How does Affirm use AI to make credit decisions?
Affirm underwrites every purchase individually, in real time, with models it builds in-house. Its fiscal 2026 10-K says the risk model turns user identifiers and transaction details into over 1,000 data points (over 500 a year earlier), models repayment probability month by month, and is trained on data from approximately 553 million loans. On September 17, 2026 Affirm added a transformer that reads individual credit-report accounts and payment histories and feeds an XGBoost model. Affirm's engineering blog says it delivered about 1.8 times the risk-ranking gain of the next XGBoost iteration, 2.1 times among thin-file applicants with no FICO score, and +1.2 percentage points (3.4% relative) conversion in an A/B test; to explain declines, Affirm developed a proprietary attribution algorithm. On credit, 30-plus-day delinquencies on monthly installment loans, excluding Pay in X and Peloton, were 2.5% in the June 2026 quarter against 2.3% a year earlier. Substantially all US loans are originated by Celtic Bank and Lead Bank. The rulebook moved in 2024 to 2026: the CFPB's May 2024 interpretive rule treating BNPL lenders as card providers was slated for revocation in March 2025, and FCA rules requiring affordability checks for UK BNPL took effect on July 15, 2026.[1][2][3][4][5][6][7][8]
Data as of Oct 19, 2026 · Affirm profile updated Oct 19, 2026 · 16 sources cited on this page · 52 fintechs profiled on this site · Markdown
What goes into Affirm's underwriting model, and what has changed?
From Affirm's 10-K filings, releases and the September 2026 engineering post on the transformer model.
| Item | What the record shows | Date |
|---|---|---|
| Inputs | Key user identifiers and transaction details become over 1,000 data points (FY2026); in FY2025 five top-of-mind user inputs became over 500 data points | 2025–2026[1][2] |
| Pricing logic | Month-by-month repayment probabilities combined with term length, purchase size, merchant and item to price and score risk, usually in seconds | August 27, 2026[1] |
| Data beyond scores | Transaction history, credit usage and Affirm's own repayment history; data from approximately 553 million loans to date (343 million a year earlier) | 2025–2026[1][2] |
| Bank-account data | Balances and cash-flow trends from linked accounts for Affirm Card and Affirm Money Account users | January 14, 2026[9] |
| Previous architecture | Hand-engineered features and tree-based models such as XGBoost, with further gains depending on new features | September 17, 2026[4] |
| Transformer stage | Tradelines and inquiries become tokens; the transformer is trained on repayment outcomes and its embeddings feed an accelerated failure time XGBoost model with credit, merchant and user features | September 17, 2026[4] |
| Offline results | Overall AUC improvement of 3.83% for the hybrid against 2.16% for the next XGBoost iteration (about 1.8x); about 2.1x among thin-file users | September 17, 2026[4] |
| Online results | +1.2 percentage points (3.4% relative) end-to-end conversion among eligible new users; added loans performed better than a comparable expansion under the previous models | September 17, 2026[4][3] |
| Rollout | Now running a first-look experiment with full approve-and-decline decisions for new users, and launching the next iteration to repeat and new users | September 17, 2026[4] |
How are the loans Affirm's models approve performing?
Affirm reports delinquency on monthly installment loans excluding Pay in X and Peloton; a loan is charged off at 120 days past due.
| Metric | Value | Period |
|---|---|---|
| 30+ day delinquencies | 2.8%, up 29 basis points year over year; allowance for credit losses 6.0% of loans held for investment | Q3 fiscal 2026 (March quarter)[10] |
| 30+ and 60+ day delinquencies | 30+ days 2.5% (2.3% a year earlier); 60+ days 1.5% (1.4%); 90+ days unchanged at 0.6% | Q4 fiscal 2026 (June quarter)[5] |
| Charge-off expectations | Recent monthly installment cohorts tracking toward roughly 3.5% ultimate net charge-offs; recent Pay in 4 vintages below 1% of GMV | August 27, 2026[5] |
| Rate caps | Celtic Bank and Lead Bank generally allow borrowers to agree to annual rates up to 36%; Cross River Bank up to 30% | August 27, 2026[1] |
| Credit reporting | All pay-over-time loans issued from April 1, 2025, including Pay in 4, reported to Experian | March 19, 2025[11] |
Which banks and regulators shape Affirm's credit models?
