Intelligent Customer Support for the Masses: How Nubank Does It
summary
Nubank describes a production system that routes and auto-answers support using Pusher, Zendesk, AWS, a custom queue manager próximo, and an internal web app shuffle. Text models use TF‑IDF and logistic regression, stacked with a contextual model; routing hits ~85% accuracy for email and ~75% for chat. Auto-reply handles about 10% of incoming email (13k/day total), cutting minutes per reply. The pipeline records human corrections to retrain models and shifts simple macros into automated responses, freeing agents for complex cases.
AI-written summary · about
chapters (19)
- 0:00Introduction to Nubank
- 2:19Why customer support matters
- 5:12Customer support design team goals
- 6:59Self-help and human support channels
- 9:08Queue management system Próximo
- 10:55Squad-based support organization
- 13:25Routing emails and chats efficiently
- 16:15Rules vs machine learning routing
- 18:43Extracting info from customer text
- 20:28Training the routing model
- 21:53Contextual model for routing
- 23:16Routing model results and impact
- 24:39Reducing wait times in support
- 25:22Handling simple questions with macros
- 28:49Auto-reply system for simple queries
- 30:58Building a modeling cycle early
- 32:23Benefits of using rules first
- 33:50Auto-reply for chat interactions
- 35:14Future improvements and hiring
related talks
32mThe AI Agent Path to Prod: Building for Reliability
45mFinding the balance between human and bot interaction
21mHow nib transformed customer experience
36mAI Agents in Automation | 3 Real Production Cases with ROI
19mAutomating 90% of support tickets: How we built Vercel's support agent
33mOptimizing Bank Operations: Design Patterns, NFR, and GenAI Fusion
42mCurated Panel: A Practical Approach to AI Transformation Across Player Support
30mAI Driven Transformation in Contact Centers: Enhancing Customer Experience