‹ All case studies
Running AI at bank scale · Banking
Jana Small Finance Bank: 50+ voice agents and 70+ dashboards for one bank
Collections, EMI reminders and customer service on voice AI, in four languages, at 10M+ calls a quarter. The AI collection agents now collect more than the human ones. Now extending to WhatsApp and inbound.
Teams servedCollections, Customer Service
ChannelVoice, WhatsApp
My roleCustomer success lead, solution design
What was hard
One agent is a project. Fifty is an operation. Every campaign, language and customer segment needed its own agent, every team at the bank wanted its own view of the numbers, and at three million calls a month a small mistake gets very expensive very quickly.
What I cracked
- 50+ agents, four languages. Reminder, collections and service agents, each tuned to its campaign and segment, running in production.
- 70+ dashboards. Built for the teams that run collections and service, so each one reads its own numbers without asking us.
- Found the waste. An analysis of 16.3M dials showed 31.7% were going to numbers that had never picked up, not once. That became dialling rules: suppress dead numbers, move volume out of the worst hours.
- A model switch, measured. When the bank's post-call analyser moved to a new model, I ran a blind before-and-after check against call transcripts. Accuracy on promises to pay had dropped, so I rewrote the prompts, re-ran 13,000 calls and brought it back to 84 to 96%.
- WhatsApp, then inbound. A WhatsApp chatbot POC that is truly agentic where it needs to be and flow-based where the bank needs control, and now an inbound use case on top.
The number
42% vs 35%collection rate, AI agents against human agents
50+voice agents live, in 4 languages
10M+calls a quarter
5.17Mwasted dials found in 16.3M
At this scale the model is the easy part. Running fifty of them, and knowing which calls not to make, is the job.
Get in touch
I can walk you through how this was built. Email is quickest.