False Declines Remain a Costly Checkout Problem
False declines—legitimate transactions that card issuers mistakenly flag as fraud—continue to drain revenue from online retailers. Industry data cited by Stripe indicates that over half of US consumers have encountered a false decline in the past 90 days, and 43% of retailers view the problem as significant. The financial impact is substantial: false declines cost US online sellers an estimated $81 billion in lost sales in 2023, often outpacing losses from actual fraud.
Stripe’s Adaptive Acceptance product is designed to counter this by automatically identifying and retrying falsely declined payments in real time, without the shopper ever seeing the initial rejection. In 2024, the company reports the system recovered a record $6 billion in falsely declined transactions, a 60% year-over-year jump in retry success rate.
Model Shift Improves Retry Precision
The core of Adaptive Acceptance relies on AI to parse complex transaction patterns—combining bank policies, routing rules, messaging formats, and issuer preferences—to determine when a decline is likely a mistake. Stripe previously used a gradient-boosted tree model (XGBoost) for this task. The company has since moved to a TabTransformer-based deep neural network, referred to as TabTransformer+, which is better suited to modeling interactions among hundreds of factors that influence transaction outcomes.
A key architectural addition is high-dimensional embeddings that function as detailed maps of payment behavior. These embeddings allow the model to capture subtle signals that affect approval chances, enabling more nuanced decisions about which declines to retry and how to adjust those requests.
According to Stripe, the new model achieves 70% greater precision in identifying genuine false declines. This improvement drove the record revenue recovery while simultaneously cutting total retry attempts by 35%.
Faster Training Cycles Keep the Model Current
Alongside the model architecture upgrade, Stripe rebuilt its training pipeline. Training time dropped from days to hours, and the system now accommodates a larger dataset, giving the model deeper insight into varied transaction types and payment behaviors.
The accelerated pipeline enables Stripe to train and deploy new versions of Adaptive Acceptance multiple times per week. Faster iteration means the model can stay aligned with evolving false-decline patterns rather than relying on stale data.
Next Steps in AI Development
Stripe’s roadmap for 2025 includes pretraining and fine-tuning techniques for foundational models, paired with richer training data that incorporates additional features. The company suggests that continued AI progress, combined with transaction data from its network of millions of businesses, points toward a future where false-decline-related revenue loss becomes increasingly rare.
