Affirm has unveiled a transformer-based machine learning model for real-time, transaction-level underwriting in buy-now-pay-later (BNPL) lending. Showcased at the company’s Investor Forum, the model represents roughly 18 months of development and outperformed Affirm’s existing mainline underwriting system in experimental testing, creating what the company views as a new baseline for credit decisioning. Transformers, the architecture behind models such as ChatGPT and Gemini, use an attention mechanism to weigh relationships among data points. Affirm says the approach can assess borrower risk and select terms using current, transaction-specific signals rather than relying mainly on an older credit score. As of January 2026, its underwriting engine already incorporated real-time information from linked bank accounts, including current balances, which lifted approval rates while keeping delinquency under control. Founded in 2012, Affirm has built a proprietary dataset covering more than 50 million underwritten consumers and over $100 billion in originated loans. The company keeps its machine learning and credit engineering teams in-house, and says more accurate models can support higher approvals without increasing delinquency. For investors in Affirm (NASDAQ: AFRM), the rollout underscores the importance of underwriting infrastructure to lending unit economics, margins and defaults.