The Agents Economy Backbone - with Emily Glassberg Sands, Head of Data & AI at Stripe

Stripe has used machine learning for over a decade, primarily for fraud detection and internal operations. The shift to AI, specifically LLMs, became a focus when GPT-3.5 emerged (late 2022/early 2023). In the last 1.5 years, Stripe realized the need for its own domain-specific foundation models to create richer, denser payment embeddings for various applications.

Traditional ML models struggle with sophisticated fraud like card testing hidden in large transaction volumes. Stripe's foundation model transforms each charge into a dense embedding, allowing the detection of subtle clusters of anomalous behavior in real-time (under 100 milliseconds), uncovering new fraud vectors and suspicious transactions that traditional methods miss.

Friendly fraud includes non-payment abuse like free trial abuse or refund abuse, where legitimate credentials are used but no genuine revenue is generated. For AI companies, this is existentially threatening because GPUs and inference costs are high, making the marginal cost of services substantial, unlike traditional SaaS where marginal costs were near zero.

Stripe provides payments, billing suites, optimized checkout, and fraud protection adapted for AI companies. They've introduced 'token billing' to enable real-time pricing based on fluctuating LLM inference costs and support outcome-based billing, stablecoins, and Link for global reach and efficient transactions, especially given AI companies' rapid global expansion.

The ACP is a shared standard developed with OpenAI that allows businesses to interact with AI agents, enabling agents to make purchases on behalf of users. It standardizes how product catalogs, pricing, and brand appear to agents, and includes a shared payment token for secure transactions, addressing the challenge of 'good bots vs. bad bots.'

Stripe's strategy is 'buy and build,' prioritizing buying existing solutions if they meet needs, then considering building if no suitable option exists. They encourage re-evaluating existing builds against new market offerings and favor solutions that integrate with multiple models and providers, like open-source options, to avoid vendor lock-in.

LLMs can produce enticing output quickly, but the speaker argues that deep thinking is fostered by the arduous process of writing and structuring arguments. While LLMs are great for generative tasks and rapid experimentation, over-reliance can lead to 'LM slop' and hinder the development of intellectual 'depth' if not coupled with critical human oversight and validation.

Hubert is an internal AI tool that converts natural language queries into SQL to ask questions about business data stored in Hubble. While effective for well-structured data, its main challenge is data discovery (identifying the correct tables and fields) in complex, less-documented datasets. Stripe is working on deprecating low-quality tables and improving documentation.

Stripe's data shows AI companies are growing 2-3 times faster than previous SaaS cohorts, are significantly more global from day one, and have strong unit economics. While per-company customer retention might be slightly lower due to a competitive market, customers still value the problem-solving aspect. The speaker suggests that while expectations can be high, the underlying businesses show real value and significant growth potential, especially as inference costs decrease.

AI should make markets and commerce more efficient, expanding consumption and accelerating business creation. While the exact impact on GDP is debatable, increased efficiency in both production and consumption should lead to meaningful GDP growth. The speaker highlights 'tiny teams' (high revenue per employee) as a key indicator of AI-driven efficiency.