Emily Glassberg Sands

Emily Glassberg Sands led Stripe’s Data & AI organization in October 2025, overseeing data platforms, data engineering, ML and AI infrastructure, and applications. Her organization also included an Experimental Projects Team tackling opportunities across product boundaries; agentic commerce and token billing emerged from that team. She publicly announced Stripe’s Payments Foundation Model at Stripe Sessions 2025, explaining a shift from specialized payment models toward representations reusable across authorization, fraud, and disputes.[1][2]

Sands explained the foundation model through an analogy: “You can think of a payment sort of like a word.” An individual charge gains meaning from surrounding behavior, but the relevant sequence need not belong to one customer. It might connect a retailer’s charges from an IP address or activity involving a particular card on a Friday night. The model used self-supervised learning on tens of billions of transactions to generate dense charge representations, moving beyond the recent labeled data and hand-engineered features typical of smaller, task-specific models.[1][2]

Her concrete example was card testing: attackers try card credentials to find working cards, hiding small test charges within large merchants’ transaction volumes. Classifying sequences of embeddings could reveal clusters of this activity. At the 2025 keynote, Sands reported that detection of these attacks on large users rose from 59% to 97%. In her October 2025 interview, she said every Stripe transaction passed through the foundation model on the charge path with less than 100 milliseconds of latency.[1][2]

Sands also argued that inference costs made free-trial, refund, and nonpayment abuse more damaging to AI businesses than to software with near-zero marginal serving costs. She recounted meeting an unnamed founder who said fraud was no longer a problem because he had shut down free trials and heavily throttled credits until customers proved they could pay. She challenged the revenue tradeoff and said they worked together to restore free trials; that work was still in flight when she described it.[1]

Sands supported releasing the Agentic Commerce Protocol (ACP), jointly developed with OpenAI, as a shared standard spanning agents and payment providers. It standardized how merchants expose catalogs, inventory, prices, and brand information to agents. A shared payment token handled a separate problem: passing payment credentials on a buyer’s behalf while leaving the merchant to undertake the charge and maintain the customer relationship. She emphasized that merchants could use payment providers beyond Stripe and that the protocol was not limited to OpenAI. Her priority was interoperability: “all we want is a shared standard.”[1]

Token billing addressed the volatility of upstream model costs. Sands described it as an API for tracking and pricing against inference costs in real time: falling costs could make a fixed price uncompetitive, while rising costs could leave a service’s unit economics underwater. She favored embedding experimental staff with the teams that would sustain a product, citing the billing team’s involvement in token billing. She also saw outcome-based pricing as a way to lower buyers’ uncertainty by tying payment to useful results, such as resolved support cases.[1]

Sands treated reliable data discovery as a prerequisite for useful internal AI. In October 2025, she said Hubert, Stripe’s natural-language data assistant, had roughly 900 weekly users, deliberately concentrated among technical domain experts who could validate answers. Its weakest evaluation area was finding the right tables and fields. Her organization responded by deprecating low-quality tables and having domain owners document canonical datasets. Giving Hubert context about a user’s place in the organization improved offline results, she said, but was not yet in production.[1]

Her build-versus-buy principle was to look for external solutions first and revisit internal tools as the market evolved. She described replacing Stripe’s early homegrown LLM chat experience and choosing Braintrust through an evaluations procurement process. Building remained appropriate when requirements demanded it: she credited Airbnb collaboration on a feature platform after an evaluated external option did not meet Stripe’s charge-path latency and reliability needs at the time. She also favored infrastructure that allowed switching among models and providers, whose cost, latency, and performance suited different tasks.[1][4]

Sands rejected document length and lines of code as sufficient measures of AI productivity: “I don’t really want more lines of code, just like I don’t want more pages of docs.” She argued that writing helps experts structure and test their reasoning, and warned that fluent generated prose can conceal insufficient thought. She advocated disclosing LLM use and required deep human approval before publishing or submitting work. She nevertheless welcomed AI for rapid experimentation, crediting Vercel’s UI prototype for clarifying claimable sandboxes more effectively than lengthy meetings and design documents. Conversations with Replit had also helped shape that work.[1]

Her optimism about AI rested partly on revenue she could observe through Stripe. In October 2025, she reported that its 100 highest-grossing AI companies reached $1 million, $10 million, and $30 million ARR two to three times faster than a historical top-100 SaaS cohort from five years earlier. She acknowledged slightly lower company-level retention, interpreting some churn as customers switching between competing AI products rather than abandoning the underlying use case. She separated these revenue observations from inference costs and declined to give a categorical answer about an AI bubble.[1]

Before Stripe, Sands spent eight years at Coursera. That earlier work supplies concrete context for her later management decisions: she led experimentation and data products, and subsequently recalled the difficulty of replacing infrastructure the young company had built itself.[1][3]

Sands joined Coursera in early 2014 after receiving a PhD in economics from Harvard. In a 2017 interview, she described herself as the first non-engineering hire on its Analytics team. Her graduate work combined machine learning and causal inference to study network effects and included experiments on the online labor market oDesk concerning hiring referrals. Recalling her Coursera entry in 2024, she said she initially applied for a sales role, repeatedly pitched measuring the labor-market value of Coursera credentials, and joined after an initial rejection and eventual agreement from Andrew and Daphne.[3][4]

At Coursera, Sands drove analysis and experimentation for early user growth and monetization strategies, then expanded her leadership across data science and business intelligence engineering. She distinguished decision science, which tested hypotheses for product and business choices, from data products, including search, personalized recommendations, and customized learning pathways. Her business intelligence group supplied data modeling, standardization, and self-service reporting. In 2025, she recalled that replacing homegrown experimentation, analytics, and ML platforms during 2018–2020 was painful but the right decision, because those platforms were not Coursera’s core competitive advantage.[3][1]

Sands argued in 2017 that diversity mattered to data science because framing questions about human behavior required creativity and empathy, alongside empirical and computational skills. In her 2024 account, she credited CTO David Singleton’s outreach through a mutual friend with her move to Stripe, initially to lead data science. By that interview, her responsibilities encompassed the information organization and self-serve business. She cited Stripe’s payment-based incentives as a reason to join: helping its businesses grow could also help Stripe grow, allowing investment in improvements without separately monetizing each one.[3][4]