NU Holdings has embedded NuFormer, a proprietary AI model trained on over a decade of transaction data across 100 million customers, into its core operating model across underwriting, customer service and growth targeting. The latest iteration has quadrupled context length, training speed and inference speed whilst reducing production costs—a critical efficiency gain for a fintech operating at scale. Generative AI agents now handle over 60% of customer support conversations in Brazil at or above human performance levels, whilst the model powers credit decisions across Brazil and Mexico with expansion into SME and Colombian markets underway. This deployment sits atop a scaled platform generating $5.9 billion in quarterly revenue with a 19.5% efficiency ratio, demonstrating that NU is layering AI investment onto proven unit economics rather than speculating on unproven technology.
The competitive pressure is immediate and multifaceted. Itau Unibanco, NU's primary banking rival, has partnered with Google to embed Gemini across customer service and business banking, whilst MercadoLibre completed an AI-powered search rollout across its five largest marketplaces and maintains 88 million monthly active users on Mercado Pago. For CX teams evaluating vendor strategy, the question becomes whether proprietary models trained on domain-specific data—as NU has built—will outperform general-purpose LLMs integrated by larger incumbents. NU's 60% automation rate in customer support suggests the former, yet Itau's rapid deployment of Google's infrastructure indicates that scale and partnership velocity matter as much as model sophistication. The risk for mid-market CX platforms is that financial services players are now building defensible moats through AI, potentially reducing their reliance on third-party CX infrastructure.
NU's financial performance underpins aggressive AI investment: net income hit $1.1 billion in Q2 2026 with a risk-adjusted net interest margin of 12.4%, up from 9.5% the previous quarter. This margin expansion—driven partly by lower credit costs through AI-driven underwriting—creates a self-reinforcing cycle where AI efficiency gains fund further AI development. The critical question for support leaders is whether this model is replicable outside fintech, where transaction history and credit risk data provide the training foundation NU leverages. For teams already embedded in Zendesk or Salesforce ecosystems, NU's approach signals that vertical-specific AI models may eventually outcompete horizontal platforms, particularly where domain data is proprietary and abundant. The competitive edge NU is building is not merely technological but structural—it stems from data advantage, not just model architecture.
NU Holdings AI Push: Can NuFormer Strengthen Its Competitive Edge? The Globe and Mail
NU Holdings AI Push: Can NuFormer Strengthen Its Competitive Edge? theglobeandmail.com