--- title: "Affirm's Transformer Credit Model: Approvals Up, Losses Flat" slug: "affirm-transformer-underwriting-model-approvals-flat-credit-losses" author: "ibl.ai" date: "2026-11-07 12:00:00" category: "Premium" topics: "Affirm underwriting, transformer credit model, AI credit underwriting, BNPL, thin-file borrowers, credit risk model, adverse action, Regulation B, model risk management, financial services AI, proprietary data, model-agnostic AI" summary: "Affirm says the first version of its transformer-based underwriting model added 3.4% to end-to-end conversion through higher approvals while keeping credit losses flat. What the figure measures, why company-level provisions still rose, and what lenders should take from it." banner: "" thumbnail: "" linkedin: | Affirm says the first version of its transformer-based underwriting model added 3.4% to end-to-end conversion, through significantly higher approvals, while keeping credit losses flat. Two details the viral version skips. The 3.4% is conversion, not approval rate; Max Levchin put the approval gain at roughly 2 to 3 points. And "flat" describes the model's results, not the quarter: Affirm's provision for credit losses rose to $259 million from $163 million, which the CFO tied to deliberate credit experiments in September. The lesson still holds. The most valuable AI in lending is not a chatbot. It is a model trained on 14 years of the lender's own repayment data, reading credit history as a sequence. That advantage only compounds if the data, the models and the code stay yours. With ibl.ai, you own all the code and the data. #iblai #FinancialServices #CreditRisk #AI #EnterpriseAI --- ## The Short Answer **Affirm says the first version of its transformer-based underwriting model added 3.4% to end-to-end conversion through significantly higher approvals while keeping credit losses flat. The lesson for lenders is that the most valuable AI is a proprietary model trained on their own repayment data. On ibl.ai you own all the code and the data, so that advantage stays with the institution.** The figure comes from Affirm's [first fiscal quarter 2027 shareholder letter](https://www.sec.gov/Archives/edgar/data/0001820953/000162828026071005/affirmfq127shareholderle.htm), filed with the SEC alongside results on November 5, 2026. The model itself went live at U.S. checkout on September 17, 2026. Two qualifications matter. The 3.4% measures conversion, not approval rate. And "flat" describes the model's results, not the quarter: Affirm's company-wide provision for credit losses rose in the same quarter for a separate reason. ## What did Affirm announce about its transformer underwriting model? Affirm announced on September 17, 2026 that a transformer-based underwriting model was live at U.S. checkout, [as PYMNTS reported](https://www.pymnts.com/bnpl/2026/exclusive-affirm-rebuilds-underwriting-to-approve-borrowers-a-credit-score-cant-see/). Affirm says it is built on 14 years of transaction-level lending data. The model reads the order and timing of events in a consumer's credit history, not only summary measures such as balances, utilization and payment history. Its first effect was approving applications the existing system would have declined. Per that announcement, the newly approved group included consumers with thin files and no FICO scores. Those are the applicants a score-first system rejects most often, because a score has little to summarize. The shareholder letter describes the architecture as "directly inspired by, and quite similar to" the systems behind popular large language models. Before it, Affirm says, it selected model features entirely by hand. ## What does Affirm's end-to-end conversion figure actually measure? Affirm's exact wording is "an incremental 3.4% of end-to-end conversion via significantly higher credit approvals." Conversion counts completed purchases at checkout, so it captures approvals that a shopper then accepted, not the approval rate itself. On the [earnings call](https://www.fool.com/earnings/call-transcripts/2026/11/06/affirm-afrm-q1-2027-earnings-call-transcript/), CEO Max Levchin gave the approval number directly: "we're approving probably 2, 3 points more people." He added that this does not mean twice as many approvals.
