AI agents can pay for your shopping. Who gets your money back?

Amazon's Bedrock AgentCore Payments, built with Coinbase and Stripe, lets AI agents discover paid services, authenticate, and pay with stablecoins and x402 under preset spending limits.

AI agents are getting wallets before merchants get a court, and shopping agents have started placing orders across the open internet.

The harder problem arrives once the payment clears, when a buyer's agent pays correctly, and the buyer wants the money back. The Reserve Bank of Australia (RBA) put that gap on the record Oct. 6, and Edgars Nemse, CEO of the GenLayer Foundation, argued it caps what agents can buy.

Nemse told CryptoSlate that payment “is the easy part, because it's deterministic: the money moved, or it didn't,” while “the outcome isn't.” Defining if work was delivered as promised is a judgment call.

Transaction stage	What agents can increasingly do	What remains unresolved	Why it matters
Discovery	Find merchants, APIs, content, services	Whether merchants trust unknown agents	Controls who gets distribution
Authorization	Use mandates, credentials, spending limits	Whether the agent stayed within user intent	Determines who bears liability
Payment	Pay with cards, stablecoins, x402 or wallets	Payment success does not prove satisfaction	Settlement is deterministic
Fulfillment	Receive goods, services or API access	Was the outcome delivered as promised?	Requires judgment
Dispute	Submit complaints or refund requests	Who adjudicates outside a platform?	Determines whether open commerce can scale

AI agents remove the friction disputes depend on

According to Nemse, every dispute system runs on a hidden assumption that disputing is tedious enough that most people skip it. AI is already eroding that friction, with people filing the complaints themselves for now.

Complaints to the Consumer Financial Protection Bureau doubled to 6.6 million in 2025, and the regulator warned that LLMs and autonomous software can flood complaint systems with duplicative submissions.

A Nature Human Behaviour study estimates that LLM use raises the probability of favorable relief at the CFPB by 6.9 percentage points.

These figures describe complaint systems, and the CFPB data covers mostly credit reporting.

Mastercard and Datos’s 2025 outlook projected 324 million chargebacks worldwide by 2028. Mastercard’s 2026 US merchant benchmark of $128 per chargeback covers internal costs and third-party fees, excluding the lost goods or services. Applying that US benchmark to the global volume as an illustrative assumption, a 5% increase would add 16.2 million chargebacks and about $2.1 billion in operational costs; a 15% increase would add 48.6 million and about $6.2 billion.

Nemse noted that every queue behind those cases is staffed by humans, “Amazon's included,” and they are “already bending.”

Scenario	Increase in chargebacks	Added chargebacks vs. 324M baseline	Added operational cost at $128 each	What it shows
2025 outlook baseline	0%	0	$0	No additional chargebacks in this scenario
+5% case	+5%	16.2M	~$2.1B	Even small automation effects become material
+15% case	+15%	48.6M	~$6.2B	Dispute automation could become a major merchant cost
+30% stress case	+30%	97.2M	~$12.4B	Human review queues could become the bottleneck

Regulators put merchant costs on the record

The RBA's Oct. 6 summary of its payments consultation drew on written submissions from 75 stakeholders. Merchants, payment service providers and issuers said chargeback rules leave liability unclear when an agent acts outside its authority.

They said agentic commerce could raise merchant costs and that networks may struggle to tell whether an agent followed its customer's instructions. One stakeholder referred to reports of an additional 4% charge for AI-assisted purchases.

Submissions described adoption as early and evidence of harm as limited, generally favoring industry standards and monitoring. The RBA plans to announce regulatory priorities by the end of 2026.

A consumer's agent can file a dispute at negligible cost, while a merchant responds with evidence, including logistics records and processor workflows. Nemse expects agents to “dispute far more often, because disputing costs them nothing.”

Platforms keep the judge

Amazon blocked Meta's Muse shopping agent, citing unauthorized access and its own policies, and Nemse reads the block as a fight over the interface.

He said Amazon “has no doubt Meta's agent can buy something,” and it wants to keep the interface because becoming an API that another company's agent consumes would hand over the customer relationship, data, and advertising real estate worth billions.

Google faces the same problem, and in his view “they'll block outside agents and ship their own.”

Small merchants sit in the opposite position, because “an agent searches for whoever solves the problem best, not whoever bought the ad.” Shopify has moved toward admitting browser-based AI shopping agents into checkout.

Nemse argued that discovery covers half the job, since platforms own “the interface and the judge.” Agents can take the first, and the second needs a credible, neutral dispute process. He added:

“Without it, your agent finds the small merchant, and you still go back to Amazon.”

A CI&T survey of 1,011 US consumers found 27% comfortable with full AI shopping. Nemse puts the ceiling on agentic commerce at “the loss they'll accept with no recourse.”

API calls cost cents, so agents pay for API calls. For work, insurance claims, or refunds, “nobody lets an agent commit” unless recourse exists and someone is clearly liable. Nemse noted that “better payment rails don't move that ceiling.”

Who judges the machines

Google's AP2, Mastercard Agent Pay and Visa Intelligent Commerce focus on authorization, with signed mandates, tokenized credentials, spending controls and agent identity. A mandate proves what the buyer instructed and leaves the delivery judgment open.

Nemse's answer is validators running AI models that reach consensus on the outcome, with the decision enforced on-chain and open to appeal, a design his GenLayer Foundation is building.

Model	Who controls the interface?	Who decides disputes?	Strength	Weakness
Amazon-style platform	Platform	Platform support/refund system	Buyer trust and clear recourse	Keeps merchants dependent on platform rules
Open merchant web	Agent or browser	Unclear	More distribution for small merchants	Weak recourse unless standards emerge
Card-network model	Merchant, agent, wallet or network	Existing dispute/chargeback rails	Familiar liability infrastructure	May struggle with agent intent and subjective fulfillment
On-chain escrow/adjudication	Agent-facing apps or protocols	Validators / arbitration process	Can enforce escrowed funds programmatically	Cannot automatically compel off-chain refunds
GenLayer-style AI consensus	Open agent ecosystem	AI validators with appeals	Targets subjective outcomes at machine scale	Must prevent frivolous disputes and bad model decisions

GenLayer says common cases can finalize in roughly 30 minutes and fully escalated ones in about three hours.

Fees, bonds, or reputation penalties have to make frivolous disputes uneconomic when filing is free for an AI agent, and validators judge the evidence supplied, such as receipts, tracking, and task specifications.

Appeals protect against bad model outputs and add time and cost, and an on-chain verdict governs escrowed funds while an ordinary merchant's card refund sits outside its reach.

Where agentic commerce goes from here

If merchants and networks settle on standards, with verifiable mandates, merchant evidence records, and escrow that filters disputes before they become chargebacks, AI agents can move from API calls into services and unfamiliar counterparties. The long tail would gain the recourse that platforms enjoy today.

If disputes stay cheap to file and costly to resolve, merchants raise fees, restrict agent purchases, or send buyers back to trusted platforms.