When the customer is an AI: building platforms for AI agents

Key takeaways

• AI agents increasingly search, compare and buy on behalf of people. For them to use a platform reliably, it needs well-described interfaces and structured, current offer data.
• Open protocols define the common language: the Agentic Commerce Protocol (ACP) by OpenAI and Stripe, the Universal Commerce Protocol (UCP) by Google, and the Agent Payments Protocol (AP2) for payment. With ACP and UCP, the merchant remains the seller.
• Price and availability have to be unambiguous and fresh. According to Stripe, updating inventory and prices for selling through agents every 15 minutes is enough in most cases.
• The human decides before the purchase. The protocols include dedicated confirmation steps, and a good platform only passes on real, linkable offers.

As of October 2026

A customer asks her AI assistant for a rain jacket for a week of hiking: waterproof, no more than €150, delivered by Friday. The assistant searches catalogs, compares prices and delivery times and presents three suggestions, she picks one and confirms the purchase. In this case, the first visitor to your platform was a program. This guide shows what a platform needs so that AI agents can use it reliably, and which open standards exist for this.

What an AI agent does when shopping

An AI agent is an assistant that completes tasks in several steps on its own: it searches, compares, asks follow-up questions and prepares decisions. In shopping, this is already common: according to OpenAI, more and more people start their product search in ChatGPT. Since March 2026, ChatGPT has shown products side by side with price, reviews and features.

The OpenAI Help Center explains where this information comes from: ChatGPT takes into account structured product data from first-party and third-party providers, such as price and product description. How agents read and operate classic websites is covered in our article Website ready for AI agents. This article is about platforms and shops whose data and functions agents use directly.

Three open protocols for commerce with agents

Open protocols set common rules for how agents and platforms exchange data and orders, so that not every platform needs its own connection to every agent. Three of them shape commerce through agents (agentic commerce), all under the open Apache 2.0 license.

• Agentic Commerce Protocol (ACP): OpenAI and Stripe published it on September 29, 2025. It still has beta status, meaning it is being trialed. The current version of April 17, 2026 covers cart, product feed and orders, among other things.
• Universal Commerce Protocol (UCP): Google introduced it on January 11, 2026, developed with Shopify, Etsy, Wayfair, Target and Walmart, among others. It covers catalog search, cart, account linking, checkout and order management. The current version dates from August 25, 2026.
• Agent Payments Protocol (AP2): Google introduced it on September 16, 2025, working with more than 60 organizations. AP2 defines how an agent proves that the human authorized a specific purchase. Version 0.2 was released in April 2026, and at the same time Google handed the protocol over to the FIDO Alliance, an industry association for open standards.

The protocols complement each other. UCP integrates AP2 as an extension, and Stripe supports both ACP and UCP. With ACP and UCP, the merchant remains the seller (merchant of record) and keeps the relationship with its customers.

In March 2026, OpenAI wrote that the first version of buying directly in ChatGPT (Instant Checkout) had not offered the flexibility it aimed for: since then, merchants can use their own checkout, while OpenAI focuses on product discovery and has extended ACP for it. According to the OpenAI Help Center, ChatGPT may still show an option to buy directly in the chat for some eligible products and merchants.

Interfaces an agent understands

Agents use a platform directly through interfaces (APIs): they query data and hand over orders. What an interface can do therefore has to be described in a machine-readable way. ACP, for example, describes the checkout and the data structures in standard formats that machines read directly (OpenAPI and JSON Schema).

So that an agent knows what a platform can do, merchants using UCP publish a profile at /.well-known/ucp, and with ACP the corresponding document is /.well-known/acp.json. It lists the supported functions, versions and transports. Besides classic web interfaces (REST), both protocols support the Model Context Protocol (MCP), which we explain in our article MCP for software providers.

For a new platform, this means: set up products, prices, inventory and orders as a clear data model from the start. Each protocol then becomes a translation of this model, and another channel can be added without rebuilding the core.

Offer data: complete, structured and current

For ChatGPT, OpenAI describes a product feed, a regularly updated file or interface with identifiers, descriptions, prices, inventory, images and fulfillment options. The recommendation is to deliver the full catalog once a day and send changes via the API throughout the day. Onboarding is currently open to approved partners, and merchants can apply with OpenAI.

Some of OpenAI's recommendations fit any platform:

• Factual descriptions that explain the product concisely and clearly
• Stable identifiers per variant: a fixed ID for the product and a separate one for each purchasable variant, with price, availability, image and link per variant, where they differ
• Valid, durable links to product pages and seller information

Price and availability: unambiguous and fresh

So that a price quoted by an agent is still correct at checkout, Stripe sets out a clear rhythm for selling through agents: product data once a day, inventory and prices every 15 minutes, which is enough in most cases. According to Stripe, agents may keep reporting a product as unavailable until they receive an explicit signal that it is back in stock.

The platform itself decides on the order. With ACP, the merchant checks every order in its own systems, determines the fulfillment options and accepts or declines the order. If a platform uses UCP with the AP2 extension, the merchant digitally signs the state of the checkout so that price and terms do not change along the way. Google describes the goal of AP2 like this: what the customer sees is what she pays.

The human decides before the purchase

The protocols are designed so that the human stays in control. At the launch of Instant Checkout, OpenAI wrote that users explicitly confirm every step before any action is taken. The payment details reach the merchant as a token, a stand-in identifier. According to Stripe, such a token is only valid for a specific merchant, a maximum amount and a limited time.

UCP has a dedicated state in the checkout for this. If the merchant requires a review and authorization by the buyer, for example for a high order value or a first purchase, the agent hands over to the merchant's page via a link. AP2 makes consent verifiable: in a real-time purchase, the human's approval creates a digitally signed record (mandate) of the items and the price. According to Google, version 0.2 added the flows for purchases without the human being present: the agent then only buys according to instructions given in advance, with limits such as a maximum price and timing.

