By PYMNTS
September 17, 2026
Main Facts
Fintech giant Affirm has officially launched a next-generation, real-time underwriting model at digital checkouts across the United States. Driven by cutting-edge transformer neural network architecture—the same foundational technology driving modern generative artificial intelligence—this proprietary system is designed to evaluate consumer creditworthiness with unprecedented precision.
Rather than relying on traditional, static credit scores or heavily compressed summary metrics, Affirm’s new underwriting engine evaluates the chronological sequence, cadence, and patterns of events within a consumer’s credit history. By extracting deeper signals from existing data sources, the model successfully approves qualified applicants who would typically be rejected by legacy underwriting systems, including individuals with "thin files" and those entirely lacking a FICO score.
The deployment of this model comes at a pivotal operational moment for Affirm. The company recently reported a 41% year-over-year surge in fiscal fourth-quarter transactions, reaching 53 million completed purchases, even as the average order value (AOV) declined by 4%. This metric shift signals that consumers are increasingly turning to Affirm for smaller, more frequent everyday purchases rather than exclusively reserving buy now, pay later (BNPL) options for large-ticket items. By deploying a transformer model capable of safely greenlighting thin-file and un-scored borrowers, Affirm is positioning itself to capture an even larger share of daily digital commerce.
Chronology
The rollout of this transformer-based underwriting architecture represents the culmination of a multi-year technological evolution at Affirm, rooted in its foundational operational philosophy.
- Foundation (14 Years Prior): Since its inception, Affirm has built its brand on individual, transaction-level underwriting. Eschewing the blanket revolving credit lines common in traditional credit cards, the company engineered a system where every single purchase is evaluated independently in real-time, utilizing proprietary in-house machine learning models. The core promise has remained consistent: no late fees, no hidden costs, and custom decisioning for every transaction based on real-time affordability.
- Early 2026: Affirm executives, including CEO Max Levchin, began publicly emphasizing the company’s hyper-granular approach to risk assessment. In a widely discussed LinkedIn post earlier in the year, Levchin highlighted the company’s commitment to underwriting "every transaction in addition to every single person, every single time."
- Summer 2026: Following a robust fiscal fourth-quarter performance—which saw active consumers climb 21% to 27.8 million and gross merchandise volume (GMV) rise 36% to $14.1 billion—internal validations of the transformer model reached completion. Affirm finalized proprietary algorithms designed to ensure the model maintained regulatory explainability while retaining the processing speeds necessary for lightning-fast digital checkouts.
- September 17, 2026: Affirm officially announced the live deployment of the transformer-based underwriting model across U.S. merchant checkouts. Initial pilot metrics revealed a 3.4% boost in completed purchases relative to a control group, alongside superior loan performance compared to previous machine learning expansions.
Supporting Data
The quantitative impact of Affirm’s new underwriting architecture challenges long-held industry assumptions regarding the trade-off between credit expansion and default risk. Historically, lenders attempting to capture unserved or underserved segments—such as consumers with thin credit files—had to accept a measurable degradation in portfolio performance or risk management efficacy. Affirm’s early deployment data points to a different outcome.
- 3.4% Lift in Completed Purchases: When deployed against a standard control group, the transformer model enabled Affirm to securely approve eligible applicants who would have otherwise been declined by legacy systems. This marginal approval lift directly translated into a 3.4% increase in successfully completed checkouts for merchant partners.
- Superior Portfolio Performance: Crucially, the newly approved loans did not carry a higher default penalty. In fact, these loans demonstrated better overall performance than comparable credit expansions achieved under Affirm’s prior generations of machine learning models.
- Volume and Transaction Shifts: The launch coincides with significant volume changes recorded in Affirm’s fiscal fourth-quarter financial disclosures:
- Total Transactions: Rose 41% year-over-year to 53 million.
- Gross Merchandise Volume (GMV): Climbed 36% to reach $14.1 billion.
- Average Order Value (AOV): Fell 4%, indicating a structural migration toward lower-ticket, high-frequency consumer purchases.
- Active Consumer Base: Expanded by 21% to hit 27.8 million users.
- Engagement: Transactions per active consumer grew 20% over a trailing 12-month window to reach an average of 7.0 transactions per user.
Official Responses
Executives at Affirm have underscored that the new model is not an exercise in loosening credit standards, but rather an optimization of analytical clarity.
Libor Michalek, President of Affirm, emphasized that the fundamental mission of the company’s data science team is to extract deeper insight from existing variables rather than simply casting a wider, unrefined net.
"Seeing a credit history more clearly means we can responsibly say yes to more people," Michalek stated in the official announcement. "We’ve steadily accelerated the amount of data we use to train each generation of our underwriting models. What’s exciting about the transformer model architecture is that we can now find new information within the data we already have."
Addressing concerns regarding portfolio risk, Michalek reiterated that volume growth must never compromise fiscal discipline:
"Underwriting is the heart of what we do. The goal isn’t to approve every transaction, it’s to make the right decision for each one. We don’t benefit from extending credit that can’t be repaid, which means saying yes to more people only works when we get even better at saying no."
Chief Executive Officer Max Levchin has similarly championed the granular nature of the firm’s data strategy. By forcing machine learning algorithms to evaluate the complex temporal relationships within a borrower’s financial history—such as the exact timing between account openings, payment fluctuations, and credit utilization spikes—Affirm avoids the blunt-instrument limitations of traditional three-digit FICO scores.
Implications
The integration of transformer models into consumer credit underwriting carries profound implications for the broader fintech, banking, and retail ecosystems.
1. The Obsolescence of the Traditional FICO Score for Thin-File Borrowers
For decades, the FICO score has served as the gatekeeper of American consumer credit. While effective for individuals with long, established borrowing histories, the system systematically penalizes or shuts out younger demographics, gig economy workers with non-traditional income cadences, and immigrants who have not spent decades building institutional credit. By demonstrating that sequence-based, transformer-driven models can safely evaluate un-scored individuals, Affirm is providing a working blueprint for alternative credit scoring that could eventually influence auto lending, mortgages, and traditional banking.
2. Reinventing Buy Now, Pay Later as Everyday Commerce
The historical use case for BNPL was centered on big-ticket discretionary purchases—electronics, luxury goods, and travel. However, as Affirm’s data shows, average order values are dropping while transaction frequencies climb. Consumers are utilizing installment options for groceries, apparel, and day-to-day household goods. A real-time underwriting engine capable of instantly assessing micro-risk allows BNPL providers to seamlessly embed themselves into the fabric of daily retail, competing directly with debit cards and traditional revolving credit cards.
3. Regulatory Compliance and Model Explainability
One of the primary historical hurdles of applying deep learning and transformer architectures to financial services has been the "black box" problem—the inability to clearly explain why an algorithm made a specific lending decision to regulators and consumers. Affirm’s engineering team addressed this by developing a proprietary ancillary algorithm that translates the transformer’s multi-layered findings into clear, auditable explanations. This technological breakthrough ensures compliance with fair lending laws while preserving checkout speed, setting a new standard for explainable AI in financial services.
4. Competitive Pressures on Traditional Issuers
Traditional credit card issuers rely heavily on high interest rates and revolving debt traps (such as late fees and compounding interest) to monetize riskier or lower-income consumer segments. Affirm’s transparent model—operating on upfront terms with zero late fees—proves that responsible, AI-optimized underwriting can safely capture these segments without predatory pricing structures. As fintech competitors race to match Affirm’s technological leap, legacy financial institutions will face mounting pressure to modernize their own antiquated underwriting infrastructures.
