The release of OpenAI’s payment system has store owners demanding immediate code updates. Businesses see news about a user base of 1.2 billion weekly users linked here and want to let AI agents purchase items directly. Engineers are now forced to parse new authorization specs and rewrite cart logic for cloud browsers.
Using agentic checkouts today means dealing with system failures, high token costs, and legal liability.
The wallet works only for adult subscribers in the US
Businesses expect a global rush of buyers. However, the documentation limits payments. The feature is open only to adult ChatGPT Pro subscribers in the US with cards from American issuers like Chase, Amex, Mastercard, and Visa linked here. A global release outside America is not announced. Launching in Europe or Asia requires meeting strict frameworks like PSD3, GDPR, and the AI Act, and there is no timeline for such compliance.
The dream of fully autonomous background shopping hits a wall. Every transaction requires biometric verification via Face ID or Touch ID linked here. The model cannot spend money on its own. The absence of micro-transaction limits without confirmation makes the process feel like a standard mobile payment within the chat interface.
User demand is also low. Surveys show only 23% of Americans are willing to trust AI with financial operations linked here. A developer might spend weeks supporting this protocol for a fraction of a niche segment of American subscribers.
Agentic checkouts fail in 42.8% of cases
For now, OpenAI charges no transaction fees linked here. This looks attractive compared to standard processing, but the main cost is the LLM’s inferencing price.
Multi-step site navigation requires massive compute resources. A script typically sends one POST request with parameters. An AI works differently. It loads full HTML, takes screenshots, analyzes them with computer vision, and attempts to fill out forms. Every action sends the full session history into the context. Each button click triggers a new API request.
Consider the inferencing costs. An average store page contains many nested elements. When converted to text, the DOM tree takes up about 50,000 tokens. An agent needs at least four steps for a successful order: finding the item, selecting the size, going to the cart, and entering a shipping address. This totals 200,000 input tokens per session. Complex visual checks require flagship models. The cost of ChatGPT API calls with that context volume is about one dollar per user. If a client buys a ten-dollar phone case, traditional processing takes about 30 cents. Paying for tokens costs three times more than the bank. You can optimize costs by using model routing, sending simple text tasks to cheaper versions. You cannot fully abandon heavy models, as they are needed for captchas and complex UI.
The price tag is worsened by high failure rates. Visa tests of 22 AI models in real shopping scenarios showed a failure rate of 42.8% linked here. Almost half of all sessions break because the agent cannot pass a captcha, find a button, or enter a zip code correctly. The merchant pays for the server and tokens but gets no conversion. Out of a thousand attempts, 428 will fail, burning budget for nothing.
Merchants carry full legal liability for model errors
Agent behavior is unpredictable. The model might add the wrong size, mess up the color, or send an order to an address from an old chat. This creates a compensation issue during hallucinations. A human buyer checks details before paying. An agent sends the data directly to the payment gateway.
The service terms shift liability away from the AI developers. OpenAI states it is not a bank, lender, issuer, or merchant linked here. Dispute resolution, order cancellations, and refunds fall on the merchant and the issuing bank.
The store is left alone with the unhappy client. Automated ways to dispute AI errors do not exist. Each wrong purchase becomes a chargeback. The client calls the bank and disputes the transaction. The bank pulls money from the store's account. High chargeback rates lead to merchant account fines and the risk of losing the ability to accept cards. Adopting this wallet transfers the burden of immature models onto the retailer.
Antifraud systems block agents as hostile bots
Technical integration faces security screens. Major platforms like Amazon, United Airlines, and eBay actively block agentic traffic linked here. Protection systems analyze IP addresses, browser headers, and click speeds. Cloud servers used by models get the highest risk scores. The firewall sees useful bots as hostile scrapers. The session dies at the cart-addition stage.
OpenAI published a Commerce Partner Playbook suggesting sites allow cloud browsers and tune antifraud systems to permit these purchases linked here. In practice, this requires rewriting server-side security rules. Engineers must create exceptions for OpenAI IP pools.
Verifying requests is done via the Web Bot Auth spec. It uses the HTTP Message Signatures standard with Ed25519 signatures linked here. A developer must implement complex header verification and key validation. The server must extract the auth header, download the public key from OpenAI, calculate the hash of the body, and verify the signature. Any error here leaves the agent without access.
OpenAI is partnering with providers like Forter, Riskified, Sardine, and Signifyd to help scoring algorithms distinguish legitimate agents from attacks linked here. Until this is finished, most requests will be rejected by network infrastructure.
How cart handoff works via MCP
Integration is built around the Model Context Protocol. Documentation suggests using tools for the catalog, cart, and a parameter to send an incomplete checkout to the user’s browser linked here.
Direct interaction is still in development. The model gathers items, then a human completes payment in a separate window. This requires middle-layer logic.
import json from openai import OpenAI client = OpenAI(api_key="YOUR_API_KEY") tools = [{ "type": "function", "function": { "name": "checkout_init", "description": "Initialize checkout and return continue_url", "parameters": { "type": "object", "properties": { "cart_id": {"type": "string"}, "items": { "type": "array", "items": {"type": "string"} } }, "required": ["cart_id", "items"] } } }] response = client.chat.completions.create( model="gpt-4o", messages=[ {"role": "user", "content": "Buy two bags of coffee from my cart 12345"} ], tools=tools ) print(response.choices[0].message.tool_calls[0].function.arguments)
The model returns a function call. Your server accepts this, generates a payment link, and returns it to the chat. The agent shows the link as a button.
{ "cart_id": "12345", "continue_url": "https://store.com/checkout?token=xyz_secure" }
Creating an endpoint for this takes time. You need to manage session state between chat and browser, handle abandoned carts, and sync inventory. Models often fail when generating complex JSON structures linked here. A missing comma or wrong variable type breaks the parser. The user clicks the link and finds a blank page. You end up writing complex systems to recover cart states using URL tokens.
Risk of losing the client relationship
The product is a partnership with American Express, Mastercard, and Visa linked here. They use specialized agent-tokenization networks. Each transaction generates a dynamic network token with its own number, CVV, and expiration date linked here.
Real card data is hidden from the store. You get paid, but you lose contact with the buyer. Adapting to this wallet turns the store into a commodity backend. OpenAI controls the interface, purchase history, and user preferences. You lose the ability to upsell, collect emails, or invite users to loyalty programs.
What to do next
Supporting agentic payments is justified only for a narrow business segment. Consider this only if four conditions are met.
- You operate only in the US with American card issuers.
- Your processor supports dynamic network tokens from Visa, Mastercard, or Amex.
- Your security team can implement Ed25519 signature checks using RFC 9421.
- Your backend is ready to accept carts via
continue_urland render them for the user.
Missing even one of these leads to rising server costs and support complaints. Until protocols stabilize, focus on optimizing your standard web checkout.