Nubank Integrates OpenAI GPT-4o for Customer Service and Fraud Detection

Nubank has integrated OpenAI's GPT-4o and GPT-4o mini to automate customer support, streamline internal knowledge retrieval, and enhance fraud detection for its 114 million customers across Brazil, Mexico, and Colombia. These implementations have reduced chat response times by 70% and increased query resolution speed by 2.3x.

AI-Powered Customer Support and Automation

Nubank uses GPT-4o to power both a customer-facing AI Assistant and an internal Call Center Copilot, significantly reducing the burden on human agents while maintaining high Transactional Net Promoter Scores (tNPS).

AI Assistant

An AI Assistant powered by GPT-4o handles over 2 million monthly chats and emails. It is designed to resolve up to 55% of Tier 1 inquiries automatically, managing up to five automated interactions before escalating the conversation to a human agent.

Call Center Copilot

Over 45% of Nubank agents utilize a Call Center Copilot built with GPT-4o. This tool leverages real-time multimodal capabilities (text and speech) to provide agents with:

  • Next-reply suggestions: Recommended best answers to ensure accuracy and empathy.
  • Chat summarization: Immediate context for past or ongoing conversations.
  • Step-by-step guidance: Simplified workflows for complex technical queries to reduce staff cognitive load.

Enterprise Knowledge Search

Nubank deployed a custom enterprise search solution using GPT-4o and GPT-4o mini to provide over 5,000 employees monthly with instant access to internal policies, brand guidelines, and FAQs.

To ensure accuracy and avoid siloed information, the system employs Retrieval-Augmented Generation (RAG) and semantic search. The models are fine-tuned on Nubank's domain-specific content to prioritize the most relevant results for developers, customer support agents, and new hires during onboarding.

Fraud Quality Assurance via GPT-4o Vision

Nubank is piloting a vision-based solution using GPT-4o vision to improve the consistency and quality of fraud incident responses. By combining natural language processing with image recognition, the system analyzes transaction records, customer communications, and submitted documents to identify patterns and anomalies indicative of fraud.

This process is conducted under the oversight of dedicated teams and in compliance with regulatory requirements and fraud prevention policies.

Sources