Case study · Lending tech · Customer support
Automated customer support for a leading lending tech platform.
Lending technology · NBFC and bank loans
A maxed-out support team where more headcount was cost-prohibitive. We automated intake and drafting across local languages and English.
90%
Accuracy including multi-lingual cases
3.6 min
Average email response, down from 4.8
2,400
Additional emails handled per day
30%
Lower mean ticket resolution time
The opportunity
A support team at capacity, where hiring was the expensive answer.
This platform connects borrowers with personal loan providers-NBFCs and banks registered with the RBI. As the business grew, its customer support team hit capacity, and response times slipped with it. Adding staff was cost-prohibitive.
Automation had to clear a higher bar than throughput, though: responses needed to work in local languages as well as English, and accuracy and cultural nuance were non-negotiable for interactions that are personal by nature.
Support team at capacity
Operational constraints tightened as the business grew.
More headcount was cost-prohibitive
Scaling the team was not a viable answer to response times.
Multilingual requirement
Replies had to work in local languages as well as English.
Accuracy and nuance imperative
Personalised interactions could not tolerate approximate answers.
The impact
2,400 more emails a day, at 90% accuracy.
Event-driven intake
EventBridge connects ticketing events to downstream processing, with SQS keeping message handling orderly.
Categorised and understood
A Bedrock foundation model categorises messages, extracts entities and reads sentiment and context.
Local languages supported
Hindi, Telugu, Tamil, Kannada and others, with 90% accuracy including multilingual cases.
Drafts, not sends
Bedrock agents retrieve and validate user information, then generate a draft in the identified language.
The stack
What it runs on.
- Event workflow
- Amazon EventBridge, AWS Lambda preprocessing
- Queueing
- Amazon SQS for scalable asynchronous handling
- Model layer
- Amazon Bedrock foundation models and agents
- Integration
- Customer API for information retrieval and validation
The road ahead
Intake is handled. Next is the rest of the conversation.
With categorisation, entity extraction and drafting automated, the same context extends to channels beyond email and to resolving cases rather than only replying to them.
- 01
Context shared across voice and chat
- 02
More languages in production
- 03
Automated resolution for standard cases
Talk to us
Bring one process. Leave with an outcome architecture.
30 minutes on one function, mapped to the architecture behind it. No slides.