False declines—legitimate transactions mistakenly rejected due to suspected fraud—are a significant problem for online businesses. Recent studies show that more than half of US customers have experienced false declines, and 43% of retailers consider it a major issue. In 2023, false declines cost US online retailers an estimated $81 billion in lost sales. In many cases, revenue losses exceed those from actual fraud.

Stripe addresses this problem through Adaptive Acceptance, our AI-powered product designed to increase payment acceptance, in part by recovering false declines. We’ve recently made significant improvements to its AI architecture that allow Adaptive Acceptance to more accurately identify which transactions to retry to maximize your revenue. In 2024, this led to Adaptive Acceptance recovering a record-high $6 billion in falsely declined transactions, reflecting a 60% year-over-year increase in the retry success rate.  

AI upgrades to spot complex decline patterns

Adaptive Acceptance uses AI to optimize initial payment requests, and automatically identify and retry false declines in real time—without the end customer ever seeing the initial decline. The product’s effectiveness depends on its ability to recognize complex patterns in transaction data that indicate a legitimate payment was mistakenly rejected by issuers as suspected fraud. These patterns involve the intricate interplay of factors such as bank policies, routing, messaging, formatting, and issuer preferences.

Our previous approach, based on a gradient-boosted tree model (XGBoost), saved Stripe users billions of dollars in revenue. However, recent AI advances offered potential for improvement. We transitioned to a TabTransformer-based deep neural network, which we call TabTransformer+. This system excels at modeling complex interactions among hundreds of factors that influence transaction success.

A key enhancement in this new architecture is the addition of high-dimensional embeddings. These act as detailed maps of payment patterns, allowing our model to capture and analyze subtle signals affecting payment outcomes. This enhancement enables Adaptive Acceptance to make more nuanced decisions about which declined transactions to retry and how to adjust them for a higher approval chance.

Based on these improvements, Adaptive Acceptance’s new AI model achieves 70% greater precision in identifying legitimate transactions that have been falsely declined. This increased precision allowed us to recover more revenue than ever last year while reducing retry attempts by 35%.

Fast retraining, continuous improvement

To complement our model architecture upgrades, we also developed a more efficient training pipeline. This new system reduced our model training time from days to hours, and it allowed us to use a larger dataset—helping our model better understand the complexities of different transaction types and payment behaviors.

These advancements have significantly increased our iteration speed. We can now train and deploy new versions of our Adaptive Acceptance model multiple times per week. The faster cycle allows us to keep our models current with the latest trends in false declines, adapting quickly to new patterns as they emerge.

More AI advancements ahead

In 2025, we plan to enhance our model architecture with new techniques for pretraining and fine-tuning foundational models, while expanding our training data with additional features. The rate of progress in AI technology, combined with insights from the millions of businesses and billions of transactions on the Stripe network, make it possible to envision a near future in which sales lost due to false declines become increasingly rare—meaning fewer frustrated customers and more profit for you.

Learn more about Adaptive Acceptance.