Nubank
Nubank uses OpenAI models for a customer chat assistant and an agent copilot
Brazilian digital bank Nubank built a GPT-4o customer assistant and a support-agent copilot with OpenAI. OpenAI's case study says the assistant resolves 55% of Tier 1 inquiries and cut chat response times by 70%.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
55%
Share of Tier 1 inquiries the AI assistant resolves1
over 2 million
Monthly chats handled by the AI assistant1
70%
Reduction in chat response times1
over 45%
Share of Nubank agents using the call center copilot's key features1
What it took
Cost, team size and timeline were not disclosed in the sources.
The problem
Nubank serves a very large customer base and wants fast, accurate support as it grows. Support agents also need quick access to policies and product answers.
What they built
Three things, per OpenAI. An internal search tool for staff, a copilot that suggests replies and summarises chats for human agents, and an AI assistant that answers customers directly. The assistant and the copilot run on OpenAI's GPT-4o models.
How they did it
The internal search tool came first and uses retrieval, where the system looks up company documents before answering, with GPT-4o and GPT-4o mini. The copilot draws on Nubank's knowledge base and chat history. The customer assistant hands over to a human agent after up to 5 automated interactions.
Results
OpenAI reports the assistant resolves 55% of Tier 1 inquiries and handles over 2 million chats a month, with chat response times down 70%. Over 45% of Nubank agents use the copilot's key features. Baguete reports that up to 50% of calls are resolved without a human agent. All numbers are company or vendor claims.
All reported figures
| Measure | Figure | Citation |
|---|---|---|
| Share of Tier 1 inquiries the AI assistant resolves | 55%1 | |
| Monthly chats handled by the AI assistant | over 2 million1 | |
| Reduction in chat response times | 70%1 | |
| Share of Nubank agents using the call center copilot's key features | over 45%1 | |
| Share of calls resolved without a human agent, as reported by Baguete (Portuguese: até 50%) | up to 50%2 |
Lessons for your project
Pair a customer-facing assistant with a tool for human agents. Nubank built both, so the hard cases that reach people arrive with summaries and suggested replies.
Cap the automation. The assistant makes up to 5 automated interactions before handing over to a person.
Start with an internal use. Nubank first built an enterprise search tool used by over 5,000 employees monthly before the customer-facing work.
Check headline numbers against each other. Sources disagree on the resolution rate (55% vs up to 50%), so state who reports which.
Caveats
- The figures come from OpenAI's case study and a news article that repeats them. No Nubank-owned source was read. The page is vendor-reported.
- The sources disagree: OpenAI's page says 55% of Tier 1 inquiries, while it also says up to 50% in another passage, and Baguete says up to 50%.
- The OpenAI page returned a block to automated access, so a web-archive copy was read. The page itself carries no date, so the Baguete article date (12 March 2025) is used and the deployed date is the announcement month.
- The OpenAI page also describes a fraud-review pilot using GPT-4o vision, which is not covered here.
Sources
- Nubank elevates customer experiences with OpenAI
OpenAI · Undated · Vendor case study · openai.com
- Nubank usa soluções da OpenAI
Baguete · 2025-03-12 · News · baguete.com.br
Summarized in our own words from the sources above. Spotted an error? See corrections.