructions produce vague calls. It should not be left to negotiate prices, give medical or legal advice, or soothe a furious customer. It can transfer a call to your team when a person should take it; otherwise, when someone asks for a person, it captures the request and flags it in the summary and webhook so your team calls back.