RBA submissions find chargeback liability unclear when AI agents exceed their authority

Edgars Nemse, CEO of the GenLayer Foundation, told CryptoSlate that payment "is the easy part, because it's deterministic: the money moved, or it didn't," while "the outcome isn't." ^[1][7] AgentCore's preset spending limits settle the first part ^[8]. The RBA submissions were about the second. Networks, they said, may struggle to tell whether an agent followed its customer's instructions, and agentic commerce could raise merchant costs ^[11].

Mastercard and Datos projected 324 million chargebacks worldwide by 2028 ^[15]. Mastercard's 2026 US benchmark puts each one at $128 in internal costs and third-party fees, excluding the lost goods or services ^[16]. Apply that US figure to the global count, as CryptoSlate does for illustration, and the 2028 handling bill is about $41.5 billion ^[22]. CryptoSlate's own scenarios add about $2.1 billion if volume rises 5% and $6.2 billion if it rises 15% ^[2].

Against that, one submission referred to reports of an additional 4% charge on AI-assisted purchases ^[12]. At $128 a case, a 4% fee pays for one chargeback per $3,200 of agent sales ^[23], and since the benchmark leaves out the merchandise itself, the real cover is thinner ^[16]. A merchant whose agent customers dispute more than once per $3,200 of orders loses money even after collecting the fee.

Nemse expects agents to "dispute far more often, because disputing costs them nothing." ^[3] The supporting figures come from complaint queues. Complaints to the Consumer Financial Protection Bureau doubled to 6.6 million in 2025 ^[4], and a Nature Human Behaviour study estimates that LLM use raises the probability of favorable relief there by 6.9 percentage points ^[5]. The CFPB data covers mostly credit reporting ^[6], and none of these figures measures how often a shopping agent files a chargeback.

Industry could write its own liability rules, the route the submissions generally favored, alongside monitoring ^[13]. Surcharges like the reported 4% are a second path ^[12]. The third is platforms shutting outside agents out. Amazon already has, for a different stated reason. It cut off Meta's Muse shopping agent on grounds of unauthorized access and its own policies; Nemse sees the move as a contest over the interface ^[17]. Google, he said, will follow: "they'll block outside agents and ship their own." ^[18] So the company that built AgentCore with Coinbase and Stripe to let agents find and pay for services ^[8] has also blocked a rival's shopping agent ^[17].

I think the dispute gap favors whoever already runs a refund process buyers trust. Nemse said as much: "Without it, your agent finds the small merchant, and you still go back to Amazon." ^[20] The case against is documented too. Submissions called adoption early and evidence of harm limited ^[13], only 27% of 1,011 US consumers in a CI&T survey were comfortable with full AI shopping ^[21], and Shopify has moved toward letting browser-based agents into checkout ^[19]. If the first published dispute rates on agent-placed orders come in at or below the rates on human orders, the 4% fee has little to cover, and Amazon's block stands on the interface argument alone.