Inbound support on voice AI · Life insurance
HDFC Life: insurance support calls, answered by an AI agent
I built and solutioned the inbound agent that now takes 16,000+ HDFC Life support calls a month and resolves 41% of them with no human, one of the first of its kind in Indian BFSI. Together with renewal calling in 11 languages, the program runs 180K+ calls a month.
What was hard
Inbound is harder than outbound. You don't choose who calls, or why. A policyholder might want their fund value, a premium receipt, a surrender quote or a revival, and in a regulated insurer every answer has to be right, or handed to a person who can make it right. On the renewals side, a human agent ends a call in about ninety different ways, and each one means a different next step.
What I cracked
- An inbound agent that knows the policy. It sits behind HDFC Life's contact-centre routing and answers the high-volume queries itself: premium due and payment, fund value and NAV, policy status, surrender and revival.
- Clean hand-offs. When a query needs a person, the transcript and intent go with it, so the customer never starts again. If they were just called about a renewal, that context is already loaded.
- Ninety outcomes, two questions. For renewals, sitting with the collections team rather than IT, we reduced every call outcome to two: is this customer still callable, and when? Then wrote the retry rules that follow.
- Personal at scale. Over 100 policy fields per call, from grace date to revival quote, in 11 languages.
- A judge on every call. An LLM scores every inbound call on 16 parameters plus a hallucination check, against the policy data the agent actually pulled. We checked it against HDFC Life's own audits, then tuned the agent week by week: resolution went from 23% in September to 41% in the first week of November.

The number
The model was the easy part. The rules came from people who had been answering these calls by hand for years.
Get in touch
I can walk you through how this was built. Email is quickest.