So build the confirmation step firmly into the flow, with a clear summary of item, price, shipping and merchant.

Real offers only: protection against made-up details

OpenAI points out in its Help Center that ChatGPT interprets user intent and can occasionally make mistakes. AP2 names as one of its core questions whether an agent's request truly reflects the user's intent, without errors or details invented by the AI (hallucinations). Good platforms answer this question with their architecture.

On such platforms, every recommendation is based on a real record with an identifier, a current price and a working link. Anything that cannot be clearly matched to such a record is not passed on by the platform.

In practice

With byclaw.io, our own product, we explored one question: what does a shop look like when the customer is an AI agent? The starting point: a good recommendation needs live prices and real catalogs. Our answer is an agent-first shop, an AI shopping agent that searches for products, compares them and makes recommendations via WhatsApp, directly in Claude and ChatGPT via MCP, or on the web.

Behind it is a catalog from affiliate feeds, meaning product lists from merchants' partner programs, supplemented by a web search as a fallback. Claude turns every request into search terms, filters by relevance and picks the best match. Every recommendation has to correspond to a real, linkable product, otherwise the hallucination guard does not pass it on.

Five agent personalities weigh price, quality and taste differently, but what the user wants always takes priority. The result is a recommendation with a direct link to the shop. Whether and where to buy is up to the human.

Frequently asked questions

Is my existing shop system enough for AI agents?
According to OpenAI, Shopify product data reaches ChatGPT through Shopify Catalog, with no additional work for individual merchants. A custom platform with its own workflows, offers or partners connects agents through its own interface and one of the open protocols.

Which protocol should my platform support?
OpenAI uses ACP for ChatGPT, Google uses UCP for buying in AI Mode in Google Search and in the Gemini app. Both protocols are open and build on the same core elements: product, cart, checkout and order. If you map these elements cleanly in your own data model, you can serve both.

Can merchants in Germany take part already?
OpenAI first launched buying directly in the chat for users and merchants in the US, and since March 2026 its focus has been on product discovery. Google is introducing direct purchasing in AI Mode with US merchants first, as described in our article Shopping in Google's AI Mode. According to Stripe, its building blocks for selling through agents are already available to sellers in Germany and other European countries.

Conclusion

AI agents are becoming a customer group of their own: they work with data and present a selection to the human. A platform that serves them well offers described interfaces, complete and current offer data, unambiguous prices and a clear confirmation step. With a clean data model, it can connect to open standards such as ACP, UCP and AP2 step by step.

Should your platform be ready for AI agents?

Describe your project in our AI project estimator. You will receive a free estimate by email, and we will get back to you within 24 hours.

Learn more about our services for platforms and marketplaces and SaaS products. Or simply write to us at our email address.

Sources

1. OpenAI: "Powering Product Discovery in ChatGPT", March 24, 2026. https://openai.com/index/powering-product-discovery-in-chatgpt/
2. OpenAI Help Center: "Shopping with ChatGPT Search". https://help.openai.com/en/articles/11128490-shopping-with-chatgpt-search
3. OpenAI: "Buy it in ChatGPT: Instant Checkout and the Agentic Commerce Protocol", September 29, 2025. https://openai.com/index/buy-it-in-chatgpt/
4. Agentic Commerce Protocol: website. https://www.agenticcommerce.dev/
5. Agentic Commerce Protocol: specification on GitHub with README, maintainers list, changelog for version 2026-04-17 and MCP binding. https://github.com/agentic-commerce-protocol/agentic-commerce-protocol
6. OpenAI for Developers: "Key concepts" (Agentic Commerce). https://developers.openai.com/commerce/guides/key-concepts
7. OpenAI for Developers: "Get Started" and "Best practices" for product feeds. https://developers.openai.com/commerce/guides/get-started, https://developers.openai.com/commerce/guides/best-practices
8. Google: "New tech and tools for retailers to succeed in an agentic shopping era", Google Blog, January 11, 2026. https://blog.google/products/ads-commerce/agentic-commerce-ai-tools-protocol-retailers-platforms/
9. Universal Commerce Protocol: website and "Announcements". https://ucp.dev/, https://ucp.dev/latest/documentation/announcements/
10. Universal Commerce Protocol: release v2026-08-25, August 25, 2026. https://github.com/Universal-Commerce-Protocol/ucp/releases/tag/v2026-08-25
11. Universal Commerce Protocol: specification "Overview" and "Checkout". https://ucp.dev/specification/overview/, https://ucp.dev/latest/specification/shopping/checkout/
12. Universal Commerce Protocol: "UCP and AP2". https://ucp.dev/latest/documentation/ucp-and-ap2/
13. Google Cloud: "Announcing Agent Payments Protocol (AP2)", September 16, 2025. https://cloud.google.com/blog/products/ai-machine-learning/announcing-agents-to-payments-ap2-protocol
14. Google: "Google donates Agent Payments Protocol to FIDO Alliance", Google Blog, April 28, 2026. https://blog.google/products-and-platforms/platforms/google-pay/agent-payments-protocol-fido-alliance/
15. Agent Payments Protocol: documentation and release v0.2.0 on GitHub (April 28, 2026). https://ap2-protocol.org/, https://github.com/google-agentic-commerce/AP2/releases
16. Stripe: "Agentic commerce", documentation. https://docs.stripe.com/agentic-commerce
17. Stripe: "Sell through agents", documentation. https://docs.stripe.com/agentic-commerce/sellers/use-cases/retail
18. Stripe: "Shared payment tokens", documentation. https://docs.stripe.com/agentic-commerce/concepts/shared-payment-tokens