The Rise of Agentic Commerce — Emily Glassberg Sands (Stripe)

The MAD Podcast with Matt Turck
Guest: Emily Glassberg Sands, Head of Information, Stripe
Date: July 10, 2025

Episode Overview

This episode features a deep-dive conversation with Emily Glassberg Sands, Head of Information at Stripe, exploring the company's AI-driven transformation, the rise of "agentic commerce" (software agents transacting on our behalf), and the broader implications for payments infrastructure and the AI startup economy. The discussion covers Stripe’s proprietary foundation model, its unique approach to AI, real-time infrastructure demands, business model evolution, the global spread of AI startups, and the emerging landscape where autonomous agents conduct commerce.

Key Discussion Points & Insights

1. Stripe’s Scale, Mission & The Expanding Role of AI

• Stripe’s scale and impact:

  • Processes about $1.4 trillion yearly (1.3% of global GDP), with 50,000 new transactions every minute ([00:00]–[02:05]).
  • Stripe supports businesses from small entrepreneurs to Fortune 100 companies with a programmable financial infrastructure—no longer simply a payments API ([02:05]).
  • Stripe’s tools provide a "structural tailwind" for Internet economy growth ([03:39]).

    "Businesses on Stripe grew seven times faster last year than the S&P 500." — Emily ([04:00])

• Role of the Information Org:

  • Uses data end-to-end, powers internal/external ML, drives product-led growth, and runs Stripe’s experimental projects team ([04:17]).

• Emily’s journey:

  • From academic economist to Coursera (AI-powered learning), to Stripe for mission alignment, scale, and a real-time macroeconomic dataset ([05:56]).

2. Stripe’s Foundation Model: Why & How They Built It

• Why build a proprietary foundation model?

  • Stripe’s dataset is unique; general LLMs (OpenAI, Anthropic) lack Stripe’s detailed transaction data ([09:39]).

    “Stripe is a little bit different. We have really differentiated data… OpenAI doesn't have that data.” — Emily ([09:58])

• Language-like signals in payments:

  • Payments, at scale, share structures with language: sequences, semantics, relational context ([09:39], [13:55]).
  • Inspired a fully unsupervised embedding-based approach: each charge gets its own vector in a contextual space ([13:55]).

• Embeddings power unsupervised learning, overcoming label limitations ([13:55]):

  "It doesn't require any labels... you can actually use all of the tens of billions of transactions." — Emily ([14:36])

• Traditional ML vs. Foundation Model:

  • They currently ensemble foundation and traditional models; foundation model dramatically boosted recall in fraud detection ([17:18]).

    "Detection rate on large margins went from 59% to 97%." — Emily ([19:43])

  • Some tasks (e.g., card testing) are only possible with foundation-style models.

• Development journey:

  • Initial instinct—train with bigger transformers on single payments—was wrong ([20:18]).

    "Our first instinct was actually full on wrong." — Emily ([20:18])

  • Key breakthrough: model short sequences of payments, not individual records ([20:41]).
  • BERT encoder backbone outperformed GPT-style architectures for understanding tasks ([21:57]).
  • Built by a small team working in a quasi-research structure ([22:00]).
  • Required custom tokenization, data loaders, and robust checkpointing—Stripe as an “AI lab” ([23:21]).

• Operational challenges:

  • Shadow mode deployment was straightforward; real-time product requirements (latency, reliability) were more challenging ([24:36]).

3. Explainability, Product Deployment, and Use of AI at Stripe

• Explainability and regulatory demands:

  • LLMs are increasingly used for explanations; Stripe balances transparency, rule-based speed/clarity, and nuanced ML ([25:51]).

    "LLMs are actually getting quite good at explainability... we will continue to use rules and models in parallel." — Emily ([25:51])

• Where to apply AI?

