In October 2024, OpenAI and Stripe announced the Agentic Commerce Protocol (ACP), an open-source standard enabling AI agents to discover products and complete purchases within conversational interfaces. The partnership represents the first major infrastructure solution for a category that has existed theoretically but lacked technical standardization: commerce where AI agents, not humans, execute transactions.

The Agentic Commerce Protocol defines how AI agents interact with merchant systems to complete purchases programmatically. It standardizes product discovery, checkout negotiation, and payment processing so any compliant agent can transact with any compliant merchant without custom integration work. For the e-commerce industry, this creates a new transaction channel where conversations, not web browsers, mediate purchases.

The significance lies in scale and timing. ChatGPT serves 300 million weekly users. Shopify enabled over one million merchants automatically through platform-level integration. The protocol launched with production implementations, not pilot programs. Organizations operating e-commerce infrastructure now face questions about API readiness, real-time data requirements, and measurement gaps when transactions happen outside traditional web analytics.

This article explains what ACP is, why OpenAI and Stripe built it together, how the technical architecture works, what's currently live, and which infrastructure challenges emerge when agents mediate commerce.

What is the Agentic Commerce Protocol?

Traditional e-commerce architecture assumes humans navigate websites. Systems render HTML pages, track clicks, manage browser sessions, and process form submissions. AI agents can't operate this way. They need structured data feeds and programmatic APIs. Without a standard protocol, every agent-merchant connection requires custom integration: OpenAI building unique implementations for thousands of merchants, each merchant creating separate systems for multiple AI platforms.

This fragmentation would prevent agentic commerce from scaling. The protocol solves this by creating a standardized framework any compliant participant can implement once and use with all other compliant participants.

ACP's solution

The protocol provides three core capabilities: machine-readable product discovery through structured feeds, conversational checkout flows with stateful negotiation, and secure payment handling using delegate tokens. It's open-source under Apache 2.0 license, meaning any company can adopt it without licensing fees or proprietary restrictions.

The architecture shifts integration complexity from N×M (every agent needing custom work for every merchant) to N+M (each agent implements ACP once, each merchant implements ACP once, and they interoperate).

Key design principles

The protocol includes specific constraints reflecting both technical requirements and business model considerations:

Merchants maintain direct customer relationships. The agent facilitates transactions, but customers purchase from merchants, not from OpenAI. This preserves brand relationships and customer data ownership.

No intermediary payment capture. Money flows directly from customer payment methods to merchant accounts. The agent platform doesn't take custody of funds, reducing regulatory complexity and keeping it in a facilitation role.

Merchant control over order acceptance. Merchant systems validate transactions, check fraud, calculate taxes, and decide whether to fulfill orders. Agents request permission; merchants approve or decline based on their own logic.

Conversational optimization. The checkout flow is stateful and iterative. Agents can negotiate with merchants (changing shipping, updating quantities) before finalizing payment.

What ACP intentionally doesn't address: pre-transaction discovery behavior, attribution data, and measurement infrastructure. It defines transaction completion, not customer journey tracking.

Why OpenAI and Stripe partnered

ChatGPT reached 300 million weekly users by mid-2024, with users already requesting product recommendations and comparison shopping advice. Completing transactions within conversations was the natural progression. But OpenAI isn't a payment processor, lacks merchant relationships, and doesn't operate as a retailer. Building commerce required partnering with infrastructure handling billions in transactions, maintaining merchant relationships at scale, and solving payment security challenges.

Stripe processes hundreds of billions of dollars annually for millions of businesses. They have merchant relationships, payment infrastructure, fraud detection, and regulatory compliance already built. Their systems are API-first by design, making them adaptable to agent-driven transactions without fundamental redesign.

Stripe's relationship with Shopify provided strategic value. Shopify uses Stripe for payment processing, giving immediate access to over one million merchants. This solved the cold-start problem where ACP needs both agents and merchants to have value. Shopify's auto-enablement created instant merchant supply.

The open-source approach prevents lock-in and encourages adoption. Merchants implement standards they know won't trap them in proprietary ecosystems. AI platforms beyond OpenAI can adopt without licensing negotiations. Network effects accelerate as each additional participant increases value for existing ones.

How the technical architecture works

Merchants provide product catalogs in structured formats (CSV, JSON, XML, TSV) containing product names, descriptions, SKUs, prices, inventory levels, images, and shipping options. Feeds update every 15 minutes for current pricing and availability. ChatGPT indexes these products for answering availability and pricing questions.

When users express purchase intent, a three-request sequence occurs:

CreateCheckoutRequest: ChatGPT sends the merchant API an initial request with shipping address and selected product. Merchants validate addresses, check inventory, calculate sales tax, and return shipping options with costs.

UpdateCheckoutRequest: If customers change orders (faster shipping, adjusted quantities), ChatGPT sends updates. Merchants recalculate totals. This iterates until customers commit to payment.

CompleteCheckoutRequest: When customers confirm, ChatGPT sends a final request including a Shared Payment Token. Merchants create payment intents with Stripe, process charges, and return confirmations.

