AI-assisted shopping is no longer a futuristic pipe dream. Consumers can already buy through Perplexity directly in the feed, and while OpenAI’s own checkout experiment was short-lived, the Agentic Commerce Protocol it built with Stripe is very much alive and spreading across the ecosystem, as Forrester’s recent analysis of agentic payments makes clear. For payments leads at merchant companies, the operational question is no longer whether you will one day accept agentic payments. It’s whether you understand the liability position you’re already in from such orders.
The answer depends heavily on which platform your agentic sales are coming through. And right now, the two dominant models couldn’t be more different.
The Merchant of Record Split
When a consumer buys through Perplexity’s Buy With Pro, Perplexity is the Merchant of Record (MOR). Perplexity’s name is on the transaction, not yours. At first read that sounds like a win: someone else absorbs the chargeback liability. But it comes with a significant trade-off. Because you are not the MOR, you don’t control the payment flow or the checkout data, and miss the insights a chargeback provides. This eliminates the beneficial role chargebacks play, i.e. notifying merchants that they have a problem with true, third-party fraud or an issue in customer service. If a chargeback surfaces through the Perplexity leg of the transaction, you may have little visibility into what happened.
OpenAI’s Agentic Commerce Protocol (ACP) takes the opposite position. Under ACP, the merchant remains the MOR. You retain control over payment processing and you bear full liability for disputes. That’s a more workable model for chargeback management in one important sense: you sit inside the dispute process when something goes wrong.
Since OpenAI withdrew Instant Checkout in March 2026, the picture has become even clearer. Merchants using ChatGPT apps handle their own checkout, making their MOR status unambiguous. The convenience of agentic discovery comes with full liability attached. Since ChatGPT agents cannot be used for payments they do not create a chargeback data issue that a service like Perplexity’s Buy With Pro, where some of the evidence you would normally use to fight a chargeback – things like user IP address – are no longer available.
Misleading Data from Agent Assisted Shopping
Shoppers using ChatGPT to locate items for purchase will generate problems for third-party fraud prevention platforms because some of the behavioral signals now associated with fraudster behavior are also characteristic of agents. For example, a user immediately jumping onto the checkout page of a website without previous browsing history and then buying a large amount of a specific item, might normally trigger a high fraud risk score. However, it could just be a consumer referred to a merchant site by ChatGPT or another LLM.
This is very similar to what happened in the early days of cryptocurrency, when certain behavioral characteristics of crypto buyers triggered high risk scores on legacy anti-fraud platforms. For example, buyers’ use of virtual private networks (VPNs) is quite common in the world of crypto but not nearly as much in traditional eCommerce.
Over time new fraud models and new anti-fraud platforms took these behavioral characteristics into account. They will have to do the same with LLMs whether it is used in the search phase of shopping or for actual agentic commerce.
