Conversational Commerce has been front and center in our work at Opus Research since Dan Miller coined the term in 2011. With the advent of Large Language Models (LLMs), the concept and underlying tech have both come a long way. I’ve written previously about how Conversational Commerce has morphed into “agentic shopping,” and I examined two recent retailer-owned AI shopping assistants: Walmart’s Sparky and Williams-Sonoma’s AI culinary companion. The agentic shopping pitch was simple: brands that know their customers can suggest more relevant products and deepen loyalty. But that center of gravity just shifted. 

With “Buy it in ChatGPT,” OpenAI has turned ChatGPT into a native checkout surface. They’ve open-sourced an Agentic Commerce Protocol (ACP) that any merchant can implement. The logic is straightforward: ChatGPT already knows what customers want because they’re asking it questions all day. ACP lets ChatGPT work with merchant systems to turn conversation into completed orders. 

Customers discover products in chat, tap Buy, confirm details, and the merchant fulfills using their existing stack. They remain merchant of record. Payments move through the merchant’s provider, with Stripe’s Shared Payment Token as the first supported option. Instant Checkout started with U.S. Etsy sellers and is coming soon to over a million Shopify merchants. 

How ACP works 

ACP defines a contract between three parties: the buyer and their agent (ChatGPT), the merchant, and the payment provider. 

Merchants expose a product feed to OpenAI so ChatGPT can surface accurate SKUs with pricing, availability, and fulfillment options. ChatGPT renders the checkout UI, gathers shipping and payment details, and calls merchant-hosted endpoints to create or update a checkout session. The merchant validates the order, runs risk checks, calculates tax and shipping, and decides whether to accept. 

For payments, OpenAI doesn’t become merchant of record. Instead, it passes a one-time payment credential to the merchant’s Payment Service Provider (PSP) through ACP’s “Delegated Payment” spec. The PSP returns a token the merchant uses to charge as usual. Stripe’s Shared Payment Token is the first compliant implementation. State lives on the merchant’s systems, while ChatGPT handles presentation and orchestration. 

Stripe’s write-up adds that ACP is Apache-licensed, designed to avoid bespoke integrations with each agent, and aims to support more complex flows like subscriptions over time. Merchants keep control of branding, fraud models, and fulfillment, while AI agents never see raw payment credentials. 

Enter Google’s AP2 

Two weeks before OpenAI’s announcement, Google introduced the Agent Payments Protocol (AP2). While ACP handles commerce and checkout, AP2 focuses specifically on payment integrity when an AI initiates a charge. 

AP2’s answer is a chain of cryptographically signed “mandates”: an Intent Mandate capturing what the user authorized, a Cart Mandate locking items and price, and a Payment Mandate flagging agent involvement to networks and issuers. It’s payment-method agnostic, with a roadmap spanning cards, real-time bank transfers, and crypto. Google positions it as standards-first, aimed at issuers and networks as much as merchants. 

Why two protocols? 

They emphasize different layers. ACP is a pragmatic commerce integration that lets an AI agent and merchant coordinate through existing PSPs. It’s tightly scoped to make ChatGPT a safe point-of-sale without re-platforming. 

AP2 is a payments-integrity layer standardizing proof of user intent across channels and payment types. Its mandate credentials are designed to be inspected by PSPs, networks, and issuers, which makes it as much about liability and dispute resolution as initiating charges. 

Put simply: ACP answers “how does an AI check out on my site today,” while AP2 answers “how do all parties trust agent-initiated payments across the ecosystem tomorrow.” 

They can coexist. An ACP checkout could eventually attach AP2-style mandates when needed. But they could also compete for mindshare. If merchants standardize on ACP because demand shows up in ChatGPT first, AP2 will need concrete network adoption to become essential. If networks adopt AP2 mandates as a risk signal, ACP implementers may add AP2 support for better approval rates. 

What this means for retailers 

The control point is moving from search boxes and brand websites to persistent personal agents. OpenAI’s product makes this tangible by turning conversation into completed orders without leaving chat, while letting merchants keep the customer relationship and existing payment stack. 

The trade-off is clear: merchants gain a high-intent channel but must accept that discovery, shortlisting, and checkout UX may live in a third-party agent. 

The near-term to-do list is straightforward: 

• Stand up a high-quality product feed and ACP endpoints so your catalog can be discovered inside agents.  
• Tighten risk controls and fulfillment SLAs, as agent-led purchases will compress decision cycles and raise expectations for speed and accuracy.  
• Evaluate PSP readiness for delegated payments and watch AP2’s mandate model as issuers recognize agent presence as a distinct signal. 

The bigger shift 

Amazon’s advantage has been aggregation, logistics, and habit. Agentic commerce challenges the habit part. If buyers increasingly ask their AI to find, compare, and buy within a trusted conversational surface, loyalty shifts toward the agent that best understands their preferences and context. 

OpenAI’s move is a bid to become the operating system for shopping. Google’s AP2 is a bid to ensure payment trust scales across rails and providers. Neither replaces Amazon’s logistics, but both reduce the friction of jumping between tabs. 

Over the next buying cycles, expect three outcomes:  

• More merchants selling inside agents without ceding merchant of record. 
• Payment networks creating separate fraud models and approval rules for AI-initiated transactions.
• A new attribution fight over who gets credit; the model, the marketplace, or the merchant. 

The winners will be retailers who show up where the agent lives, keep their data clean, and deliver when the agent says “buy.” And what works for shopping may soon apply to every industry where customers delegate decisions to their AI. 

Categories: Articles