The Agents Economy Backbone - with Emily Glassberg Sands, Head of Data & AI at Stripe
Timestamps are as accurate as they can be but may be slightly off. We encourage you to listen to the full context.
Podcast Summary
Emily Glassberg Sands, Head of Data & AI at Stripe, explores how the company is building economic infrastructure for AI while processing $1.4 trillion annually (~1.3% of global GDP). She shares insights on Stripe's domain-specific foundation models that dramatically improved fraud detection rates from 59% to 97%, the launch of the Agentic Commerce Protocol (ACP) with OpenAI creating shared standards for AI agent commerce, and how AI companies are managing new fraud vectors while scaling globally faster than any previous wave of startups. (00:27)
- Core themes include AI's transformation of financial infrastructure, the evolution from traditional ML to foundation models, agentic commerce protocols, and the real economic impact of AI companies showing 2-3x faster growth than previous SaaS cohorts
Speakers
Emily Glassberg Sands
Emily is the Head of Data & AI at Stripe where she leads efforts to build financial infrastructure for the internet and leverage AI to power Stripe's products. An economist by training, she previously spent eight years at Coursera during its growth from under 40 people to a major educational platform. At Stripe, she oversees data platform, ML infrastructure, AI infrastructure, and experimental projects including the Agentic Commerce Protocol and token billing initiatives.
Alessio Fanelli
Alessio is the founder of Kernel Labs and co-host of the Latent Space podcast. He focuses on AI and technology investing, bringing practical insights from the startup ecosystem to discussions about emerging AI trends and business models.
Swyx (Shawn Wang)
Swyx is the editor of Latent Space and a prominent voice in the AI developer tools community. He brings extensive experience in developer tooling and AI infrastructure to explore how AI is transforming software development and business operations.
Key Takeaways
Build Domain-Specific Foundation Models for Real-Time Applications
Stripe developed their own domain-specific foundation model that processes every transaction in under 100 milliseconds, creating dense "payments embeddings" from tens of billions of transactions. (03:24) This approach enabled them to detect sophisticated card-testing attacks that traditional ML models missed, improving detection rates from 59% to 97% at large users. The foundation model captures sequences and relationships in payments data, similar to how language models understand word context, allowing them to identify anomalous patterns that would be invisible to smaller, task-specific models.
Address AI-Specific Fraud Vectors with Real-Time Solutions
AI companies face unique fraud challenges including free trial abuse, refund abuse, and nonpayment abuse that can be "existentially threatening" due to high GPU and inference costs. (09:59) Unlike traditional SaaS where marginal costs were near zero, AI services have substantial compute costs that make friendly fraud particularly damaging. Stripe developed radar extensions specifically for AI businesses, enabling them to flag suspicious transactions even when they don't result in chargebacks, helping companies maintain healthy unit economics while still offering free trials and flexible pricing models.
Embrace Protocol-First Thinking for Market Expansion
The Agentic Commerce Protocol (ACP) demonstrates the power of creating open standards rather than proprietary solutions. (23:09) By partnering with OpenAI to establish a shared language for agent-to-merchant interactions, Stripe enabled major retailers like Walmart and Sam's Club to participate in agentic commerce. The protocol approach allows merchants to integrate once and reach multiple AI agents, while agents can access diverse product catalogs through a standardized interface, creating network effects that benefit the entire ecosystem.
Implement Token Billing for Dynamic Cost Management
AI companies need flexible billing that adjusts to real-time inference costs as underlying LLM prices fluctuate dramatically. (16:42) Stripe's token billing API allows businesses to track and price to inference costs in real time, protecting unit economics when model costs increase 3x or capturing competitive advantage when costs drop 80%. This is crucial for wrapper businesses whose service pricing depends heavily on upstream LLM costs, enabling them to maintain healthy margins while staying competitive in rapidly changing markets.
Scale Global Operations from Day One
AI companies are going global faster than any previous wave, with the top 100 AI companies on Stripe operating in a median of 55 countries by the end of their first year. (15:09) This global reach is enabled by accepting diverse payment methods, stablecoins for high-value international transactions, and optimized checkout experiences. Companies like Shadeform see 20% of their volume coming through stablecoins, saving significant margin on international transactions while accessing incremental revenue from global customers who prefer alternative payment methods.
Statistics & Facts
- Stripe processes approximately $1.4 trillion annually, representing about 1.3% of global GDP, giving them unique visibility into economic trends and transaction patterns. (00:27)
- The top 100 AI companies on Stripe reach revenue milestones 2-3 times faster than comparable SaaS companies from five years prior and operate in twice as many countries by the end of their first and second years. (1:22:02)
- 8,500 Stripe employees (out of approximately 10,000 total) use LLM-based tools daily, with 65-70% of engineers using AI coding assistants regularly, demonstrating widespread internal AI adoption. (49:59)