Merchant systems run standard business logic throughout: fraud checks, inventory verification, tax calculation, order validation. Merchants control approval or rejection.

Payment security model

Customers save payment methods in ChatGPT (cards stored securely by Stripe). When authorizing purchases, ChatGPT doesn't send raw card data. Stripe creates Shared Payment Tokens: limited-scope delegate tokens restricted by amount, merchant, and expiration time.

Merchants use tokens to create Stripe PaymentIntents for exact amounts. Attempts to charge different amounts or use expired tokens fail. Post-purchase, merchants send order status updates to webhook endpoints ChatGPT exposes.

Traditional e-commerce redirects customers to merchant-hosted payment pages. ACP checkout happens entirely within ChatGPT's interface without redirects, merchant-hosted forms, or separate confirmation pages. Stateful negotiation occurs through API calls. This creates faster, more conversational experiences but removes traditional measurement touchpoints. Merchants don't see how long customers spent comparing options or what alternatives they considered.

Current implementation status

OpenAI launched Instant Checkout in October 2024 for U.S. ChatGPT users (Plus, Pro, Free tiers). Over one million Shopify merchants were enabled automatically, including Glossier, SKIMS, Spanx, and Vuori. Etsy sellers are available. The implementation supports single-item purchases, with multi-item carts on the roadmap.

Shopify merchants were enabled through platform-level integration. Merchants on other platforms need to implement ACP directly or wait for their provider to add support. OpenAI reviews merchants for quality and policy compliance before making products available, creating curated rather than open access.

Technical limitations include real-time API requirements. Merchants whose systems can't respond in milliseconds struggle. Inventory systems updating on slow batch schedules create availability mismatches. Geographic availability is U.S.-only at launch, with international expansion requiring navigation of different payment regulations and tax systems.

Infrastructure implications for merchants

Technical requirements

Implementing ACP requires APIs serving product feeds in required formats with 15-minute update frequency. Systems must handle CreateCheckoutRequest, UpdateCheckoutRequest, and CompleteCheckoutRequest with fast response times.

Real-time inventory and pricing become non-negotiable. Batch job updates showing availability for unfulfillable products create problems. Fraud detection runs server-side during checkout request handling without client-side signals like device fingerprints or browsing behavior. Address validation, velocity checks, and risk assessment use structured API request data.

Webhook infrastructure handles post-purchase updates, sending order status changes to ChatGPT-exposed endpoints for customer notifications.

Operational shifts

Agentic commerce becomes a channel alongside websites, mobile apps, and physical retail. Inventory, pricing, and promotions require management across surfaces with different customer behaviors.

Brand visibility changes in conversational contexts. Recommendations show text descriptions, not product imagery, brand stories, or marketing copy. Competition occurs on specifications, price, reviews, and availability rather than visual presentation.

Direct customer relationships persist. Customers buy from merchants, not ChatGPT. Merchants handle returns, refunds, and support, preserving business relationships despite transactions happening in external interfaces.

Data and measurement challenges

Traditional e-commerce provides rich analytics: page views, time on site, products compared, cart abandonment rates, checkout funnel drop-offs. These metrics inform pricing, inventory planning, and marketing optimization.

Agentic commerce removes these touchpoints. Merchants see completed transactions but not the discovery and consideration processes leading to them. Questions customers asked ChatGPT, alternatives ChatGPT suggested, competitor price comparisons: all happen in conversations invisible to merchant systems.

Event-level compliance becomes challenging when agents mediate transactions. Traditional e-commerce captures consent through customer-clicked checkboxes during checkout, creating explicit records. In agentic commerce, agents complete checkout on customers' behalf. Proving customers consented to marketing or verifying which privacy regulations apply to traveling customers creates complexity.

First-mile data collection requires new approaches when discovery happens outside merchant control. Systems need to capture intent signals from conversational contexts: what customers asked for, which criteria mattered, what reasoning led to purchases. Data needs consistent structure across ChatGPT, other agent platforms, and traditional channels.

Attribution complexity increases without clear paths tracing sales to marketing efforts. Did search ads influence ChatGPT's training data? Did website product reviews affect agent recommendation logic? Budget decisions rely on correlation rather than attribution.

Identity resolution across sessions becomes harder without cookies. Recognizing that customers purchasing through ChatGPT today browsed websites last week requires approaches working without browser-based tracking while respecting privacy regulations.

Agentic commerce and infrastructure evolution

The OpenAI-Stripe partnership represents infrastructure-level change in how transactions occur, shifting commerce from requiring human navigation through visual interfaces to supporting programmatic transactions through conversational ones.

For merchants, this creates opportunity and challenge. The opportunity: access to 300 million weekly ChatGPT users purchasing products without leaving conversations. The challenge: systems supporting real-time APIs, data infrastructure adapting to missing measurement touchpoints, operations handling channels where brand visibility works differently.

The protocol solves transaction mechanics. What remains unsolved: measurement, attribution, and compliance challenges emerging from agent-mediated commerce. These require separate infrastructure approaches.

‍