Nubank screened its highest-volume support agent in 16,000 simulated conversations, then raised self-service rate 8.82 points live

arXiv·medium signalXBlueskyLinkedInCopy link

The paper runs the Snowglobe simulator on Nubank's Card Delivery and Card Management chat agents in Brazil. Simulated evaluator scores tracked production across 4 deployed versions, and simulation-guided iteration raised transactional NPS by 36.69 points in a live A/B test. An open-weight configuration chosen from 16,000+ simulated conversations lifted self-service rate by 8.82 points to Nubank's highest level, with no significant tNPS change. It is one of the few production-scale reports showing that pre-deployment simulation predicts live results for a regulated-industry agent.