Latent Space: The AI Engineer Podcast

Episode Summary: "The Agents Economy Backbone"

Guest: Emily Glassberg Sands, Head of Data & AI at Stripe

Date: October 30, 2025

Overview

This episode explores Stripe’s central role in the emerging “AI agents economy”—detailing how foundational AI and ML infrastructure at Stripe powers global-scale financial systems, fraud detection, and next-generation agent-driven commerce. Emily Glassberg Sands, Stripe’s Head of Data & AI, speaks with hosts Alessio and Swyx about the rapid evolution in payments, billing models, fraud vectors, and why agent-to-agent protocols are the next step for automated online commerce. The episode offers a front-row view of how Stripe leverages huge volumes of global payments data and builds infrastructure for both startups and Fortune 500s navigating the AI revolution.

Key Discussion Points and Insights

1. Stripe’s Data & AI Org: Mission and Scope

• [00:15-02:18]
  • Stripe’s mission expanded from payments to broader financial infrastructure; now includes economic infrastructure for AI.
  • Emily leads data, ML, AI platform, and experimental projects teams.
  • The experimental projects team (≈24 people) focuses on 0→1 innovation across the company, yielding products like agent commerce and token billing.
  • Stripe processes $1.4T annually (≈1.3% global GDP), yielding vast, unique data for product optimization and fraud detection.

2. Stripe’s AI Evolution & Foundation Models

• [02:32-04:37]

  • Serious investments in AI began post-GPT3.5 (late 2022–early 2023).
  • Early ML used for fraud and onboarding, but LLMs/AI are now integrated across Stripe for production-grade applications and “domain-specific foundation models.”
  • At Stripe’s scale, needs evolved from “single-task ML” to rich “payment embeddings” enabling downstream applications.
  • Stripe moved toward its own foundation models to serve dense, high-volume payment inference—e.g., detecting card-testing fraud.

  “We actually need to have our own domain specific foundation model… move from these, you know, single task point solution ML models to, you know, a much richer, denser payments embeddings that can then power [the] downstream applications.”
  — Emily, [03:15]

3. Real-time AI for Fraud Detection

• [05:08-08:08]

  • Stripe’s foundation model processes every transaction—e.g., catching “card-testing” attackers who hide in large ecommerce flows.
  • Payment data is encoded as embeddings, with contextual/sequential data (like language models), boosting the detection of subtle, anomalous behaviors.
  • Card-testing detection rate jumped from 59% to 97% with the new system.
  • Faster deployment: The embedding model enables rapid clustering to flag “suspicious but not disputed” transactions—meeting unique needs of AI companies.

  “Each charge becomes this dense embedding. You start to see these clusters sort of pop out and you know in real-time that they’re card testing and you can block them. So yes, it is happening on the charge path in less than 100 milliseconds of latency.”
  — Emily, [05:17]

4. The New Fraud Landscape for AI Companies

• [09:35-13:49]
  • “Friendly fraud” (i.e., nonpayment, free trial, and refund abuse) is now existential—especially with high-cost inference and global access.
  • Marginal cost of AI services (GPUs, compute) makes abuse far more costly than in SaaS.
  • Real-world examples:
    • Some AI startups have resorted to disabling free trials, “choking” their own revenue out of fraud fear.
    • “Free trial credit cards” are widely marketed – a problem Stripe is now tackling at the source.
    • Refund abuse particularly targets mid/high-tier AI subscriptions.
  • Stripe now offers explicit fraud solutions for new abuse cases seen in the AI economy.

5. Billing Models and Stripe as the “Skeletal System” for AI Businesses

• [14:11-20:50]

  • AI startups are more global from day one—median operates in 55 countries in first year.
  • Startups require “turnkey” monetization: payments, global methods, fraud, billing (including fixed, usage, and outcome-based).
  • Token Billing API:
    Stripe’s new API tracks and prices LLM inference in real time, so services can dynamically adjust pricing as underlying model costs fluctuate.
  • Outcome-based billing:
    Beta adoption (e.g., pay for “support cases resolved”) helps overcome enterprise buyer friction and speeds AI product adoption.
  • Stablecoins:
    Rapidly growing among global/high-value AI businesses. Example: Shade Form, where 20% of their payment volume is stablecoins, reducing margin-eating international card fees.