  • Always starts with identifying user pain/business need, then checks for abundant Stripe-specific data, then evaluates amenability to AI ([29:18]).
  • Example: Smart Disputes automates chargeback evidence gathering, boosting recovery rates without added labor ([29:18]–[34:10]).

• Revenue generation through AI:

  • AI is used across the payments lifecycle (checkout, auth, fraud, disputes).
  • Optimized Checkout Suite personalizes payment methods, boosting conversion and revenue (e.g., Turo recovered $100M+, 5% boost) ([34:32]).

    "Businesses that show at least one relevant payment method beyond just cards see like a 12% increase in revenue." — Emily ([36:55])

4. Data Infrastructure & Real-Time ML at Stripe

• ML and data infra stack:

  • Databricks notebooks, Flyte for orchestration, Nvidia GPUs, PyTorch, and Shepherd (now open-sourced as Cronon) for feature serving ([38:08]).
  • Crisis of divergence solved by building a unified infra layer; enables feature sharing, efficiency, and global scalability ([38:45]).

• Real-time requirements:

  • Stripe operates at "56 nines" reliability; real-time and security demands sometimes require in-house innovation ([41:35]).

    "You can't have downtime... even Radar API downtime is super, super costly." — Emily ([41:35])

5. The Emergence of Agentic Commerce

• Definition and current applications:

  • Agency shifts from humans to AI agents that execute tasks/end-to-end purchases ([43:13]).

    "AI is no longer just about getting answers to your questions. It's starting to do things for you." — Emily ([43:17])

  • Early use cases: Barista Agent buys coffee; Perplexity allows in-app hotel booking; Hipcamp bookings via AI ([43:13]–[46:20]).

• Behind the scenes:

  • Agentic money flows are akin to existing controlled virtual card architectures (e.g., DoorDash)—now done by bots instead of humans ([46:37]).
  • Agents enable "in situ commerce"—buying within apps/developer tools, creating new channels/convenience ([48:27]).

• Systemic implications:

  • As agents transact, the product interface becomes structured data ("intent as the interface").

    "Humans click around, agents just declare what they want." — Emily ([49:50])

  • Product data must be machine-readable; merchants must expose APIs for agentic buying ([49:50]).

• Fraud and risk changes:

  • Credentials move from being "owned" to being permissioned (time/amount limited, scoped tokens) ([51:24]).
  • Need for better bot distinction and real-time log observability ([52:15]).

6. MCP Server: Protocol for Agent-to-Agent & LLM-to-API Integration

• What is MCP and why it matters:

  • MCP server lets LLMs/agents interact with Stripe’s APIs—bridging AI models with deterministic SaaS actions ([56:02]).
  • Stripe’s MCP server enables low-risk, high-frequency admin tasks (e.g., issuing refunds, retrieving invoices).

    “It’s pretty clear that MCP is becoming the default way any single service… talks to an LLM.” — Emily ([58:54])

  • Case study: Decagon reduced support costs 65% by automating Stripe-admin tasks via MCP ([57:19]).

• Future:

  • MCP will incorporate monetization (MCP payments) ([59:17]).

7. The New AI Startup Economy: Unprecedented Growth

• Stripe’s vantage point:

  • 78% of Forbes AI 50 are Stripe users ([60:06]).
  • AI companies monetize up to 3x faster than previous SaaS cohorts (median: $30M ARR in 1.5 years vs. 5.5 years for SaaS) ([61:38]).

    “They are monetizing faster than any previous generation of startups that we’ve seen.” — Emily ([60:28])

• Hyper-global from day one:

  • Median AI startup sells in 55 countries their first year, 80 by their second ([63:11]).

    "They already sell into 184 markets... you don't have to go that deep into your geography background to know there aren't that many more markets to sell into." — Emily ([63:41])

  • Enabling factors: LLMs’ translation ability and Stripe’s globalized infrastructure ([63:58]).
  • AI companies are leaner: multi-million-dollar businesses with just 10-30 people ([64:44]).

