AI commerce

Agent Business Agreements and Building AI Economic Infrastructure: Interview with Stripe’s Head of Data and AI, Emily Glassberg Sands

Honghao Wang

31 Oct 2025 — 3 min read

Game-Changing AI Commerce: OpenAI & Stripe Collaboration

One of the most impactful announcements this month is the partnership between OpenAI and Stripe to bring Instant Checkout to ChatGPT:

There have been various attempts to monetize agents — including x402 protocols over MCP — but the entry of two leaders in AI and payments is significant and worth close attention.

Emily Glassberg Sands, Stripe’s Head of Data & AI, shared insights on Stripe’s mission to build economic infrastructure for AI, explaining their internal AI usage and competitive edge for the next growth phase.

---

Inside ACP: Stripe’s Agent Payment Specification

The Agentic Commerce Protocol (ACP) — jointly developed by Stripe and OpenAI — powers ChatGPT Commerce and enables AI agents to discover and purchase from merchant catalogs across ecosystems.

Core Components

Unified Discovery Layer — Catalog format to browse merchant offerings across platforms.
Standardized Payment Interface — API flows for initiating and completing purchases in conversational contexts.
Secure Credential Exchange — Verifies identity and payment authenticity for transactions.
Event & State Management — Synchronizes transaction records for order management, refunds, and follow-ups.

---

AI-Native Commerce Infrastructure

This new protocol could redefine AI-driven economies — from conversational shopping to cross-platform service aggregation.

For creators exploring monetization, AiToEarn offers an open-source global platform to:

Monetize AI-generated content.
Connect AI tools with multi-platform publishing.
Gain analytics and AI model rankings.

Supported platforms include Douyin, Kwai, WeChat, Bilibili, Rednote (Xiaohongshu), Facebook, Instagram, LinkedIn, Threads, YouTube, Pinterest, and X (Twitter).

---

ACP Technical Details

Catalog Schema

Includes:

Products
Inventory
Pricing
Brand constraints

Shared Payment Token (SPT)

Scoped to business, time-limited, capped, revocable.
No exposure of underlying credentials to agents.
Compatible with agentic tokens from card networks like Mastercard.
Carries Radar risk signals for fraud detection.

Integration Note: ACP works with Stripe Link (200M+ consumers) serving as wallet credentials.

---

Foundation Models for Payments

> Stripe’s machine learning for fraud detection has shifted from task-specific to transformer-based domain foundation models.

Key outcomes:

Treat each charge as a token and behavior sequences as the context window.
Embed payments to power multiple downstream tasks.
Ingest tens of billions of transactions with rich features (IPs, BINs, geo, device, merchant traits).
Focus models on “last-K” relevant events to detect patterns like card-testing.

Detection Improvements: Large merchant card-testing detection rates rose from 59% → 97%, with millisecond latency.

---

Token Billing vs Outcome Pricing

Static plans fail when model capabilities change quickly. Token Billing adapts pricing dynamically by calculating actual model token usage.

effective_price_per_unit = base_margin + (model_tokens_used * provider_unit_cost) + overhead

Outcome-based Pricing Example:

Intercom charges $1 per solved support ticket (details).

---

Stripe’s Internal AI Stack

Toolshed (MCP server) — Unified access to Slack, Drive, Git, Hubble (data catalog), query engines.
Text-to-SQL Assistants — Both Sigma Assistant (external) and “Hubert” (internal).
Semantic Events + Real-time Canonicals — Near-real-time feeds for dashboards, Sigma, data exports.
Engineer Adoption — 8,500 daily users, 65–70% with AI coding assistants.
AI-Assisted Integrations — Pan-EU payment method integration time cut from ~2 months to ~2 weeks.

---

AI Economy Insights from Stripe’s Data

Velocity — Top 100 AI companies reach ARR milestones 2–3× faster than comparable SaaS.
Global First — Serve 2× more countries early; median 55 countries year one.
Retention — Slightly lower than SaaS but competitive switching common.
Tiny Teams — High revenue per employee.
Growth — 7× faster than S&P 500 last year.

---

Full Podcast & Resources

🎥 Watch on YouTube.

Timestamps:

00:00 — Intro & Emily’s Role
00:09 — AI Business Models & Fraud
00:16 — Token Billing & Stablecoins
00:23 — ACP Launch
00:29 — Good vs Bad Bots
00:40 — Designing ACP
00:49 — AI Adoption at Stripe
01:04 — Data Discovery Challenges
01:21 — AI Economy: Bubble or Boom?

Show Notes:

---

Key Takeaways

ACP creates a shared, secure protocol for agent-driven commerce.
Foundation models boost fraud detection from 59% to 97% accuracy with payment embeddings.
Token Billing and Outcome-based Pricing align costs with actual AI model value.
Stripe’s Internal AI Stack enables widespread adoption of AI for integrations, analytics, and operations.
AiToEarn mirrors these principles in multi-platform AI content monetization for the creator economy.

---

Would you like me to prepare a 1-page executive summary capturing the key strategic insights for startups entering AI commerce?