Insurance sales coach

This was an ambitious one. An assistant for insurance agents that could help them prepare, practice & get through a real customer conversation.

There was a lot of resistance. People were worried about AI taking their jobs. We had to work with them and help them understand what the system actually did.

The BDM (the insurance agent) was still the person having the conversation. Our AI listened and offered help in the chat window when they needed it.

The help started before the call#

A BDM could ask for a pitch script before speaking to a customer. Given the customer’s information, the assistant could also prepare a tailored brochure.

During the call, it listened for places where the BDM missed a detail, moved away from correct policy information, or just got stuck. It could suggest a response or help with an objection in chat.

After the call, it generated a sales coaching report for the supervisor. So the same system supported preparation, the live conversation & the review afterward.

Listening was only half the job#

The assistant had to understand what was happening in the conversation and decide when an interjection would help.

We ingested policy documents and used retrieval to ground its responses. If an agent missed a policy detail, the suggestion needed to come from that policy.

We used a lightweight harness with speech processing around it, similar to the STT/TTS sandwich in the voice agent. For live coaching, the interjection appeared as text in the BDM’s chat window.

We used WebRTC for the live session and its data channel for nudges. Recording, model work and nudges ran in parallel so the session could keep moving while the assistant worked.

Getting that loop to work in real time was one challenge. Helping people understand why it was there was another. Both were part of building the system.

The constraints went beyond the conversation#

For this project, the data had to stay in India. That requirement shaped the choices we could make about where data went and which services we used.

We implemented PII redaction, went through vulnerability assessment and penetration testing (VAPT), and hardened the system based on that work.

These were part of getting the assistant ready for real customer conversations. A working demo was only a portion of the job.

What stayed with me#

Was a hell of a ride. We were working through real-time audio, policy grounding, security requirements & people’s concerns about AI, all in the same project.

It also made me understand the value of owning the model. When you have a specific use case and constraints on where data can go, control over the model starts to matter a lot.

Fine-tuning right-sized models for specific use cases is a direction I’d like to explore further. This project gave me a very practical reason to care about it.