Affirm's transformer model lifted approvals 3.4% with no added credit losses

Affirm's first transformer-based underwriting model produced a 3.4% gain in end-to-end conversion by approving a larger share of applicants, all without pushing credit losses higher. That combination matters because approval-rate gains typically come at a cost: lenders who chase volume usually accept worse credit risk, and that trade-off is what has historically separated durable BNPL businesses from fragile ones.

CEO Max Levchin framed the result as lending profitably to a greater pool of responsible borrowers, with direct benefits flowing to merchants (more completed sales) and to Affirm's capital partners (more high-quality assets on their books). The claim, if it holds through subsequent quarters, would mean the model is finding creditworthy borrowers that the prior system was incorrectly turning away.

Transformer architectures are the same class of machine-learning model that underpins large language models, applied here to credit decisioning. Affirm explicitly described this as the first version, which signals the company views this as an iterable system. Early iterations in model development tend to capture the most obvious gains, so the trajectory of version two and beyond will be worth tracking closely.

Despite the encouraging underwriting data, the stock slipped 3% in the period since the earnings call. Investors need to resolve that gap: either the market doubts the durability of the flat-loss-rate claim, or other factors from the same earnings report outweighed it.