I built an AI voice agent that calls new leads back within a minute

Most people who fill in an enquiry form are ready to talk for about an hour. After that they've called someone else, or they've forgotten why they asked. So I wanted to see what happens when every new lead gets a phone call within a minute, at any time the rules allow.

This is my second official demo: an AI voice agent built for an insurance advisor I know. It calls people who asked to be called, has a real conversation with them in Hindi or English, and books a time for the advisor to call them back properly.

The problem I started with

The advisor sells life and health insurance on his own. He has no steady flow of leads, so he finds customers the old way: cold-calling from his personal mobile. That's slow, it wears people down, and India's rules on promotional calls from ordinary 10-digit numbers keep getting stricter.

The fix isn't more calls. It's calling the right people at the right moment: people who raised their hand and said "call me", reached before their interest fades.

What the agent does on a call

1. Says who it is. It's an AI assistant calling on behalf of the advisor, the call is recorded, and the person can ask for a human at any time.
2. Checks it's a good time. If not, it offers to call back later.
3. Asks what they're looking for. Term cover, health insurance, saving for a child's education, investing, retirement.
4. Asks only the questions that matter. Age, who depends on them, a rough income range, salaried or self-employed, and what cover they already have. Budget and risk comfort come up only if they want to save or invest.
5. Answers their questions in plain language. What term cover is, how waiting periods work, why the cover amount matters.
6. Books the advisor's call. The person names a time that suits them, the agent checks it, and offers the soonest free times if they have no preference.

Everything it learns is written up as a short brief, so the advisor starts the real conversation already knowing what the person needs.

The rules I gave it

An insurance call is not the place for a chatbot that says yes to everything. So the agent has hard rules:

• It only calls people who ticked a box asking for a call.
• It never quotes a premium, a return or a tax benefit, and never says a plan is "the best" for someone. That decision stays with the human advisor.
• If someone says "don't call me", it apologises, ends the call and never calls that number again.
• It only calls within set hours, Monday to Saturday.

I tested these on purpose: asking for a guaranteed return, asking which company is best, pretending to be the wrong number, asking it to stop calling. It held the line every time.

Why I rewrote it to ask questions first

My first version had the lead fill in a long form (age, income, family) and then the agent called to explain the result. Then I asked myself a simple question: if the form already has everything, why call?

There was no good answer. Real leads, like the ones from ads, arrive with just a name and a phone number. So I flipped it. The form now asks only for a name, a number and permission to call. The call is where the discovery happens. It's also a better conversation, because people say far more when they're asked than when they're filling in boxes.

What went wrong along the way

• Hindi numbers. The agent understood Hindi and English sentences fine, but numbers spoken in Hindi or Hinglish ("पैंतालीस", "आठ लाख") were often misheard the first time. It recovered by confirming them back, but that's clumsy. The fix was simple once I found it: give each agent one language. A Hindi agent hears Hindi numbers correctly. So there will be two agents, and the lead picks Hindi or English on the form.
• Days and times. At one point it told a caller that today was Wednesday when it was Thursday, and called 5 pm "outside hours" when it wasn't. A language model is not a calendar. Checking dates and free slots now happens in a small tool the agent calls, not in its head.
• Loops. Early versions repeated the introduction, or asked the same question three times. Shorter instructions and clear steps fixed most of it.

What it costs

At about ₹6.8 a minute, a four to five minute discovery call costs roughly ₹27 to ₹34. For a solo advisor, a few hundred of these calls a month comes to under ₹10,000, and the agent never forgets a follow-up or skips a Sunday enquiry.

But cost isn't the point. Speed is. The agent's real job is making sure nobody who asked for a call waits until tomorrow.

What's next

Next it goes to work for the advisor as a pilot: connected to his calendar, saving every lead with the call summary, and sending the lead a confirmation of the booked call. I'll share what happens, including the parts that don't work.

If you run a business where leads go cold because nobody calls back fast enough, this is the kind of thing I build.