| Body | What | Date |
|---|---|---|
| Partner banks | Celtic Bank (Utah industrial bank) and Lead Bank (Missouri) originate substantially all US loans; underwriting is automated under their policies and those of Affirm's direct lending entity | August 27, 2026[1] |
| Nevada FID and FDIC | Applications for Affirm Bank, a Nevada industrial loan company | January 23, 2026[12] |
| CFPB | Interpretive rule that BNPL lenders are credit card providers with dispute and refund duties (May 22, 2024); the Bureau told a court in March 2025 it planned to revoke it | 2024–2025[6][7] |
| State attorneys general | A coalition of seven began coordinated inquiries into BNPL providers in December 2025, as the CFPB signaled reduced prioritization of BNPL enforcement, per Affirm's 10-K | August 27, 2026[1] |
| FCA | Unregulated BNPL comes under FCA rules from July 15, 2026, including proportionate affordability checks and the Consumer Duty; Affirm U.K. Limited is already authorised for consumer credit | February 11, 2026[8][1] |
| ECOA and Regulation B | Listed in the 10-K among laws that apply; the transformer post says underwriting models must give the principal reasons for a decline | 2026[1][4] |
What does Affirm not disclose about its credit AI?
Affirm reports the transformer model's gains against its own baselines and a control group; it has not published approval rates, loss rates for the loans the model added, or results by demographic group, and the offline AUC figures measure ranking, not losses.[4][3]
The explainability method is proprietary; Affirm says validation and ongoing monitoring check that its reasons stay accurate, but it has not described the method or any external review.[4][13]
What does the Affirm profile record on this topic?
The rows below are the entries of the Affirm AI strategy profile that bear on this page, as of Oct 19, 2026; the profile holds the full record of 13 dated moves, 8 use cases and 15 sources.
| Use case | What the record shows | Status | Source |
|---|---|---|---|
| Transformer-based real-time underwriting | Every purchase is underwritten individually in real time; the fiscal 2026 10-K says the risk model turns key user identifiers and transaction details into over 1,000 data points (over 500 in the fiscal 2025 10-K). The transformer model, live at US checkout since September 2026, added 3.4% completed purchases in initial deployment. | In production | [3][1][2] |
| Explainable decisions for thin-file applicants | The model approves some applicants with limited credit histories and no FICO scores; Affirm built a proprietary algorithm for explainability, with validation and ongoing monitoring of the explanations. | In production | [3] |
| Real-time bank-account signals | Underwriting uses account balances and cash-flow trends for Affirm Card users who link a bank account or hold an Affirm Money Account, with expansion to more consumers planned. | In production | [9] |
| AdaptAI personalized credit offers | AI-powered platform matching promotions such as 0% APR or special repayment terms to a consumer's purchase and profile, in use on Affirm's own products and offered to merchants since April 2025. | In production | [14][1] |
| Date | Move | Detail | Source |
|---|---|---|---|
| Sep 18, 2026 | Levchin: no staff cuts to fund AI | On CNBC's Closing Bell Overtime, Levchin discussed the new AI underwriting system and said Affirm has not had to cut staff to implement AI. | [15] |
| Sep 17, 2026 | Transformer underwriting model goes live | Affirm launched a transformer-based model that reads the order and timing of events in a credit history; in its first deployment it produced 3.4% more completed purchases than a control group, with a proprietary algorithm providing explainability at checkout speed. | [3] |
| May 12, 2026 | Investor Forum: transformer models outperforming | At its 2026 Investor Forum, President Libor Michalek said Affirm had processed $150 billion of volume and tracked 2.3 billion repayments over 14 years, and that its newer transformer-based underwriting architecture was already outperforming previous model generations. | [16] |
| Jan 23, 2026 | Affirm applies for a Nevada industrial bank charter | Affirm submitted applications to the Nevada Financial Institutions Division and the FDIC to establish Affirm Bank, an FDIC-insured industrial loan company, with John Marion selected as its president. | [12] |
| Jan 14, 2026 | Underwriting adds real-time balance and cash-flow signals | Affirm's underwriting began using account balances and cash-flow trends from linked bank accounts for Affirm Card and Affirm Money Account users; early results showed an average lift of about 12% when purchasing power rises. | [9] |
Does Affirm use AI to approve loans?
