“Send one of your people to our Mumbai office. Let me see what they can do, and then I'll decide.”

That was how our first meeting ended. We had gone to Gandhinagar to meet the CEO of an NBFC,¹ for what was meant to be a one-hour conversation. It lasted three. We wanted to build their entire ERP, and after three hours we had no yes, only a challenge. We didn't even know what the test would be. We would only find out when someone got there.

I had nothing to prepare from: no requirements, no brief, not even a hint. I could read up on technology, but what about the business? Lending was a completely new world to me.

So I sat with the team, one person at a time, and asked how they worked.

I had come to talk about an ERP. On the third day, they handed me something else entirely.

They had a single telecaller, and one campaign could bring in around 700 leads in two days. How could one person keep up?

My first instinct was to automate the calling with AI. I knew it could never run entirely on its own, but the more I learned about what a telecaller actually does, the worse that idea looked. Replacing her would have been a mistake: qualifying a loan takes trust and judgment, and that has to stay with a person.

To find good leads, I first had to understand what made one good. The person who knew was the telecaller, so I sat beside her and learned how she judged them. It begins with knowing who is on the line, and there are two kinds of caller:

• A DSA, or Direct Selling Agent: a broker who brings in loans for other people and earns a commission.
• A direct customer, who wants the loan for themselves.

I learned how commissions work, both one-time and recurring, and the lending policy behind them. Then I spotted a gap: there was no way to tell a DSA from a customer.

Asking directly can backfire. If the caller is a genuine customer, it leaves a bad impression. So the AI agent, whom I named Riya, would ask a few natural questions that bring out the answer. Here's that part of her instructions:

Do NOT ask "are you a DSA".Ask: "Aapko loan chahiye?"(Do you need a loan?)  For themselves → CUSTOMER  For someone else, ask:  "Wo aapke relative ya family hain, ya aapke client?"  (Are they a relative, or your client?)    Relative or family → CUSTOMER    Client, ask: "Ji, aap loan ka kaam karte hain?"    (Do you work in loans?)      Yes → DSA      No  → CUSTOMER

From there, a DSA takes one of two paths. An existing one shares their unique code, and Riya verifies it against the CRM during the call. A new one is registered on the spot, with their PAN² and details.

The tools came last. Frappe CRM is open source, so it became the home for every lead. For the voice I chose Sarvam, which handles Indian languages far better, and an NBFC's data has to stay in India.

Teaching Riya to sound human

The logic was only half the job. The other half was making a call with an AI feel natural, and that came down to many small rules:

• She says upfront that she's an AI agent and that the call is being recorded.
• She blends English into Hindi, Marathi or Gujarati, the way people really speak. Pure Hindi sounds like a government announcement, and nobody translates loan, EMI or CIBIL.³
• She always uses feminine verb forms: bol rahi hoon, never bol raha hoon. One slip and the caller notices.
• She asks the questions that decide a loan first, because most people give an AI agent about five minutes.
• She never claims a code was found unless the CRM actually returned it.

None of this helps unless the telecaller can act on it, so I shaped Frappe CRM around her work. Leads move through three stages, and the best ones stay at the top of her list:

• AI-Qualified: a strong lead, ready for her.
• Follow-up: needs a call from a person.
• AI Follow-up: needs another call, and the AI can make it.

When she opens a lead, she doesn't need to listen to a recording. The summary, the next step and the scores are waiting for her, and the full transcript is one click away.

An AI to check the AI

Nobody can listen to every call, so I added one more layer to the CRM. A second AI reads each call, checks that Riya followed the process, and scores the lead.

I stayed in Mumbai for a month, and the MVP was ready in 15 working days. Then the customer saw it working.

“Fantastic work, in such a short time.”

The CEO gave us the go-ahead, and I came back to Ahmedabad with the project won. The customer who once asked to test one person went on to trust us with their entire LOS / LMS platform,⁴ and I lead that work with my team today.

Good systems aren't built from behind a screen. They start with listening to the people who do the work every day.