If a merchant buys fraud-risk scores from one company but runs payments through another, who pays when fraud slips through?

This explores who carries the loss when a merchant buys fraud scoring from one company (Stripe) but settles the payment through another. The corpus doesn't cover that exact arrangement. No note looks at contracts for standalone risk scores or at how liability works between processors. What it does show is the principle that answers most of the question: liability follows the Merchant of Record, not the company that gave the risk advice. Does agentic commerce shift fraud liability onto merchants? argues that in every proposed agentic commerce protocol, the merchant keeps the fraud liability. A risk score is something the merchant uses to make a decision, and the merchant still owns that decision.

AI agents make this split sharper. Fraud scoring has always relied on behavioral signals such as typing rhythm, mouse movement and session patterns. When an agent does the checkout, those signals disappear. Boboev's point is that merchants end up carrying more risk while their detection tools see less. So a merchant using an outside risk score on agent-driven traffic may be trusting a model that is quietly missing the inputs it was built on, and the merchant still pays for any mistake.

The protocols try to fill this gap in different ways, and none of them moves liability onto the scorer. Stripe's Shared Payment Tokens are limited to one merchant and one cart total, which caps what a stolen credential can do (Can payment tokens let AI agents checkout safely?). Google's AP2 chains signed 'mandates' together, so there is a tamper-proof record of what the user actually approved (How can signed mandates prevent payment fraud by agents?). That kind of evidence matters most in a dispute because it shows who authorized what. Boboev also describes these protocols as layers a merchant stacks, with each layer taking on a different level of risk (Do agentic commerce protocols compete or complement each other?). In practice, a merchant mixing vendors is building their own stack, and the gaps between layers are theirs too.

An IMF analysis offers a useful lens. It separates probabilistic judgment, such as an AI deciding or a model scoring, from deterministic authorization and settlement, which is where legal enforceability sits (Should AI agents make payment decisions independently?). A risk score belongs to the probabilistic side. It is advice. Liability attaches to the settlement side, which here is the other processor and the merchant's contract with it. Consent adds a further wrinkle. If the user ticked an 'allow always' box, a disputed agent purchase may hinge less on fraud than on what the user actually authorized (Do broad consent settings pose bigger risks than prompt injection attacks?).

The answer the corpus supports is that the merchant almost certainly bears it. If you want the specifics of how Stripe's scoring terms or card-network chargeback rules split losses across vendors, the collection doesn't have that material yet.