• Market specialization and business model innovation:

  • The global reach makes "vertical" (niche) AI products economically viable ([65:52]).

    “Specialization is rewarded. Because the markets are bigger and so even a very specialized niche is a very large business.” — Emily ([66:49])

  • Pricing models evolve from per-seat (classic SaaS) to usage-based or outcome-based to reflect inference costs and value alignment ([67:58]).

    "As products get more AI centric… companies are shifting from this sort of per seat billing… to usage based billing… and experimenting with outcome based pricing." — Emily ([68:00])

  • Example: Intercom moved from per-seat to per-resolved-case pricing ([69:49]).

8. Building AI Literacy and Culture at Stripe

• Experimentation as the foundation:
  • Internal LLM explorer gave thousands of employees safe, model-agnostic access to LLMs ([71:32]).
  • Culture encourages the creation/sharing of prompts, reusability, and internal agent development ([72:54]).

    “We enabled these things called presets, which are basically shareable prompts. And basically overnight the Stripe community developed hundreds of these of reusable LLM interaction patterns.” — Emily ([72:18])

9. The Roadmap and Looking Ahead

• Expansion includes:
  • Rolling out foundation models across applications
  • "Risk as a service" offerings
  • Advancing AI-driven commerce tools and agentic infrastructure ([74:14])
• Expect ongoing work on MCP, order intent, and agentic commerce scenarios ([74:14]).

Notable Quotes

• "Businesses on Stripe grew seven times faster last year than the S&P 500." ([04:00])
  — Emily Glassberg Sands

• “Stripe is a little bit different. We have really differentiated data … OpenAI doesn’t have that data.” ([09:58])
  — Emily

• “Our first instinct was actually full on wrong.” ([20:18])
  — Emily, on Stripe's initial approach to their foundation model

• "LLMs are actually getting quite good at explainability... we will continue to use rules and models in parallel." ([25:51])
  — Emily

• "Companies are shifting from this sort of per seat billing… to usage based billing… and experimenting with outcome based pricing." ([68:00])
  — Emily

• "Specialization is rewarded. Because the markets are bigger and so even a very specialized niche is a very large business." ([66:49])
  — Emily

• “Humans click around, agents just declare what they want.” ([50:30])
  — Emily

• “It’s pretty clear that MCP is becoming the default way any single service… talks to an LLM.” ([58:54])
  — Emily

• “They are monetizing faster than any previous generation of startups that we’ve seen.” ([60:28])
  — Emily

Timestamps of Key Segments

| Topic | Time | |---------------------------------------------|---------------| | Stripe’s scale and infra | 00:00–04:06 | | The Information Org's scope | 04:17–05:37 | | Emily’s background and joining Stripe | 05:56–08:53 | | Why Stripe built its own foundation model | 09:39–13:16 | | Foundation model vs. traditional ML | 16:24–20:07 | | Model development & challenges | 20:18–24:25 | | Explainability and transparency | 25:24–28:37 | | Where to use AI (Smart Disputes) | 29:18–34:10 | | Revenue-generating AI (Optimized Checkout) | 34:32–37:46 | | Stripe's ML/data infrastructure | 37:46–41:22 | | Real-time and reliability considerations | 41:22–42:50 | | The agentic commerce paradigm | 43:13–49:35 | | Technical requirements for agentic commerce | 49:50–55:44 | | MCP protocol and server | 56:02–59:30 | | AI startup economy trends | 60:06–64:56 | | Specialization, verticalization | 65:52–67:18 | | Pricing and billing transformations | 67:58–70:57 | | AI literacy & internal culture at Stripe | 71:32–74:04 | | Roadmap and next 12–18 months | 74:04–74:44 |

This conversation is a must-listen for anyone tracking the intersection of AI, fintech infrastructure, and the next frontier of digital commerce. Emily brings clarity, specificity, and candor, with real-world examples from Stripe’s vantage point at the center of the new AI economy.