How Is Stripe Using AI to Power Payments and Fraud?

TL;DR

Stripe launched what it calls the first AI foundation model for payments, a proprietary system trained in-house on all of Stripe's historical payment data to handle fraud prevention and conversion optimization. The company also expanded stablecoin money management to 100 new countries and released a toolkit for agentic buying, letting purchases happen inside AI tools like ChatGPT.

Transcript

Stripe this week holding its annual product event here in San Francisco. Stripe Sessions, the payment giant out with some AI news among other things. Here to talk about that and a lot more is John Collison himself, president and co-founder of Stripe. John, great to see you. I'm glad to be here. Yeah, it's really great to have you here in person. We... Read More

Key Insights

• Stripe's payments foundation model is trained in-house on all of the company's historical payment data, replacing single-purpose models to make inferences for fraud prevention and conversion optimization across merchants.
• The foundation model is proprietary and built from scratch, using open-source AI infrastructure but not importing models from OpenAI or Anthropic, according to co-founder John Collison.
• Card testing fraud, described as a particularly pernicious type, has dropped 80% over the past two years on Stripe through its earlier single-purpose AI fraud models.
• Fraud prevention is a constantly evolving adversarial game, and new technology like AI creates new venues for fraud such as deep fakes, much as automobiles once aided both police and bank robbers.
• Stablecoin money management, including balances and payouts, is now available in 100 new countries from Argentina to the Vatican, and Stripe announced a stablecoin card backed by Visa.
• The total dollar value in the stablecoin system grew 40% over the past year, with real-world businesses like X, Scale AI, and remote.com using stablecoins for global payouts rather than speculation.
• Stripe acquired Bridge, a stablecoin orchestration platform, late last year, and now frames the technology as digital dollars rather than using crypto jargon like NFTs or minting.
• AI companies are among the fastest-growing segments on Stripe, exemplified by Cursor reaching a $9 billion valuation and growing to $300 million in ARR within two years of launch.

Questions & Answers

Q: What is Stripe's AI foundation model for payments?

Stripe launched what it calls the first payments foundation model, trained on all of Stripe's historical payment data. Just as large language models have a deep understanding of language, this model has a deep understanding of payment data. It replaces earlier single-purpose models and can be used to make inferences on behalf of merchants for tasks like fraud prevention and improving conversion, bringing multiple capabilities together into one system.

Q: Is Stripe's payments foundation model built on OpenAI or Anthropic models?

No. According to co-founder John Collison, the model is proprietary and built in house. While Stripe uses lots of the open-source AI infrastructure that is available, the model itself is Stripe's own, trained from scratch on all of the payment data the company has. It is not importing anything from OpenAI, Anthropic, or other external model providers.

Q: How much has AI reduced fraud on Stripe?

Stripe has seen an 80% reduction in card testing, which Collison describes as one particularly pernicious kind of fraud, over the past two years. This came from Stripe's earlier approach of using specific single-purpose AI models, such as a dedicated fraud prevention model. Fraud prevention remains a constantly evolving adversarial game, especially as AI enables new kinds of fraud like deep fakes.

Q: Why does Stripe compare AI fraud to bank robbers and automobiles?

Collison uses the analogy that new technology tends to help both good and bad actors. The automobile was very useful for police but also very useful for bank robbers, and it took a while to figure out the bank robbing situation. Similarly, the internet created a new venue for fraud, and AI now enables new kinds of fraud such as deep fakes, making fraud prevention a constantly evolving adversarial game.

Q: What stablecoin features did Stripe announce?

Stripe added stablecoin-powered money management functionality across 100 new countries, from Argentina to the Vatican, letting people keep balances and do payouts. It also announced a stablecoin card backed by Visa. Collison noted the total dollar value in the stablecoin system grew 40% over the past year, and that Stripe acquired Bridge, a stablecoin orchestration platform, late last year.

Q: Who is actually using stablecoins on Stripe?

Collison said the usage is not retail or speculative but real-world businesses. Companies like X, Scale AI, and remote.com use stablecoins for paying out users around the world. Companies like Chipper Cash and the Dolar app allow users in Latin America or Africa to manage balances. Stripe describes this functionality as digital dollars rather than using crypto jargon like NFTs or minting.

Q: How is the AI boom affecting Stripe's business growth?

AI companies are one of the top segments among the fastest-growing companies on Stripe. Collison cited Cursor, which announced a $9 billion valuation and grew from launch to $300 million in ARR in just two years. He said this takeoff is much faster than previous trends like SaaS. Beyond AI-native firms, existing companies are also using AI tools to power their own growth.

Q: What is agentic buying and how does it work?

Agentic buying moves commerce out of the web browser and into AI tools. Currently payments happen when someone clicks a link, fills out details in a browser, and hits buy. Stripe launched a toolkit for agentic buying, which it believes is the first, enabling purchases inside AI tools like ChatGPT. This relies on MCP, a standard letting AI tools interact with and manipulate the real world, such as making a purchase.

Summary & Key Takeaways

• At Stripe Sessions in San Francisco, which drew 7,500 attendees, Stripe unveiled what it calls the first AI foundation model for payments. Trained on all of Stripe's historical payment data, the proprietary model replaces earlier single-purpose models and makes inferences for fraud prevention and conversion optimization on behalf of merchants.

• Stripe reported strong stablecoin growth, adding money management functionality across 100 new countries and a Visa-backed stablecoin card. Total stablecoin value grew 40% over the past year, driven by real-world business payouts from companies like X, Scale AI, and remote.com rather than speculation, following Stripe's acquisition of Bridge.

• Stripe launched a toolkit for agentic buying, aiming to move commerce from web browsers into AI tools like ChatGPT, Claude, and Perplexity using the MCP standard. Collison compared adjusting to AI-driven purchases to riding in a Waymo, where initial nervousness quickly becomes normalized.