Agentic Commerce Survived. Instant Checkout Did Not.

The Product Died. The Payment Rail Is Still Coming for You.

On September 29, 2025, OpenAI launched Instant Checkout inside ChatGPT, letting people buy from U.S. Etsy sellers without leaving the chat window, powered by a new open standard called the Agentic Commerce Protocol (OpenAI announcement). The pitch was enormous: more than 700 million people turn to ChatGPT each week, and OpenAI wanted every one of them to be able to complete a purchase inside the conversation.

But here is the part the breathless coverage missed. By approximately March 2026, OpenAI had quietly scaled back the in-chat checkout to a discovery-and-redirect model, after only roughly a dozen Shopify merchants ever shipped against it and sales were near-zero. The consumer-facing product stalled. The infrastructure did not.

That distinction matters more than the product launch ever did. The Agentic Commerce Protocol survived as an active open specification. Stripe, PayPal, and others are still building on it. For any business that touches payments, the real question is not whether AI agents will buy things in a chat window. It is who holds legal liability when the agent gets the order wrong, and which regulated party sits in the payment flow when it does.

What Actually Shipped: A Protocol, Not Just a Feature

The Agentic Commerce Protocol (ACP) is an open standard co-developed by OpenAI and Stripe, released under the Apache 2.0 license on September 29, 2025 and maintained as an open-source project on GitHub (Stripe ACP documentation). It defines how an AI agent, a merchant, and a payment processor exchange product data, place an order, and settle a transaction on a consumer's behalf.

The pieces that are still live

• The specification. ACP remains an active, beta open standard governed jointly by OpenAI and Stripe as Founding Maintainers, with a stated path toward broader community governance.
• The payment partners. On October 28, 2025, PayPal announced it would adopt ACP to let ChatGPT users check out with PayPal and to connect PayPal's network of tens of millions of merchants to the protocol (PayPal press release).
• The direction. Even as the in-chat purchase button retreated to a redirect, the underlying rail continued to attract processors and networks.

The lesson for legal and compliance teams is that infrastructure outlives features. A single consumer product can flop while the standard it introduced becomes the plumbing for an entire category. If you plan around the product, you plan around the wrong thing.

Payment Rail Risk: Who Is Liable When the Agent Buys the Wrong Thing

When a customer clicks buy, the chain of legal responsibility is well settled. When an AI agent clicks buy on the customer's behalf, that chain fractures. This is the core of payment rail risk in agentic commerce, and it is where most businesses are unprepared.

Three liability questions ACP does not answer for you

1. Authorization. Did the consumer actually authorize this specific purchase, or did the agent infer intent from a loose instruction? Under existing card-network and Regulation E frameworks, unauthorized transactions carry chargeback and reimbursement consequences. An agent acting on ambiguous intent is a dispute-generation machine.
2. Error attribution. If the agent orders the wrong item, the wrong quantity, or ships to the wrong address, who eats the cost — the merchant, the processor, the AI platform, or the consumer? ACP moves data between parties; it does not allocate fault among them.
3. The regulated party in the flow. Payment processing is a licensed activity. When an AI platform inserts itself between buyer and merchant, businesses must ask whether the platform, the processor, or the merchant is the party of record for money movement, refunds, and dispute resolution.

The 700-million-user figure is the tell. OpenAI cited more than 700 million weekly ChatGPT users as the market opportunity. That same figure is the scale at which small liability-allocation gaps become systemic disputes. A one-percent agent error rate across hundreds of millions of users is not a rounding error. It is a chargeback flood.

The Regulators Are Already Circling Agentic AI

Agentic commerce is not developing in a regulatory vacuum, and the enforcement posture is sharpening faster than most merchants realize.

Two 2026 developments that bear directly on agentic checkout

• The FTC's AI accuracy policy statement. On July 1, 2026, the FTC published a proposed policy statement addressing AI accuracy and manipulation of AI outputs, with the public comment period closing July 31, 2026 (FTC press release). An agent that recommends and purchases a product is making representations to a consumer. Accuracy is no longer a product-quality concern. It is a Section 5 concern.
• The AI AGENT Act discussion draft. On June 29, 2026, Senator Mark Warner released a discussion draft of the Artificial Intelligence Access, Gatekeeper Exchange, and Nondiscriminatory Transfer Act (DLA Piper analysis). As of that date it had not been formally introduced and carries no bill number, but it signals that Congress is now drafting around agent authentication and authorization directly.

Neither the FTC statement nor the Warner draft targets ChatGPT checkout by name. Both reach the conduct at its heart: an autonomous system acting on a consumer's behalf in a commercial transaction. The message is unmistakable. The rules for agentic commerce are being written now, and businesses that wait for a final rule will be building on sand.

What to Do Before Your Business Sits on an Agentic Rail

Whether or not you integrate ACP tomorrow, the standard is likely to reach your payment flow through a processor or platform partner. Prepare now.

Action items for merchants and platforms

First, map the party of record in every agentic flow. Identify precisely which entity authorizes payment, holds refund liability, and answers a chargeback when an agent transacts. If your contracts do not name that party, your contracts are incomplete.

Second, rewrite your merchant and processor agreements for agent-initiated orders. Standard e-commerce terms assume a human clicked buy. Add explicit provisions allocating liability for agent error, defining what counts as valid authorization, and setting dispute procedures for transactions no human directly confirmed.

Third, treat agent accuracy as a compliance control, not a UX metric. In light of the FTC's July 1, 2026 policy statement, document how product representations reach the consumer through the agent and who is accountable if they are wrong.

Fourth, do not build your legal strategy around a single product. Instant Checkout scaled back within roughly six months of launch. The ACP standard did not. Build around the rail and the liability model, not the interface.

The firms that treat agentic commerce as a payments-and-liability problem rather than a marketing opportunity will be the ones still standing when the rail matures.

Key Takeaways

• The product failed but the rail survived. OpenAI scaled Instant Checkout back to a redirect model by roughly March 2026 after only about a dozen Shopify merchants shipped it, yet the Agentic Commerce Protocol remains an active Apache 2.0 open standard.
• Agentic checkout fractures the liability chain. When an AI agent buys on a consumer's behalf, authorization, error attribution, and the regulated party of record all become open legal questions that ACP does not resolve.
• Scale turns small error rates into systemic disputes. OpenAI cited more than 700 million weekly ChatGPT users; a one-percent agent error rate at that scale is a chargeback flood, not a rounding error.
• Regulators are already drafting. The FTC's July 1, 2026 AI accuracy policy statement (comments close July 31, 2026) and Senator Warner's June 29, 2026 AI AGENT Act discussion draft both reach autonomous commercial conduct directly.
• Contract for the agent, not the human. Merchant and processor agreements written for a human clicking buy leave you exposed the moment an agent transacts on the rail.