Affirm Credits an LLM-Style Model for a 3.4% Conversion Lift

Affirm told shareholders on 5 November that a new underwriting model built on a transformer architecture, which it describes as “directly inspired by, and is quite similar to, the ones behind some of today’s most popular LLMs”, “delivered an incremental 3.4% of end-to-end conversion via significantly higher credit approvals, while keeping credit losses flat”, according to its fiscal first-quarter shareholder letter, filed with the SEC. The quarter to 30 September brought revenue of $1,262.2 million, up 35%, on gross merchandise volume (GMV) of $14.4 billion, up 34%. Over the same quarter, the provision for credit losses on the income statement rose 59%, to $259.4 million from $162.8 million (our arithmetic).

The letter says Affirm used to choose its credit model’s inputs by hand: “For years, we carefully selected model features (data that helps predict credit outcomes, such as repayment consistency) entirely by hand.” The new system, it says, combines that decade of manual work “with an entirely new approach for automatically learning subtle patterns in relevant data.”

Business Pill · Financial Conditions

An evergreen Business Engineer explainer on how the cost and availability of credit change. It teaches the general idea only and does not discuss Affirm.

What the Letter Says About AI

• Underwriting. “The first version of the transformer-based model delivered an incremental 3.4% of end-to-end conversion via significantly higher credit approvals, while keeping credit losses flat.” The letter adds that “all decisions are deterministic, transparent, and explainable to consumers.”
• An LLM gateway. “A double-digit (and growing) share of our AI inference traffic now runs through an Affirm-developed gateway, allowing us to route between leading frontier and open-weight models behind a single interface, improving speed of adoption and token cost management.”
• Agentic commerce. “As we integrate Affirm into agentic commerce experiences, we can already see that agents will indeed free shoppers from the drudgery of filling out forms and clicking buttons.” The letter does not give agentic volumes.
• The Card risk model. Affirm says it rolled out “the latest iteration of our Card-specific risk model”, which it says unlocks “more Card GMV without sacrificing unit economics.”

The Quarter in Numbers

• Total revenue $1,262.2 million, up 35%; revenue was 8.79% of GMV, above Affirm’s 7.5% to 8.5% medium-term range.
• GMV $14.4 billion, up 34%, its twelfth straight quarter above 30% growth. Transactions rose to 55.6 million from 41.4 million.
• Revenue less transaction costs (RLTC, a non-GAAP measure) $617.1 million, up 36%, or 4.30% of GMV.
• Operating income $187.1 million, against $63.7 million a year earlier; net income $164.2 million, against $80.7 million; diluted earnings per share $0.47, against $0.23.
• Active consumers 28.7 million, up 19%. Active cardholders rose 112% to 6 million and Card GMV grew 117%; the letter says Card accounted for 23% of GMV.

Where the Credit Cost Shows

The letter says RLTC as a share of GMV rose 7 basis points, with gains in revenue, funding and servicing “mostly offset by a 29 basis point increase in provision for loan losses.” The allowance for credit losses rose to 6.3% of loans held for investment, from 5.9% a year earlier and at 30 June. Thirty-day-plus delinquencies on monthly installment loans, excluding Peloton and Pay in X loans, rose 9 basis points from a year earlier and 35 basis points from the prior quarter, to 2.9%, the letter says.

Affirm says recent monthly installment cohorts “are tracking towards approximately 3.5% ultimate net charge-offs as a percent of cohort GMV, which is in-line with expectations and consistent with the performance of historic loan cohorts.” Its average funding cost fell about 69 basis points from a year earlier to 6.0%, but rose 27 basis points from the prior quarter, which the letter attributes to higher benchmark rates; fixed-cost funding was more than 70% of on-balance-sheet funding.

The Outlook

• Fiscal second quarter: GMV of $17.9 billion to $18.2 billion and revenue of $1,435 million to $1,465 million.
• Fiscal 2027: GMV of more than $65.25 billion, with revenue and RLTC similar to fiscal 2026 as a share of GMV, and an operating margin of more than 14.75%.
• The Wall St Engine card that flagged the results put the analyst consensus at $1.21 billion of revenue and $0.38 of earnings per share; those estimates are the card’s, not Affirm’s.

Connecting the Coverage

Affirm says it is integrating into agentic checkout. Visa’s own tests found AI shopping agents failed 42.8% of runs and often still bought, in Visa: AI Shopping Agents Fail 42.8% of Runs, Often Still Buy; start-ups building those agents raised $30 million, in Agent-Shopping Firms Raise $30M as Uber Logs 4 Agent Rides; and the largest platforms are keeping agent sales in-house, in AI Assistants Sell In-House: Amazon’s 400M vs Uber’s 4 Rides.

The Structural Read

As we read the letter, Affirm is positioning AI as an underwriting engine rather than a product feature: the model sits where its revenue is decided, at the approval.

As we read the income statement, the two numbers measure different things. The 3.4% is Affirm’s estimate of the model’s effect; the 59% provision growth covers the whole loan book, and the letter reports provision up 29 basis points as a share of GMV, which mostly offset gains elsewhere in RLTC.

Three Implications

FOR LENDERS Affirm says the model lifted approvals without lifting losses; its provision line will show whether that holds as the book grows.

FOR AGENTIC COMMERCE Affirm says it is integrating into agentic checkout; the letter gives no agentic volumes yet.

FOR AI VENDORS A double-digit share of Affirm’s inference traffic now routes through its own gateway across frontier and open-weight models, the letter says.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Four Intelligence Moats.

The framework says: “The transformer is a commodity now; what changes between paradigms is where intelligence accumulates and who can capture it.”

As we read it, Affirm’s letter makes the same distinction: the architecture is borrowed from LLMs, and the edge it claims comes from its own repayment data. The framework does not discuss Affirm.

What Is Not Established

The 3.4% conversion lift and the “flat” credit losses are Affirm’s own figures for its model; the letter does not say against what baseline, over what period or on what share of volume they were measured, and they are not independently verified. The provision figure on the income statement covers the whole loan book and is not a measure of the new model. The letter does not say how much of the quarter’s approvals the transformer model handled. The quarterly 10-Q’s risk factor section refers readers to the risks in the annual report for the year to 30 June 2026.

The Bottom Line

Affirm says a model it likens to LLMs approved more borrowers without raising losses; the same quarter’s accounts show credit provisions growing faster than revenue.

95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. Figures and quotations are from Affirm’s fiscal first-quarter 2027 shareholder letter, filed as Exhibit 99.1 to its Form 8-K on 5 November 2026, and its Form 10-Q for the quarter to 30 September 2026. Consensus estimates are as cited by Wall St Engine on X. Growth rates in the chart are our arithmetic from the income statement.