The case for AI-driven payments optimization

Payments require hundreds of real-time microdecisions across a data-rich landscape, making the domain well-suited for AI-powered optimization. According to a recent Stripe survey, 43% of companies already use AI or machine learning tools to optimize payments, and another 32% plan to adopt them within two years.

Stripe's Payments Intelligence Suite, introduced at Stripe Sessions, applies AI across several product areas to automate these decisions. The suite includes products that have been available for years alongside new announcements such as Authorization Boost, Smart Disputes, and multiprocessor support for Radar.

Recovering declined transactions

False declines—legitimate transactions rejected due to suspected fraud—cost US online retailers an estimated $81 billion in lost sales annually. Authorization Boost addresses this by increasing authorization rates and reducing network costs through:

  • Network tokens available in more than 45 markets
  • Real-time card account updater across major card networks
  • Proprietary messaging, formatting, cost, and retry optimizations

The product includes excessive retry prevention, which blocks transactions likely to be rejected and reduces network costs. Improvements to the AI architecture powering retry logic helped Stripe recover more than $6 billion in legitimate declined transactions for users in 2024—a record amount. Businesses using Authorization Boost's real-time optimizations see an average acceptance rate increase of 3.8%, with some seeing up to 7%.

Strengthening fraud prevention

Radar, Stripe's fraud prevention tool, has been trained on Stripe data for over a decade. In the past year, dispute rates for Radar users fell by 17% even as industry-wide ecommerce fraud increased 15%. Recent upgrades to Radar include:

  • Intelligent 3DS authentication triggers for transactions that are risky but below the block threshold, backed by a new multihead model and decisioning layer. Early users saw over 30% fraud reduction on eligible transactions.
  • Risk scores and recommended actions for payments processed by other PSPs, enabling fraud protection across all transactions regardless of processor. DoorDash reported a 10% reduction in chargeback costs using these scores.
  • Screening and blocking for fraudulent ACH and SEPA transactions without reducing conversion for legitimate users—Radar users see an average 42% reduction in SEPA fraud and 20% reduction in ACH fraud.
  • Account fraud detection using models trained on data from more than 14,000 platforms, with custom account-level rules, suspicious-transaction interventions, and advanced analytics.

Disputes still cost businesses about $55 billion annually. Smart Disputes, a new AI-powered tool, automatically compiles and submits evidence for each dispute, tailored using Stripe's data. Vimeo and Squarespace recovered 13% more chargebacks using the tool, and Stripe covers the cost of fighting a dispute if the counter fails.

Gaining visibility into payments performance

Many businesses lack an end-to-end view of their payments conversion funnel, making it difficult to identify improvement opportunities. Payments analytics provides that visibility, including quantifiable recommendations such as potential revenue gains from adding a digital wallet at checkout. Disputes analytics offers insight into dispute rates by card brand and reason, along with evidence submission and dispute win rates.

Anomaly alerts use AI to catch authorization changes with over 90% precision, proactively notifying enrolled businesses before those changes impact revenue.

The underlying model architecture

Historically, Stripe built specialized AI models for individual tasks like fraud detection, authorization, and dispute management. The Payments Foundation Model takes a different approach: trained on tens of billions of transactions, it distills information about each payment into a single embedding that captures hundreds of subtle signals. These embeddings enable real-time predictions about each transaction.

The impact is visible in card testing defense. While attacks increased industry-wide, they are down 80% on Stripe. For novel attack types scattered across the volumes of large companies—which resisted traditional detection methods—the Payments Foundation Model raised detection rates for attacks on large users from 59% to 97% overnight by classifying sequences of embeddings.

Stripe plans to expand the Payments Foundation Model's use across its products for further performance and revenue improvements.