Claim in circulation What Affirm's sources say
"+3.4% conversion" Confirmed: an incremental 3.4% of end-to-end conversion, from the first version of the model
"More approvals" Roughly 2 to 3 points more people approved, per Levchin on the call
"Credit losses flat" Model-level: "while keeping credit losses flat"
"Same risk for the whole company" Not what was reported: quarterly provision rose to $259M from $163M, tied to separate September credit experiments
"New" Live since September 17, 2026; the results were reported on November 5, 2026
## Did Affirm's credit losses really stay flat with the transformer model? Affirm's credit losses stayed flat for the model's newly approved population, per Affirm. Levchin described it on the call as lending "with no incremental loss whatsoever." The company-wide credit picture for the quarter moved, for a different reason. Provision for credit losses was $259 million against $163 million a year earlier, per the shareholder letter. On the call, CFO Rob O'Hare tied the increase to deliberate credit experiments "localized to the month of September." The shareholder letter also reports 30-plus-day delinquencies, excluding Peloton and Pay in X loans, at 2.9%: up 35 basis points on the quarter and 9 basis points on the year. None of that contradicts the model result. It does mean "same risk, more approvals" is a claim about one model's population, not about Affirm's balance sheet. Affirm has not published the population size, loss curves or long-run performance behind the 3.4%. ## Why would a transformer read credit history better than hand-built features? A transformer can read credit history better than hand-built features because it learns patterns from the sequence of events itself. A hand-built feature compresses that sequence into one number, such as utilization, and discards the order in which things happened. Affirm President Libor Michalek told [American Banker](https://www.americanbanker.com/payments/news/why-affirm-is-incorporating-llm-tech-into-its-underwriting) that the model looks at "credit and payment and repayment behavior as a sequence." The same article notes Parafin and Plaid also run transformer-based underwriting models. Levchin framed the gain as automation of feature discovery: the model "figures out how to learn these subtle patterns without any human involvement," work that previously needed "really smart humans." The same idea sits behind Revolut's banking foundation model, which we covered in [Revolut Built Its Own Foundation Model. Most Banks Can't.](/blog/banks-building-own-ai-foundation-models-revolut-pragma) Revolut reported a 2.3x uplift in identifying credit default risk on historical data. ## Can a transformer credit model meet adverse action explanation rules? A transformer credit model can be used in U.S. lending only if the lender can still state specific reasons when it declines someone. [Regulation B, section 1002.9](https://www.ecfr.gov/current/title-12/chapter-X/part-1002/section-1002.9), requires a statement of specific reasons for adverse action, whatever model produced the decision. The CFPB withdrew its 2022 circular on complex algorithms in May 2025, among 67 guidance documents, but the regulation itself did not change. Model complexity is not an exemption. Affirm says it addressed this directly: its letter states that "all decisions are deterministic, transparent, and explainable to consumers." PYMNTS reports a proprietary method that matches the explainability of its earlier machine learning models. That claim is Affirm's own, and the method is not public. For any other lender, the explanation layer, its validation and its monitoring are exactly what a model risk team will ask to inspect, as we set out in [SR 11-7 Compliant AI for Banks](/blog/sr-11-7-compliant-ai-for-banks). ## What should banks and lenders take from Affirm's underwriting model? Banks and lenders should take one structural point from Affirm: the 3.4% came from a model trained on 14 years of the lender's own transaction data. No vendor can sell that data, and no general-purpose chatbot carries it. The wrong conclusion is that every lender should train a transformer next quarter. On the call, Affirm President Michael Linford described "hundreds of millions of decisions every quarter" feeding its models, and a corpus that size is the precondition. The transferable conclusion is about ownership. A lender's repayment history, its features and its models are the asset. Anything that moves them onto a vendor's platform, or locks them to one model provider, gives away the part that compounds. The same pattern shows up in narrow finance tasks generally, where [fine-tunes on private data beat frontier models](/blog/finserv-fine-tunes-beat-frontier-models-not-junior-analysts) at a fraction of the inference cost. ## Where does an AI agent platform like ibl.ai fit around an underwriting model? ibl.ai does not build credit scoring models, and Affirm is not an ibl.ai customer. Where ibl.ai fits is the agent layer around a lender's own models and data: risk write-ups, credit memo drafting, policy lookup and analyst workflows that read from systems the institution controls. On ibl.ai, you own all the code and the data. The platform is model-agnostic, runs inside your own perimeter, and has no per-seat pricing, so a lender can point agents at its in-house models and swap language models without rewriting anything. For a regulated lender that matters for the same reason Affirm's model does. The data and the decisions stay where the model risk team can validate them, and the advantage stays with the institution that earned it. See how this works for [financial services](/solutions/financial-services). ibl.ai is family-owned and operated from New York, NY, and 1.6M+ users across 400+ organizations run the platform this way. ## Why does owning the AI stack matter? **ibl.ai is the agentic AI platform where you own all the code and the data. You self-host the entire stack inside your own perimeter, run it model-agnostic across any LLM and switch anytime, and pay by usage with no per-seat pricing — so you can deploy anywhere: your cloud, on-premise, GovCloud, or fully air-gapped.** - **You own all the code and the data.** Full source code under a perpetual license, running on your infrastructure. Not API access to someone else's platform — the stack itself is yours. - **Model-agnostic.** Run any LLM — Claude, GPT, Gemini, Llama, Command, or your own fine-tune — and switch providers without rewriting the platform. - **No per-seat pricing.** Usage-based billing against a budget cap you set. Cost tracks what your organization actually uses, not how many people you employ. - **Deploy anywhere.** Your cloud, your VPC, on-premise, GovCloud, or a fully air-gapped network with no outbound connectivity. 1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University. ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on.