Affirm's AI approves shoppers FICO can't see, yet no one can explains its rules yet.

On 17 September, Affirm switched on a new brain at U.S. checkout: a transformer model, built on 14 years of transaction-level data↗, that reads the order and timing of events in a credit history instead of the summary a three-digit score captures ¹⁴. Its first job is approving shoppers the old system would have declined, thin files and no-FICO↗ borrowers included ¹.

A wider funnel, at what the company calls comparable risk

Affirm says the model produced 3.4% more completed purchases than a control group, at comparable risk, with the extra loans performing better than a similar expansion under earlier models ¹⁴. The result is company-reported and unaudited, and the loss metrics and borrower mix behind it are unpublished ².

Investors priced it anyway. The stock rose 5.5% premarket on announcement day↗ ²⁹, and Mizuho analyst Dan Dolev has floated over $10 in GAAP earnings per share by 2030↗, roughly double consensus and above management's own $3-to-$4 target, in a scenario that puts the stock above $200 ⁷.

The rulebook is a patchwork

The approval layer went live before any public rulebook could explain it. The CFPB said on 6 May 2025 it would not prioritize enforcement↗ of its 2024 rule treating BNPL lenders like card issuers, and is contemplating rescinding it ⁵. Affirm furnishes repayment data to all three national bureaus; Klarna and Afterpay do not report routine on-time payments ³. Same payment record, two different credit futures, decided by which checkout button you tapped.

Senators Warren, Blumenthal, Duckworth and Hirono demanded bureau answers by 18 May; none has been released publicly ³. FICO says a yearlong study with Affirm found its new model moved most scores fewer than 10 points↗, but it covered only the one lender furnishing data voluntarily ³. The same firm, whose 1989 score is used by 90% of top U.S. lenders ⁸, ships BNPL-inclusive models to lenders this fall ³.

The losses would land at home

The machine is being asked to widen a funnel that already runs hot. Revenue has grown at least 29% year over year for ten straight quarters ⁶, and fiscal fourth-quarter GMV rose 36% to $14.1 billion on 53 million transactions, up 41%, at an average order value that fell 4% ¹².

Levchin's own guardrail for the machine: "a human decides what's worth looking at, what's not worth looking at" ⁴. If the risk math breaks, losses land on Affirm's own book while its Nevada industrial-bank application, filed 23 January, waits on the FDIC ², and card delinquencies sit at a 15-year high ³.

The borrower-side loss lands either way. About 15% of U.S. adults used BNPL in the past year, Federal Reserve data shows, and usage skews toward borrowers with limited credit access ³. A Klarna shopper's flawless pay-in-four record never reaches the bureaus, so it cannot help at a mortgage or card desk; the first explanation of the new rules arrives as a denial letter or a higher rate ³.

If Affirm's math holds instead, it compounds on borrowers a 1989 score marks invisible ². The first full credit cycle, the bureaus' answers and FICO's fall rollout will say which ³.