  “I think of Stripe as the skeletal system for AI companies… All of the Forbes AI50 who monetize online monetize through Stripe.”
  — Emily, [14:11]

6. Network Effects and Measurement

• [21:36-22:38]
  • Network density (e.g., over half of Lovable’s payments through Stripe Link) and spread (Herfindahl index) are actively measured.
  • Stripe’s goal: global diversification, not just dominance in one space.

7. Agentic Commerce Protocol (ACP): The Backbone for Agent-Driven Transactions

• [23:09-29:16]

  • Stripe and OpenAI launched the Agentic Commerce Protocol—the first open, cross-platform standard for agents to query inventories, transact, and hand off payments securely at scale.
  • Decouples agent action from merchant risk: agents get “shared payment tokens” to represent buyer credentials without holding funds or risk.
  • Designed for any payments provider and agent system—intentionally not Stripe or OpenAI exclusive.
  • Major merchant traction: Shopify (1M+ stores), Walmart, Sam’s Club, and others joining.
  • Fraud prevention includes agent “goodness scores” to help merchants distinguish good/bad bots.

  “It’s not about Stripe… ACP works no matter who your payments provider is. We can pass the shared payments token over to any other PSP… It’s also not just about OpenAI.”
  — Emily, [27:28]

8. Fraud, Scalpers, and Explainability in the Agent Era

• [29:50-34:17]
  • The line between “good bots” and “bad bots” is blurring—as agents become commerce participants.
  • New challenges: ticket scalping (bots buying up all supply), deeper fraud models, and the need for proactive fraud signals before checkout.
  • Importance of explainability and appeals (“Can I speak to the manager, please?”) for user trust as AI-driven systems gate commerce.
  • Goal: separate “sheep from goats” so good users aren’t penalized.
  • Upholding market efficiency via innovation—Stripe’s incentives align to help their businesses grow.

9. Protocol vs. Product: Why ACP is Open

• [36:38-41:55]
  • The protocol route was chosen to maximize market growth and interoperability, not lock-in.
  • ACP and payment tokens are orthogonal; multiple approaches (e.g., agent wallets, one-time use cards for travel/Doordash, stablecoins) will coexist for different scenarios.
  • Stripe’s design: allow flexible, evolving money flows — both for “spend” (user to merchant) and “receive” (monetize agent services).

10. Stripe’s Internal AI - Adoption and Impact

• [49:32-54:46]
  • Rapid experimentation was bottoms-up: company-wide LLM tools (Go LLM chat interface), prompt sharing presets, etc.
  • LLM proxies for production engineer workflows (e.g., merchant onboarding/verification).
  • Around 8,500 Stripes per day (almost entire company) engage with LLM tools.
  • Heavy use of AI coding assistants (≈65–70%) but unclear how best to measure productivity impact.
  • Recent integration: LLMs used for rapid payment method integrations, documentation tasks, and leveraging internal data.

11. Social & Organizational Shifts from AI Usage

• [54:46-59:30]
  • Memo writing used to enforce “deep reasoning;” LLMs risk short-circuiting this by making it easier to produce convincing but shallow docs.
  • Emily emphasizes the need for citation (“if an LLM was used… please cite the LLM!”).
  • Depth of thought will become more, not less, important.

12. Internal RAG (Retrieval-Augmented Generation) and Tool Use

• [60:19-62:04]
  • Toolshed: Centralized “MCP server” (multi-tool access), internal chat/knowledge system, LLMs connected to internal code/data sources.
  • RAG isn't dead—combining information retrieval and tool-calling is key.

13. Data Platform Trends & Discovery

• [65:12-71:01]
  • Stripe’s “Hubert” (NLQ interface on internal data) is effective when data is well-structured, but discovery remains hard with messy/undocumented tables.
  • Strong push for semantic events and canonical datasets to power real-time insights—anticipating a future where static dashboards fade away in favor of live, agent-fed analytics.

14. Build vs. Buy Philosophy

• [71:40-78:43]
  • Stripe’s approach is layered: buy when possible, run RFP/Spotlight programs for emerging vendors, build only when nothing suitable exists or for core infra (feature platforms, charge path needs).
  • Acknowledges trade-offs of homegrown infra (e.g., Coursera’s pain) and why “rip and replace” can be essential.
  • AI solutions should work across multiple model providers for resilience.

15. The Macro View: The “AI Economy” and Bubble Debate

• [81:02-85:34]

  • Stripe sees AI companies moving “2–3x faster” to revenue milestones than top SaaS startups (from Stripe’s transaction data).
  • They’re twice as global, and operate with tiny teams/high revenue per employee & healthy unit economics (despite high inference costs).
  • Churn is high—but is intra-vertical (customers flip between providers), suggesting a highly competitive market.
  • Overall: “These are companies that are growing very quickly—faster than anything we’ve ever seen.”

