No PriorsNo Priors Ep. 50 | With Stripe Head of Information Emily Glassberg Sands
CHAPTERS
- 0:00 – 0:34
Emily’s path to Stripe and what drew her in
Elad introduces Emily Glassberg Sands and her background leading data science at Coursera. Emily explains why Stripe’s scale and payments data created a unique opportunity to learn from the economy and help businesses succeed.
- •Emily’s prior work at Coursera on AI-powered personalized learning and skill measurement
- •Why Stripe’s payments footprint implies uniquely rich data about businesses and the economy
- •Stripe’s ability to turn insights into product interventions that improve customer outcomes
- 0:34 – 2:21
Inside Stripe’s “information” org: data foundations + product impact
Emily outlines the remit of Stripe’s information organization and how it spans internal decision-making, data-powered products, ML infrastructure, and corporate tech. She also describes her second responsibility: growing Stripe’s self-serve SMB/startup business.
- •Two hats: enabling data use across Stripe and owning the self-serve business
- •Investments in ML infrastructure and data organization as critical foundations
- •Serving SMBs/startups across payments plus adjacent products (billing, tax, rev rec, etc.)
- •Focus on efficient onboarding/integration and expanding product adoption
- 2:21 – 3:42
From traditional ML to the LLM “what do we do now?” moment
The conversation shifts from Stripe’s long-standing ML use (fraud/risk) to the emergence of generative AI. Emily describes the mix of awe and overwhelm when LLMs hit the mainstream and how Stripe framed the opportunity and risk.
- •Stripe has deep ML roots in fraud, risk, and payments optimization
- •LLMs created a clear opportunity—but required fast, safe execution
- •Early question: what concrete user value exists and how to discover it quickly
- 3:42 – 4:42
LLM Explorer: rapid internal enablement and safe experimentation
Emily shares the origin story of LLM Explorer—built by three engineers in weeks—to put a ChatGPT-like tool in every employee’s hands. The focus was bottoms-up learning with guardrails like PII handling and secure usage.
- •Bottoms-up experimentation as a cultural advantage at Stripe
- •LLM Explorer launched with GPT-3.5/4 and expanded to many models
- •Security by design (e.g., stripping and rehydrating sensitive data)
- •Goal: let thousands of employees discover high-value use cases
- 4:42 – 8:16
Driving adoption: community mechanics, presets, and viral internal usage
They discuss how Stripe encouraged broad internal usage and learning. Emily explains how shared “presets” (reusable prompt patterns) helped institutionalize best practices beyond Slack threads and accelerated cross-team reuse.
- •Slack channel + hackathon used, but enthusiasm drove adoption naturally
- •Usage scaled quickly (large share of company within days; sustained WAU)
- •“Presets” system for saving/sharing prompts; search and upvote to surface best patterns
- •Example: Stripe tone/style-guide preset used across roles (marketing, sales, execs)
- 8:16 – 10:33
Applied ML accelerator teams: seeding bets with 1–2 pizza teams
Sarah asks about Stripe’s accelerator model for moving from exploration to real systems. Emily describes ring-fenced teams funded for ~6 months to create durable prototypes (including LLM Explorer) and to rotate experienced internal talent into new AI work.
- •Accelerators create protected capacity for experimentation without disrupting core teams
- •Run from the CTO’s office with clear charters and iterative milestones
- •Multiple accelerators span infra and applied use cases (e.g., support tooling)
- •Rotation model develops internal talent; many participants are long-tenured Stripes
- 10:33 – 10:39
User-facing LLM products: assistants that generate code and retrieve info
Emily highlights two primary user-facing patterns: code generation and faster information retrieval. She introduces Radar Assistant and Sigma Assistant as betas designed to democratize capabilities for non-technical users and speed up technical workflows.
- •LLMs most useful today: automating code writing + accelerating information retrieval
- •Assistants turn natural language into actionable configurations or queries
- •Democratization: enabling non-technical users to access advanced tooling
- 10:39 – 13:27
Radar Assistant: natural-language custom fraud rules to fight faster
Radar Assistant helps customers generate custom fraud rules using plain English rather than writing code. Emily explains why speed matters in fraud and how this expands who can contribute to fraud operations, including SMBs without developer resources.
- •Built on Stripe Radar’s fraud/risk foundation plus customer-defined rules
- •Natural language → custom fraud rules without needing coding skills
- •Faster iteration helps businesses stay ahead of fraudsters
- •Expands usage to less-technical teams and smaller merchants
- 13:27 – 14:29
Sigma Assistant: natural-language analytics on Stripe revenue data
Sigma Assistant turns natural-language questions into SQL/reporting outputs so customers can analyze their Stripe data without knowing SQL. Emily describes example questions from basic revenue reporting to behavioral/payment-delay analysis.
- •Sigma is Stripe’s SQL-based reporting/insights product on Stripe data
- •Assistant enables NL querying for revenue, retention, churn, payment behavior
- •Targets business users who need insights but lack SQL expertise
- •Planned broader rollout beyond beta
- 14:29 – 16:58
3–5 year vision: foundation models on financial data and an “economic OS”
Elad asks how gen AI could change Stripe in the coming years. Emily lays out a bigger vision: foundation models trained on financial data to materially improve payments optimization and potentially evolve Stripe into an economic operating system for businesses.
