The Pulse
Affirm Deploys Transformer Model for Real-Time Credit Decisions
Affirm has deployed a transformer-based underwriting model at U.S. checkout. The company says the system produced 3.4% more completed purchases while maintaining comparable credit risk.
AI.info Team ·
“What’s exciting about the transformer model architecture is that we can now find new information within the data we already have.”
Libor Michalek, president, Affirm
Affirm has put a transformer-based machine learning model into production for real-time credit decisions at checkout in the United States. The buy-now-pay-later company announced the deployment on September 17, saying the system approved additional eligible applications that its previous underwriting models would have rejected.
In an initial deployment measured against a control group, Affirm said the model generated 3.4% more completed purchases. The additional loans performed better than a comparable expansion made with the company’s previous machine learning models, according to Affirm. The release did not provide the underlying approval, delinquency or loan-volume figures behind the comparison.
Affirm’s announcement describes the model as the latest stage in 14 years of transaction-level underwriting. The company evaluates each purchase individually rather than extending a single revolving credit line, using information about the consumer, the transaction and repayment risk at the time of purchase.
Affirm Reads Credit Histories as Sequences
Traditional credit data often arrives as a set of summaries: account balances, credit utilization, account counts and payment history. Affirm says its new model retains those inputs but examines the order and timing of events across a consumer’s credit history.
That allows the transformer to identify patterns within individual credit accounts and across multiple accounts as they change over time. The company says the system can find those signals without requiring engineers to design a separate measure for every possible pattern in advance. Affirm has specifically used the model to expand access for people with limited credit histories and consumers without FICO scores.
Michalek said the company has increased the amount of data used to train each generation of its underwriting models. “Seeing a credit history more clearly means we can responsibly say yes to more people,” he said.
The Checkout Constraint
Deploying a more complex model created a practical problem: a credit decision still has to arrive quickly enough for a customer completing a purchase. Affirm said it built a proprietary algorithm around the transformer that preserves the same level of explainability as its traditional machine learning models while keeping the system fast enough for real-time use.
The company also said it uses validation and ongoing monitoring to check whether the model’s explanations remain accurate and reliable. Affirm did not disclose the model’s latency, architecture size, training compute or the specific explanation method used in production.
Explainability carries particular weight in lending because an automated approval or denial must be tied to reasons that can be reviewed. Affirm’s announcement makes a product claim about its internal system rather than presenting an independent audit or research paper, so the company’s reported performance figures remain the main public evidence available for the launch.
More Approvals, Not Every Approval
Affirm framed the rollout as a credit-risk decision rather than an approval-maximization exercise. “The goal isn’t to approve every transaction, it’s to make the right decision for each one,” Michalek said. He added that the company does not benefit from extending credit that consumers cannot repay.
That distinction reflects Affirm’s business model. The company says it underwrites every purchase in real time and offers payment plans based on what a consumer can responsibly repay, with no late or hidden fees. Its earlier underwriting materials describe the use of external credit-reporting data, internal repayment history and transaction details such as the requested purchase amount.
The transformer deployment extends that approach by treating credit history as a changing sequence rather than a static collection of account metrics. Affirm’s reported 3.4% increase in completed purchases shows the immediate commercial result it is seeking: more transactions at what the company describes as comparable levels of risk. The release does not disclose whether the model will be expanded beyond U.S. checkout or how its performance varies across consumer groups.