Episode Summary
Executive Summary: Emily Glassbrook-Sands describes how Stripe’s information org combines data science, ML infrastructure, and business applications to improve both internal operations and customer products. The conversation centers on Stripe’s rapid, bottom-up adoption of LLMs, the creation of internal tooling and accelerators, the launch of AI assistants for fraud and analytics, and Stripe’s broader vision of using financial data to become an economic operating system for businesses.
Main Topics: Stripe’s information org and mandate (Priority: 5/5): Emily explains her dual role: enabling data use across Stripe through foundations, ML infrastructure, and GenAI applications, while also owning the self-serve business and helping SMBs integrate and expand their Stripe usage. Bottom-up LLM adoption inside Stripe (Priority: 5/5): Stripe’s early internal LLM Explorer began as a three-engineer, three-week beta and quickly spread across the company. The team used presets, Slack, and security controls to turn experimentation into shared organizational learning. From experimentation to production AI products (Priority: 5/5): Stripe created internal accelerators and an internal API to help teams move from prototypes to production systems. This centralized some infrastructure while still letting teams choose models based on cost, latency, and quality trade-offs. AI assistants for fraud and analytics (Priority: 5/5): Emily highlights Radar Assistant and Sigma Assistant as user-facing examples of generative AI. Radar Assistant turns natural language into fraud rules, while Sigma Assistant turns natural language into SQL-based business analysis. Stripe’s data advantage and future economic platform (Priority: 4/5): Stripe sees AI as a way to improve payments optimization, help businesses make pricing and personalization decisions, and eventually surface real-time economic signals that make Stripe more like an operating system for commerce. AI startups as a distinct customer segment (Priority: 4/5): Stripe is seeing AI-native companies monetize earlier, operate with high compute costs, serve global demand from day one, and rely heavily on subscriptions and revenue automation. Emily’s economics background and education perspective (Priority: 3/5): Her labor economics training shapes her focus on decision quality, access to opportunity, and measurable outcomes. At Coursera and beyond, she emphasizes that AI should improve both skill development and labor-market signaling.
Key Arguments: Stripe’s core advantage is not being an AI company per se, but having rich financial data that can improve decision-making, risk management, and business growth. LLMs should be introduced internally first so employees can discover use cases, share prompts, and build confidence before products are launched externally. A small, durable accelerator model is effective for seeding new AI initiatives without disrupting core teams. Centralized ML and LLM infrastructure reduces friction around model selection, enterprise access, security, and deployment, but teams still need autonomy for trade-offs. Generative AI is especially powerful when it lets non-technical users perform technical tasks, such as writing fraud rules or querying data with natural language. Stripe’s long-term opportunity is to move beyond payment processing into helping customers optimize pricing, personalization, fraud, and broader business operations using real-time data. AI startups differ from prior software waves because they have immediate compute costs, early monetization pressure, global demand, and lean teams that must operate like mature businesses. Emily’s economics lens emphasizes causal inference, labor-market outcomes, and using data to improve access, quality of decisions, and opportunity. AI may reshape education most meaningfully when it improves both learning personalization and skill credentialing/signaling in the labor market. Stripe’s financial vantage point could eventually help produce real-time macro signals and business insights that are more actionable than traditional lagging indicators.
Data Points: Stripe employees using LLM Explorer within days: about one-third of Stripe - Internal LLM Explorer adoption shortly after launch Weekly active users of LLM Explorer: almost 3,000 - Current weekly usage of Stripe’s internal LLM Explorer Company-wide usage share: just shy of half the company every week - Based on weekly active users of the internal tool Reusable LLM interaction patterns/presets: about 300 - Shared prompt patterns created after launching presets in the internal tool Models served in LLM Explorer: over a half dozen - The tool began with GPT-3.5 and GPT-4 and later expanded to more models Applications built on internal LLM API: 60 applications - Internal API for programmatic use of LLMs, including internal and external apps Accelerator team size: one- to two-pizza teams - Stripe’s internal experimental AI accelerator model Accelerator duration: six months - Funding window for durable experimental bets Radar false declines recovered: about 10% - Stripe uses ML to optimize authorization requests and recover false declines Smart Dunning decline reduction: about 30% - Recurring payments optimization in Stripe Billing Radar transaction characteristics: 1,000 characteristics - Radar evaluates many signals to decide whether billions of legitimate payments can proceed Radar decision latency: less than 100 milliseconds - Time to evaluate a payment transaction Coursera team size at Emily’s arrival: 40 people - She notes Coursera was very small when she joined Women-written U.S. stage productions: half today - Emily references an audit study that helped change theater production norms over time GenAI company growth on Stripe: massive spike over the last year - Stripe observed rapid growth in generative AI startups using its platform Forbes top 50 AI companies using Stripe: over half - More than half of Forbes’ top 50 AI companies were using Stripe AI customer examples: Otter AI, Midjourney, OpenAI, Mistral, Moonbeam, Runway - Examples of AI companies on Stripe, from foundational models to applications
Pivotal Quotes: "we’re really focused on understanding who those users are, getting them the right shape of integration efficiently, building product experiences that meet their needs" — Emily Glassbrook-Sands: Describing her responsibility for Stripe’s self-serve SMB and startup business "let’s get a chat GPT-like interface in the hands of the 7,000 talented Stripe employees and really let them figure out how to apply it to their work" — Emily Glassbrook-Sands: Explaining the origin and intent of Stripe’s internal LLM Explorer "The vision, the opportunity is much bigger than what we could do in a year" — Emily Glassbrook-Sands: Her view on Stripe’s long-term GenAI strategy and economic-platform ambitions
Implications: For listeners, the key takeaway is that AI value often starts internally, then becomes productized through infrastructure and domain-specific workflows. For fintech, the big opportunities are fraud, payments optimization, identity, and business intelligence. For AI startups, Stripe is signaling that monetization and global scale are arriving much earlier than in prior software waves.