Nu Holdings put its nuFormer AI model into production in 2025 and launched over 100 products and features across its Latin American markets while keeping risk-adjusted net interest margins stable quarter-over-quarter. The deployment demonstrates how machine learning models can accelerate product velocity without sacrificing underwriting quality.
Inter & Co welcomed 7 million new clients in 2025, marking its strongest annual performance. The fintech maintains industry-leading funding costs at 65.6% of Brazil's CDI benchmark rate. Newer client cohorts are transacting faster and more frequently than older ones, indicating improved targeting from automated credit decisioning.
The AI models handle real-time portfolio monitoring alongside automated underwriting. Traditional banks process credit applications in days using manual review processes. AI systems evaluate hundreds of data points in seconds, including transaction patterns, payment histories, and behavioral signals invisible to rule-based systems.
Brazilian fintechs face unique credit risk challenges. The country's CDI rate fluctuates significantly, compressing margins during rate spikes. AI models adjust credit lines and pricing dynamically as portfolio conditions shift, maintaining profitability across rate cycles.
Nu's stable risk-adjusted margins despite rapid expansion suggest the nuFormer model identifies creditworthy clients more accurately than previous methods. The company avoided the common tradeoff between growth speed and credit quality that plagued earlier fintech lending waves.
Inter's accelerating cohort performance reveals another AI advantage. Newer clients transact more because better targeting matches products to customer needs upfront. Manual underwriting relies on demographic proxies; machine learning identifies behavioral patterns that predict product usage.
The 2025 results contrast with 2021-2022, when Brazilian fintechs expanded rapidly but suffered rising non-performing loans. AI models trained on post-pandemic data captured changing payment behaviors that rule-based systems missed.
Client acquisition costs declined as AI improved approval rates for qualified applicants while rejecting riskier profiles. Fewer false negatives mean less marketing waste; fewer false positives reduce collection expenses.
Both companies now iterate credit models continuously. Nu launched 100+ features by testing algorithms on live portfolios and rolling out improvements monthly rather than in annual cycles. This cadence was impossible under manual underwriting governance.