The Death of Traditional E-commerce: Why AI Agents Are Taking Over

Stripe's Head of Information, Emily Glassberg Sands, recently discussed the company's advancements in AI and its impact on the future of online transactions. She highlighted how Stripe, handling 1.3% of global GDP, is leveraging its vast data to build advanced AI models, improve fraud detection, and adapt to the emerging landscape of agentic commerce. The conversation also touched on the rapid growth and unique characteristics of the current generation of AI startups.

Stripe's AI Journey: Building a Payments Foundation Model

Stripe, known for its financial infrastructure, has ventured into building its own AI foundation model. This decision was driven by the company's unique access to a massive amount of payment data—over $1.4 trillion in annual payment flow. Unlike general AI models, Stripe's model is specifically trained on transactional data, which, despite its apparent sparsity, shares similarities with language in its structured syntax and semantics.

Stripe's initial approach to building the model was to treat individual payments as isolated units. However, they quickly realized that payments, much like words in a sentence, gain meaning from their context. By stitching together short histories of transactions, the model gained a deeper understanding of payment patterns. This led to the development of a BERT encoder-based model, chosen for its effectiveness in understanding tasks rather than generating content.

Key Takeaways:

• Data Differentiation: Stripe's unique access to vast payment data (1.4 trillion annually) sets it apart from general AI labs.

• Payments as Language: Transactional data, with its agreed-upon syntax (BIN, MCC, amount) and longer-range semantics (device reuse, merchant history), can be treated like a language for AI models.

• Unsupervised Learning: The foundation model learns from tens of billions of transactions without requiring labels, allowing it to operate at very large scales.

• Efficiency in Building: Shared embeddings, available in Stripe's feature store (Shephard/Kronon), turn new model development into a quick project rather than a long one.

• BERT Encoder: Stripe chose a BERT encoder over GPT-style decoders because it's better suited for understanding payment patterns rather than generating new ones.

• Small Team, Big Impact: The foundation model was initially developed by a small team of three machine learning engineers in a "research bubble," protected from daily operational tasks.

AI in Action: Enhancing Fraud Detection and Revenue

Stripe's AI models are not just theoretical; they are actively improving various aspects of online commerce. One notable application is in fraud detection, particularly in identifying card testing. Traditional machine learning models often miss subtle card testing attempts due to their dilute nature within a large volume of legitimate transactions. However, the foundation model, by observing sequences of transactions, can identify these patterns, leading to a significant increase in detection rates.

For example, on large merchants, the detection rate for card testing jumped from 59% to 97% when combining traditional models with the foundation model's insights. This ability to catch previously undetectable fraud not only protects businesses but also indirectly boosts revenue by reducing false positives that might block legitimate customers.

Beyond fraud, Stripe uses AI to drive revenue for businesses through its optimized checkout suite. This suite personalizes the checkout experience for customers, dynamically surfacing the most relevant payment methods based on their location and preferences. Businesses using this feature have seen a 12% increase in revenue and a 7% lift in conversion rates.

Another example is Smart Disputes, an AI-powered tool that helps businesses fight chargebacks. By analyzing the likelihood of success, the system automatically gathers evidence and files responses, recovering revenue that might otherwise be lost. This is especially helpful for small businesses that lack the resources to manually dispute every chargeback.

The Rise of Agentic Commerce

Agentic commerce, where AI agents buy and sell on behalf of humans, is no longer a futuristic concept. Stripe is actively enabling this shift, recognizing that AI is moving from simply answering questions to performing actions. This has two main implications:

1. AI Agents as Buyers: AI agents are now capable of making purchases. For instance, a "barista agent" can scour the internet for coffee beans and buy them for a user. These transactions often use single-use virtual cards, similar to how human agents (like DoorDash drivers) operate, ensuring controlled and secure money movement.

2. In-Situ Commerce: Buying and selling are increasingly happening within AI tools themselves, rather than requiring users to navigate to external websites. Perplexity, for example, allows users to discover and book hotels directly within its app, powered by Stripe.

This shift demands changes in how online commerce operates. Intent, rather than clicks, will become the primary interface, with AI agents declaring their desires in structured formats. Product data will need to be machine-readable, and latency budgets will shrink to machine time, requiring extremely fast transaction processing. Furthermore, the risk landscape will evolve, necessitating new ways to distinguish between good and bad bots and to manage programmable trust through scoped, one-time tokens.

The New AI Economy: Faster Growth, Global Reach

Stripe's unique vantage point provides insights into the rapid growth of the current generation of AI startups. These companies are monetizing significantly faster than previous generations of startups. For example, AI companies hitting $30 million in annualized revenue reached that milestone in about 1.5 years, compared to 5.5 years for the fastest-growing SaaS startups a decade ago.

This rapid growth is also accompanied by an accelerated global reach. AI companies are internationalizing twice as fast as earlier SaaS companies, with more than half of their revenue often coming from overseas customers within their first year. This is partly due to the borderless nature of digital products and the translation capabilities of LLMs, but also because the tools for global expansion, like Stripe's optimized checkout suite and tax solutions, have become more accessible.

Key Trends in the AI Economy:

• Faster Monetization: AI startups are reaching revenue milestones (e.g., $30 million ARR) in a fraction of the time it took for previous tech cycles.

• Global from Day One: AI companies are expanding internationally much earlier, with a significant portion of their revenue coming from overseas customers.

• Lean Teams: These businesses are building substantial value with very small teams (10-30 people), a trend not seen before.

• Shift to Usage-Based Pricing: Due to meaningful inference costs, AI companies are moving away from per-seat billing to usage-based models. Some are even experimenting with outcome-based pricing, charging only for successful results, which aligns incentives with customers and reduces risk.

Fostering AI Literacy at Stripe

Stripe promotes AI literacy internally through a culture of experimentation. They started by providing employees with an internal LLM Explorer, a ChatGPT-like interface, to encourage safe experimentation with AI capabilities. This approach, combined with decoupling from any single model and enabling collaboration through sharable prompts, has allowed Stripe employees to quickly integrate AI into their workflows. The company also provides engineers with tools like an LLM proxy and an agent builder to create production-grade AI applications and agents that can interact with various internal and external systems.