If AI agents check out without typing patterns or mouse movements to verify them, who ends up eating the fraud?

This explores what happens to fraud, and to who pays for it, when an AI agent checks out instead of a human and the behavioral clues fraud systems depend on disappear. First, a direct caveat: the corpus has no dollar figures or loss rates for this. No source here measures how much fraud rises when behavioral biometrics go away. What it does show is something many readers won't expect. The main loss isn't a new kind of fraud. It's a change in who absorbs the fraud that already exists, with the blind spot landing on whoever is least able to see.

The sharpest version of this is Does agentic commerce shift fraud liability onto merchants?. When an agent fills in the checkout form, signals like typing cadence and mouse movement no longer tell you whether a real, legitimate person is present. Yet under every proposed agent-commerce protocol, the merchant is still the Merchant of Record, the party legally responsible for the sale. So merchants carry the risk of transactions run by software while losing the detection tools they used to manage that risk. Does agent-initiated commerce break existing payment infrastructure? describes the same gap from the infrastructure side. Payment gateways and fraud engines were built around a person whose identity and intent can be observed, and agent-initiated transactions put pressure on exactly those assumptions. Federal Reserve Governor Waller groups the problem into three barriers: proving the agent is authorized, deciding who pays when something goes wrong, and detecting fraud patterns specific to agents. He treats all three as unsolved obstacles to scaling agent purchases (What blocks autonomous agents from handling purchases at scale?).

The industry's main answer is to replace watching behavior with checking signed proof. Google's AP2 protocol chains cryptographically signed 'Intent' and 'Cart' mandates into a tamper-proof record of what the user actually authorized (How can signed mandates prevent payment fraud by agents?). Note what this shifts. Behavioral biometrics asked 'is this a real human acting normally?' Mandates ask 'did a real human sign off on this?' That helps with disputes after the fact. It doesn't help in real time if a compromised or confused agent carries out a properly signed but bad instruction. Identity is a second gap. Can credential vaulting alone solve merchant-facing agent identification? argues that safely storing the consumer's card details doesn't tell the merchant which agent is at the door or whether it has permission to be there. The Amazon–Meta Muse dispute shows what happens without such a standard: retailers fall back on blocking agents through their terms of service (What identity standards do shopping agents actually need?).

The security research widens this from 'fraud signals vanished' to 'the agent itself becomes a place where attacks can happen.' Red-teaming work finds that an agent's identity often lives in context files that can be manipulated, and that its authorization often rests on the conversation rather than on enforcement at the system level (Why do agents fail at identity verification and authorization?). An attacker therefore doesn't need to imitate a human's typing. Steering the agent is enough. There's also a less obvious cost. One proposed fix, routing agent authentication through large login providers such as Google, would give those providers a cross-service view of what agents are doing (Do OpenID Providers gain surveillance power over AI agents?). Replacing behavioral surveillance at the merchant could create a new kind of surveillance at the identity layer.

So the honest answer is this. The corpus describes the losses as a transfer of risk to merchants plus a new attack surface. It does not yet quantify a fraud-loss number. If you want to see where that risk ends up in a real market, travel is a useful case. OTAs (online travel agencies such as Expedia) are keeping their Merchant of Record role as agents arrive (Are AI agents actually removing online travel agencies from bookings?), which means they're the ones holding this liability.