Stripe's Payments Foundation Model: How Data & Infra Create Compounding Advantage, w/ Emily Sands

Episode Summary

Executive Summary: In this episode, Emily Sands, Head of Data and AI at Stripe, discusses Stripe's innovative Payments Foundation Model, a transformer-based model that processes tens of billions of transactions to create dense embeddings for fraud detection and payment optimization. The model, which treats payments as a distinct modality, has dramatically improved card testing detection from 59% to 97%. Sands also covers Stripe's modular AI deployment strategy, the use of LLMs as judges for labeling, adaptive risk thresholds, and the broader implications of AI for platform incumbency and agentic commerce.

Main Topics: Payments Foundation Model (Priority: 5/5): A transformer model that converts each payment into a compact vector (embedding) by analyzing sequences of transactions across multiple entities (buyer, card, device, merchant). It uses a BERT-style masked modeling setup with similarity fine-tuning, enabling superhuman understanding of payment patterns. Fraud Detection and Prevention (Priority: 5/5): Stripe's layered approach to fraud includes the foundation model, adaptive 3DS authentication, dynamic risk thresholds, and blending rules with ML models. The system has reduced dispute rates by 17% year-over-year despite industry-wide fraud increases. Modular AI Deployment (Priority: 4/5): Stripe exposes foundation model embeddings as features for existing ML systems, allowing rapid iteration (weekend projects vs. quarter-long projects). This modular approach is also applied to merchant intelligence and other horizontal services. LLMs as Judges and Labeling (Priority: 4/5): Using LLMs to generate and validate labels for suspicious payments where no ground truth exists, particularly for friendly fraud and free trial abuse. This helps tighten the iteration loop against fraudsters. Agentic Commerce and AI Infrastructure (Priority: 3/5): Stripe is focusing on being the economic infrastructure for AI, enabling agentic commerce (e.g., agents buying on behalf of users), embedding payments in developer tools, and supporting AI companies with billing, tax, and fraud solutions. Data Advantage and Platform Incumbency (Priority: 4/5): Stripe's scale ($1.4 trillion processed) creates a compounding data flywheel: more data leads to better models, which attract more customers, further strengthening their competitive moat. This raises questions about market competition in the AI era. Talk-to-Your-Data and Sigma System (Priority: 3/5): Stripe's natural language interface for analytics leverages well-structured data and provides explanations of SQL queries to build user trust. The system also offers smart benchmarking against peer groups.

Key Arguments: Payments foundation models are a distinct modality that can achieve superhuman performance in understanding transactions, especially when modeling sequences across multiple entities. Modular AI deployment (exposing embeddings as features) allows rapid iteration and incremental value without requiring a complete system overhaul. Blending rules with ML models (e.g., adaptive 3DS, dynamic risk thresholds) outperforms pure ML or pure rule-based approaches in fraud detection. LLMs as judges can generate reliable labels for ambiguous cases (e.g., friendly fraud) where no ground truth exists, accelerating the iteration loop. AI strongly favors incumbent platforms with large, proprietary datasets, creating a compounding advantage that may make competition difficult. Agentic commerce is still early but gaining traction, with examples like HipCamp and embedded payments in developer tools (e.g., Cursor).

Data Points: Total payments processed by Stripe in 2024: $1.4 trillion - Roughly 1.3% of global GDP Card testing detection rate improvement: From 59% to 97% - After deploying the Payments Foundation Model Dispute rate reduction for Stripe businesses: Down 17% year-over-year - Despite industry-wide e-commerce fraud being up 15% Stripe's transaction volume growth: 38% year-over-year - Driving the data flywheel Percentage of cards seen first by a merchant that Stripe has seen before: 92% - Demonstrates the density of Stripe's network Percentage of AI companies using Stripe: Two-thirds of the Forbes AI 50 - Stripe's focus on being the economic infrastructure for AI Lovable's ARR in first eight months: $100 million - Case study of a company going all-in on Stripe Failed charge recovery via smart retries: 60% - For recurring bills

Pivotal Quotes: "The more data we process, the better our models get. The better our models get, the more value we deliver to businesses. The more value we deliver to businesses, the more the businesses grow, which means the more transactions they run through Stripe." — Emily Sands: Describing Stripe's compounding data advantage and competitive moat "Our detection rate of card testing at large merchants went from 59%, which is not bad, but not great, to 97% from that change." — Emily Sands: Highlighting the dramatic impact of the Payments Foundation Model on fraud detection "Once you have a shared embedding, then spinning up a new model becomes a weekend project, not a quarter project." — Emily Sands: Explaining the efficiency gains from modular AI deployment

Implications: Stripe's Payments Foundation Model demonstrates how domain-specific AI can create superhuman capabilities and strong competitive moats. The modular deployment strategy offers a blueprint for other companies to leverage AI incrementally. The rise of agentic commerce and AI-native infrastructure will reshape how businesses operate, favoring platforms with data advantages. Policymakers must consider the implications of AI-driven market concentration.