  “At the end of their first year, [AI companies] are in twice as many countries. They have majority of revenue from outside their home market… Their revenue per employee is unlike any other business, including public companies known for efficiency.”
  — Emily, [81:20]

16. Will AI Show Up In the Macro-Economy?

• [88:38-91:14]
  • Theories abound for AI’s measurable impact on GDP—early evidence strongest in business creation ("tiny teams," new startup formation rate) and cost reduction, less so yet in classic productivity stats.
  • Outcome-based/usage-based pricing avoids the “seat-based death spiral” of SaaS disrupted by automation.
  • “Equality of adoption,” not just access, is key for balanced global economic growth.

17. Brands & Community Matter More Than Ever

• [92:52-94:48]

  • The new wave of successful AI companies pairs strong brand/community focus with technical prowess.

  “So much of the… value created in AI companies has actually come from… compelling brand and community. You also need them to be hyper-focused on the user and product experience—not just tech.”
  — Emily, [92:52]

18. Stripe’s Quality Obsession & Management

• [94:48-96:08]

  • Emily admires Head of Design Katie Dill’s relentless commitment to detail—and describes a company-wide culture of enforcing quality.

  “Katie does not give an inch on quality… There’s a clear line and if it doesn’t meet the quality bar, you just gotta fix it…”
  — Emily, [94:48]

Notable Quotes & Memorable Moments

• On the value of AI-driven fraud detection:

  “[We went from] 59% to 97% [detection rate] on large inserts… but the other thing that was really helpful was the speed at which we got it out… you can just move faster.”
  — Emily, [08:05]

• On existential threat of friendly fraud for AI businesses:

  “Now we’re in the world where GPUs are expensive, inference costs are high, and free trial… or general non payment abuse is existentially threatening for AI businesses.”
  — Emily, [09:55]

• On supporting token billing for dynamic LLM pricing:

  “If the cost of the underlying LLM 3X’s… you could have unit economics literally underwater if you don’t adjust your price.”
  — Emily, [14:11]

• On economic impact and efficiency:

  “Anything we do to help the businesses on Stripe grow helps us grow because, you know, they… run their business through us.”
  — Emily, [35:46]

• On branding and design:

  “So much of the… value created in AI companies has actually come from… compelling brand and community.”
  — Emily, [92:52]

Timestamps for Key Segments

| MM:SS | Topic | |-------|-----------------------------------------------------------| | 00:15 | Stripe’s Data & AI Org structure and mission | | 02:32 | Stripe’s history and evolution with AI/ML | | 05:08 | Foundation models in real-time fraud detection | | 09:55 | Shape and risk of new AI-driven fraud types | | 14:11 | Stripe as economic backbone; billing models and “wrappers”| | 17:30 | Usage, outcome, and token-based billing | | 20:50 | Stablecoins and global network effects | | 23:09 | Agentic Commerce Protocol (ACP) explained | | 29:50 | Bot/scalper fraud; explainability and appeals | | 36:38 | Openness: protocol vs. product at Stripe | | 41:55 | Evolving commerce and money flows for agents | | 49:32 | Internal AI/LLM adoption timeline and use | | 54:46 | Social contract shift: writing docs in the LLM era | | 60:19 | Internal RAG, Toolshed, and code/productivity | | 65:12 | Data platform, discovery, and semantic events | | 71:40 | Build vs. buy: philosophy and process | | 81:02 | The “AI Bubble” question and Stripe’s data | | 88:38 | Will AI show up in GDP and employment? | | 92:52 | Brand/community as defensibility in AI | | 94:48 | Stripe’s design culture and quality standards |

Takeaways & Closing

Stripe sits at the epicenter of the “AI agents economy”—providing the backbone for both the new wave of agentic commerce and powering healthier, more secure, and more dynamic online businesses. Emily’s perspective bridges deep technical insight (from fraud models, real-time payments, foundation models) to economic big picture (globalization, startup formation, GDP impact). Openness—through protocols, multi-model support, and a buy/build/replace philosophy—define Stripe’s unique value in a rapidly evolving market.

Hiring at Stripe:
Stripe is expanding in machine learning, backend, data engineering, AI infra, and platform roles.

For more details, visit latent.space.

End of Summary