- •Potential for foundation models to outperform incremental ML in auth/conv/fraud/cost
- •Payments optimization as a major lever with meaningful bps impact
- •Product direction: dashboards → recommendations → APIs → more automated business decisions
- •Long-term ambition: help with pricing, discounting, geo strategy, personalization
- 16:58 – 18:26
Scaling investment: when to go beyond experiments and how to organize
Sarah probes how Stripe decides to invest more deeply than initial small teams. Emily explains an iterative operating model combining more accelerator teams with embedding gen AI adoption inside core vertical product teams.
- •No single “right model” yet; approach is intentionally iterative
- •Parallel paths: grow accelerator capacity and enable vertical teams directly
- •Two north stars: empower non-technical users + make technical users dramatically faster
- •Move from prototypes to production-grade, scalable implementations
- 18:26 – 20:45
Model strategy and internal LLM platform: centralize plumbing, keep team agency
Elad asks about Stripe’s proliferation of models and selection criteria. Emily describes a centrally operated internal LLM API with defaults and safety, while leaving application teams flexibility to trade off cost, latency, and quality; expensive usage is charged back.
- •Internal tool evolved into a programmatic LLM API used by dozens of apps
- •Centralization reduces overhead (agreements, infra, safety) and creates economies of scale
- •Teams choose models based on cost, latency, and performance requirements
- •Chargeback for costly applications encourages pragmatic scaling decisions
- 20:45 – 22:01
Build vs buy for experimentation: why Stripe built its own platform
Sarah asks what infrastructure Stripe builds centrally. Emily explains Stripe’s hybrid approach—using third-party tools where possible but building internally for unique constraints—highlighting their in-house experimentation platform due to stringent charge-level latency/reliability needs.
- •Buy where strong third-party solutions exist; build where needs are unique/mission-critical
- •Stripe runs high-stakes, charge-level experiments with stringent requirements
- •Internal experimentation platform justified by latency and reliability constraints
- •Still leverages external ML tooling where appropriate
- 22:01 – 25:38
AI whitespace in fintech: identity, integration automation, and business growth layers
Elad asks where AI can have the biggest impact beyond Stripe. Emily points to merchant identity understanding, mapping business activity to regulatory constraints, making financial integrations more seamless via code generation, and building growth insights on top of payments data.
- •Identity: verifying who merchants are and what they sell in complex regulatory settings
- •Operational/supportability and network/bin-sponsor constraints as AI-relevant classification problems
- •Automating robust, merchant-specific integrations using LLM code generation
- •Using financial data to help businesses grow—aligned incentives can “grow the pie”
- 25:38 – 27:48
How Stripe already turns payments data into customer value (and what’s next)
Sarah asks for concrete examples of giving insights back to merchants. Emily details front-end checkout optimizations and back-end authorization, smart dunning, and fraud systems—showing how scale data drives conversion and revenue improvements, with room to go further.
- •Optimized Checkout Suite: dynamic payment method presentation and UX improvements
- •ML for authorization optimization: recover false declines via routing/retry strategies
- •Smart dunning in Billing reduces declines by retrying at optimal times
- •Radar evaluates many transaction features quickly to block fraud while allowing good payments
- 27:48 – 30:08
Labor economics through-line: using causal data to improve decisions and opportunity
Sarah asks how Emily’s labor economics training shapes her leadership. Emily recounts early audit-study work on gender bias in theater, her focus on causal inference to drive change, and how that mission connects to Coursera and Stripe’s “beneficent” approach.
- •Career theme: using data to understand and improve decisions by individuals and firms
- •Audit study example: pen-name gender effects; data spurred awareness and change
- •Importance of robust econometrics/causal inference to find true drivers
- •Motivation for Coursera (access to learning) and Stripe (help businesses grow)
- 30:08 – 32:41
Macro signals from Stripe’s data: internal strategy and future customer-facing insights
Elad asks how Stripe uses its macroeconomic vantage point. Emily explains Stripe uses these signals for internal decisions and is exploring ways to provide earlier, more actionable macro insights to users—potentially months ahead of traditional indices.
- •Stripe can infer near-real-time economic trends from transaction data
- •Macro insights inform internal decision-making and long-term investment choices
- •Open question: productizing macro signals for customers (earlier than CPI/indices)
- •Fits the broader “economic operating system” vision
- 32:41 – 35:33
AI and education: personalization is important, but skills signaling may matter most
Elad pivots to Emily’s Coursera experience and AI’s impact on education. Emily argues AI can personalize learning and expand instruction quality globally, but emphasizes the labor-market “pull-through”: skill measurement and credible credentials as equitable signaling mechanisms.
- •AI can improve discovery, learning personalization, and scalable teaching
- •Education’s labor-market role: skills development plus signaling/credentialing
- •Shift toward skills as a labor-market currency enabled by measurement and data
- •Equity angle: fairer opportunity via demonstrable skills rather than proxies
- 35:33 – 39:24
Unique needs of AI startups on Stripe: compute costs, global demand, and early monetization
Elad asks what distinguishes AI-native companies from prior startup waves. Emily outlines four differences: high compute costs forcing faster monetization, global demand from day one, heavy subscription skew, and “growing up” operationally earlier—driving adoption of financial automation tooling.
- •Compute costs pressure AI startups to monetize earlier than prior SaaS waves
- •Borderless distribution creates immediate cross-border payments needs
- •Strong skew toward subscription business models, especially consumer AI
- •Lean teams must operate like mature businesses early; adoption of rev/finance automation
- •Rapid growth in AI companies using Stripe; many top AI startups monetizing fast
