How a Mid-Tier Indian NBFC Cut Collections Cost 71% with a Multilingual Voicebot

The Challenge

The client ran a 200-agent collections call centre handling 400K+ outbound calls/month across English, Hindi, Tamil and Telugu. Cost per contact had crept to ₹62; recovery on Day-3 and Day-7 buckets was declining as customer-language complexity overwhelmed agents. The compliance team also flagged inconsistent disclosure language across agents.

Our Solution

Iedeo deployed a 4-language outbound voicebot for early-bucket collections (Day 1-15). The bot calls customers, identifies them via OTP, plays bucket-appropriate dialogue with consent-aware compliance scripts, captures intent, schedules promise-to-pay, and books a human callback only on genuine dispute or hardship. RBI-aligned audit logs, full call recording, and per-segment A/B testing of dialogues.

Architecture & Stack

• Inbound + outbound telephony via Exotel SIP trunk
• Whisper-based ASR fine-tuned on Indian-language collections corpus (12K hours)
• Claude Sonnet for dialogue management with custom guardrails
• Coqui XTTS voice cloning for consistent Hindi/Tamil voice
• Self-hosted on client VPC (DPDP residency requirement)
• CRM write-back to LeadSquared with promise-to-pay and dispute reasons
• Real-time supervisor dashboard with anomaly alerting

Timeline

3 weeks discovery → 6 weeks build → 4 weeks pilot (10% traffic) → 4 weeks ramp to 100%

“We expected cost savings. We did not expect CSAT to go up. Customers prefer the bot — it never gets impatient at 9 PM.”