Yes. Affirm underwrites each purchase in real time with in-house machine-learning models trained on data from approximately 553 million loans, per its FY2026 10-K. Since September 17, 2026 a transformer model that reads individual credit-report records has fed its decisions for new US users.
Can Affirm approve people with no credit score?
Sometimes. Affirm says its transformer model approved some applicants its previous system would have declined, including people with limited credit histories and no FICO score, and its engineering blog reports the largest ranking gain, about 2.1 times the next XGBoost iteration's, among those thin-file users.
How does Affirm explain declined applications?
Affirm's engineering post says underwriting models must give the principal reasons an application was declined, so it developed a proprietary attribution algorithm for its transformer-plus-XGBoost model, with validation and ongoing monitoring. It has not published the method. The FY2026 10-K lists ECOA and Regulation B among the laws it follows.
What are Affirm's delinquency rates?
Per PYMNTS' report of the June 2026 quarter, 30-plus-day delinquencies on monthly installment loans excluding Pay in X and Peloton were 2.5%, against 2.3% a year earlier, 60-plus-day were 1.5% and 90-plus-day 0.6%. Recent monthly installment cohorts were tracking toward roughly 3.5% ultimate net charge-offs.
How is BNPL credit regulated in the US and UK?
In the US the CFPB's May 2024 interpretive rule treated BNPL lenders as credit card providers, but in March 2025 the Bureau said it planned to revoke it, and state attorneys general opened inquiries in December 2025. In the UK, FCA rules requiring affordability checks for BNPL took effect on July 15, 2026.
- Affirm Holdings Annual Report (Form 10-K) for the fiscal year ended June 30, 2026
- Affirm Holdings Annual Report (Form 10-K) for the fiscal year ended June 30, 2025
- Affirm launches transformer-based machine learning model for real-time underwriting
- Learning Richer Credit Representations with Transformers at Affirm
- Affirm Makes Smaller Purchases BNPL's Next Growth Test
- CFPB Takes Action to Ensure Consumers Can Dispute Charges and Obtain Refunds on Buy Now, Pay Later Loans
- CFPB retreats from classifying BNPL like credit cards
- New protections confirmed for Buy Now Pay Later borrowers
- Affirm updates underwriting with enhanced signals to better reflect consumers' real-time finances
- Affirm FQ3'26 Shareholder Letter (Exhibit 99.1 to Form 8-K)
- Affirm Expands Credit Reporting with Experian to Include All Pay-Over-Time Products
- Affirm submits applications to establish industrial loan company
- EXCLUSIVE: Affirm Rebuilds Underwriting to Approve Borrowers a Credit Score Can't See
- Affirm launches "AdaptAI" - its AI-powered promotions platform - to merchant partners
- Never had to cut staff to buy more AI tokens, says Affirm CEO as it rolls out AI underwriting
- Affirm Targets Card Networks With Broader Payments Push
When Affirm moves on AI, you'll read it here first.
when a bank, fintech or regulator moves on AI, it is in the next morning's brief · six sourced stories · 7 am PT · free
plus every regulation, bank and agent page update, the morning after · leave any morning