Nubank Accelerates AI Development by 80% Using Large-Scale Synthetic Customer Simulations

Nubank has revealed a new methodology for developing customer-facing AI agents by testing them against synthetic personas before deployment. This simulation-first approach significantly reduces development cycles and improves customer satisfaction metrics, offering a blueprint for financial institutions looking to scale generative AI safely without risking real-world user friction or service degradation.

What was announced

Nubank’s AI Research team published findings from a study titled “Screen Before You Serve: Simulation for Production Customer Experience AI Agents at 140M Scale”, detailing a controlled testing environment where synthetic customers interact with AI agents. The study focused on two specific agents within the credit card lifecycle: Card Management, which handles address updates and customer-requested reissues, and Card Delivery, which manages delivery status and reissues. By running more than 16,000 simulated conversations across 29 model configurations, the team reduced the average development time for a new agent version from 21.2 days to just 4.4 days.

The simulation phase allowed researchers to identify a version of the Card Management agent that reduced unnecessary handoffs to human support from 22.4% to 16.0%, a decrease of 6.4 percentage points. In subsequent live A/B testing against the Card Delivery agent, the simulation-refined Card Management tool recorded a 4.9-percentage-point higher self-service rate and a 36.7-point increase in transactional Net Promoter Score (tNPS). Furthermore, testing a refined Qwen configuration with reasoning enabled led to an 8.8-percentage-point increase in self-service and a 25% reduction in p95 response latency. Following these results, the improved Card Management agent has been rolled out to the company’s customer base in Brazil.

"Every change to an AI agent is a bet on the customer experience. Simulation lets us place those bets on synthetic customers first, so only the changes that clearly win reach real people,"

Rohan Ramanath, General Manager of AI at Nubank.

The companies involved

Nubank is one of the world’s largest digital financial services platforms, serving millions of customers across Latin America. Headquartered in Brazil, the company has established itself as a dominant force in the neobanking sector by challenging traditional banking incumbents with a mobile-first, low-fee model. Since its inception, Nubank has expanded its footprint beyond its home market into Mexico and Colombia, diversifying its product suite from its original credit card offering to include savings accounts, investments, and insurance.

The company operates under the parent entity Nu Holdings Ltd. and has consistently prioritized technological autonomy, building much of its core infrastructure in-house. This focus on proprietary tech is reflected in its dedicated AI Research team, which focuses on integrating machine learning and generative AI into the customer journey. As the firm matures, it has increasingly moved toward a multi-product ecosystem, leveraging data to drive personalization and operational efficiency. Its market position is defined by high customer engagement levels and a rapid pace of product iteration, often serving as a bellwether for the broader fintech industry’s adoption of emerging technologies.

What FF News has reported before

FF News has closely tracked Nubank’s strategic shifts and leadership changes. Recently, the publication covered how Nubank CTO Eric Young to Step Down; CEO David Vélez to Lead Tech Transition, marking a significant moment for the firm’s technical leadership. The company has also been active in brand positioning, as seen in the Nubank Unveils Major Brand Refresh to Support Global Expansion and Multi-Product Ecosystem.

In regional markets, Nu Mexico Launches Score Nu to Empower 16 Million Users with Credit Insights, demonstrating the firm's commitment to financial education. Additionally, the bank has leveraged high-profile marketing, such as when Patrick Dempsey Joins Nubank Ultravioleta for High-Performance Mercedes-AMG F1 Campaign to promote its premium offerings.

What this means

This move toward synthetic customer simulation addresses one of the primary bottlenecks in fintech AI: the risk of testing Large Language Models (LLMs) on real users without a safety net. By moving the majority of the trial-and-error phase into a simulated environment, Nubank is effectively lowering the cost of failure for its engineering teams. This puts pressure on traditional banks that still rely on slower, manual QA processes or limited pilot groups. The significant jump in tNPS suggests that "reasoning" models, when properly screened, can bridge the gap between robotic automated responses and helpful human-like interaction. The industry must now consider whether synthetic testing will become a standard regulatory or compliance requirement for AI safety in finance.