How a large Indian health insurer uses voice AI agents to simulate, evaluate and improve high-stakes sales conversations at scale.
Voice agents simulate real customer personas, enabling teams to practise, assess and refine interactions before they go live.
0+
Scored competencies per interaction
each with evidence traces
0+
Languages with natural code-switching
English, Hindi, Tamil, Telugu, Malayalam and more Indic languages
0 × 6
Customer personas
across three difficulty levels
0
Turns of consistent persona memory
in one multi-turn conversation
The opportunity
Thousands of agents, and no way to practise the conversations that decide a sale.
A retail health-insurance sales force is large, distributed, multilingual and turns over. The conversations that decide a sale, a price-skeptical customer, a family that distrusts claims, a buyer who keeps postponing, follow patterns that can be practised. Classroom role-play cannot reach thousands of agents, and it cannot score them consistently. The insurer wanted every agent to practise those conversations, and their managers to see who was ready to sell.
Role-play does not scale
A trainer can run a handful of sessions a day; the sales force needs thousands.
Scoring is inconsistent
Two trainers score the same call differently, so readiness cannot be compared.
Managers cannot see readiness
Performance, capability and product knowledge were invisible until the customer conversation.
Customers switch languages mid-sentence
The system handles mixed-language conversations seamlessly.
The impact
A customer who pushes back, in the agent's own language, with a scorecard after every call.
Rehearse plays the customer, not the agent. Personas and scoring rubrics are tuned by the product team to the insurer's product context and organisational guardrails, so what gets practised and how it is scored follows the sales model your team has learnt over years.
100%
Sessions scored
Every practice call ends in a scorecard with evidence quoted from the conversation, compliance flags for misinformation and pressure-selling, and coaching tips in the language of the session.
6+
Competencies per interaction
Each scored with evidence traces, so a manager can see why, not just how much.
40
Turns of persona memory
Fixed facts, family details, quotes already received, health conditions, held consistently through a full negotiation. Objections in character; concession only when genuinely convinced. Learnings from past wins transcribed from call recordings and much more.
3
Difficulty levels per persona
The same customer, playable at easy, medium or hard, so agents move up as they improve.
The stack
What it runs on.
- Model layer
- Amazon Bedrock foundation models
- Voice layer
- Sarvam AI and ElevenLabs
- Context engineering
- Advanced retrieval-augmented generation, over the insurer's product documents
- Learning system
- Completions, scores, duration, difficulty and language into the LMS over SCORM and xAPI, with a secure link to the replay
- Access
- Browser on desktop and mobile; single sign-on for staff
- Scope
- Scalable to 100K+ sales agents
The road ahead
In a production pilot with the sales organisation. Next: rollout.
The platform is currently in a production pilot with the insurer’s sales organisation.
- 1
Assessment campaigns before product launches, with results integrated into dashboards and LMS systems
- 2
Region-specific customer personas built to match the insurer’s operating context
- 3
Outcome metrics to be shared once the evaluation concludes and approvals are in place
Client under NDA. Reference available on request.
Talk to us
Want to improve your sales conversations?
We agree the target, implement within your environment, and measure outcomes against your baseline.
