AI agents can pay for your shopping. Who gets your money back?
Amazon’s Bedrock AgentCore Payments, constructed with Coinbase and Stripe, lets AI agents uncover paid providers, authenticate, and pay with stablecoins and x402 below preset spending limits.
AI agents are getting wallets earlier than retailers get a court docket, and procuring agents have began inserting orders throughout the open web.
The more durable drawback arrives once the payment clears, when a purchaser’s agent pays accurately, and the customer desires the money again. The Reserve Bank of Australia (RBA) put that hole on the report Oct. 6, and Edgars Nemse, CEO of the GenLayer Foundation, argued it caps what agents can purchase.
Nemse informed CryptoSlate that fee “is the straightforward half, as a result of it is deterministic: the money moved, or it did not,” whereas “the result is not.” Defining if work was delivered as promised is a judgment name.
| Transaction stage | What agents can more and more do | What stays unresolved | Why it issues |
|---|---|---|---|
| Discovery | Find retailers, APIs, content material, providers | Whether retailers belief unknown agents | Controls who gets distribution |
| Authorization | Use mandates, credentials, spending limits | Whether the agent stayed inside person intent | Determines who bears legal responsibility |
| Payment | Pay with playing cards, stablecoins, x402 or wallets | Payment success doesn’t show satisfaction | Settlement is deterministic |
| Fulfillment | Receive items, providers or API entry | Was the result delivered as promised? | Requires judgment |
| Dispute | Submit complaints or refund requests | Who adjudicates exterior a platform? | Determines whether or not open commerce can scale |
AI agents take away the friction disputes depend upon
According to Nemse, each dispute system runs on a hidden assumption that disputing is tedious sufficient that most individuals skip it. AI is already eroding that friction, with individuals submitting 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 program can flood grievance techniques with duplicative submissions.
A Nature Human Behaviour examine estimates that LLM use raises the likelihood of favorable reduction at the CFPB by 6.9 percentage points.
These figures describe grievance techniques, and the CFPB knowledge covers principally credit score 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 inside prices and third-party charges, excluding the misplaced items or providers. Applying that US benchmark to the worldwide quantity as an illustrative assumption, a 5% enhance would add 16.2 million chargebacks and about $2.1 billion in operational prices; a 15% enhance would add 48.6 million and about $6.2 billion.
Nemse famous that each queue behind these instances is staffed by humans, “Amazon’s included,” and they’re “already bending.”
| Scenario | Increase in chargebacks | Added chargebacks vs. 324M baseline | Added operational value at $128 every | What it exhibits |
|---|---|---|---|---|
| 2025 outlook baseline | 0% | 0 | $0 | No extra chargebacks on this state of affairs |
| +5% case | +5% | 16.2M | ~$2.1B | Even small automation results turn out to be materials |
| +15% case | +15% | 48.6M | ~$6.2B | Dispute automation may turn out to be a serious service provider value |
| +30% stress case | +30% | 97.2M | ~$12.4B | Human assessment queues may turn out to be the bottleneck |
Regulators put service provider prices on the report
The RBA’s Oct. 6 abstract of its funds session drew on written submissions from 75 stakeholders. Merchants, fee service suppliers and issuers mentioned chargeback guidelines depart legal responsibility unclear when an agent acts exterior its authority.
They mentioned agentic commerce may elevate service provider prices and that networks might battle to inform whether or not an agent adopted its buyer’s directions. One stakeholder referred to reports of an additional 4% charge for AI-assisted purchases.
Submissions described adoption as early and proof of hurt as restricted, typically favoring business requirements and monitoring. The RBA plans to announce regulatory priorities by the top of 2026.
A client’s agent can file a dispute at negligible value, whereas a service provider responds with proof, together with logistics information and processor workflows. Nemse expects agents to “dispute much more typically, as a result of disputing prices them nothing.”
Platforms hold the decide
Amazon blocked Meta’s Muse procuring agent, citing unauthorized entry and its personal insurance policies, and Nemse reads the block as a combat over the interface.
He mentioned Amazon “has little question Meta’s agent can purchase one thing,” and it desires to maintain the interface as a result of changing into an API that one other firm’s agent consumes would hand over the shopper relationship, knowledge, and promoting actual property price billions.
Google faces the identical drawback, and in his view “they will block exterior agents and ship their very own.”
Small retailers sit within the reverse place, as a result of “an agent searches for whoever solves the issue finest, not whoever purchased the advert.” Shopify has moved towards admitting browser-based AI procuring agents into checkout.
Nemse argued that discovery covers half the job, since platforms personal “the interface and the decide.” Agents can take the primary, and the second wants a reputable, impartial dispute course of. He added:
“Without it, your agent finds the small service provider, and you continue to return to Amazon.”
A CI&T survey of 1,011 US customers found 27% comfortable with full AI shopping. Nemse places the ceiling on agentic commerce at “the loss they will settle for with no recourse.”
API calls value cents, so agents pay for API calls. For work, insurance coverage claims, or refunds, “no one lets an agent commit” until recourse exists and somebody is clearly liable. Nemse famous that “better payment rails do not transfer that ceiling.”
Who judges the machines
Google’s AP2, Mastercard Agent Pay and Visa Intelligent Commerce concentrate on authorization, with signed mandates, tokenized credentials, spending controls and agent id. A mandate proves what the customer instructed and leaves the supply judgment open.
Nemse’s reply is validators working AI fashions that attain consensus on the result, with the choice enforced on-chain and open to enchantment, a design his GenLayer Foundation is constructing.
| Model | Who controls the interface? | Who decides disputes? | Strength | Weakness |
|---|---|---|---|---|
| Amazon-style platform | Platform | Platform assist/refund system | Buyer belief and clear recourse | Keeps retailers depending on platform guidelines |
| Open service provider net | Agent or browser | Unclear | More distribution for small retailers | Weak recourse until requirements emerge |
| Card-network mannequin | Merchant, agent, pockets or community | Existing dispute/chargeback rails | Familiar legal responsibility infrastructure | May battle with agent intent and subjective success |
| On-chain escrow/adjudication | Agent-facing apps or protocols | Validators / arbitration course of | Can implement escrowed funds programmatically | Cannot mechanically compel off-chain refunds |
| GenLayer-style AI consensus | Open agent ecosystem | AI validators with appeals | Targets subjective outcomes at machine scale | Must stop frivolous disputes and dangerous mannequin selections |
GenLayer says widespread instances can finalize in roughly 30 minutes and fully escalated ones in about three hours.
Fees, bonds, or repute penalties need to make frivolous disputes uneconomic when submitting is free for an AI agent, and validators decide the proof provided, equivalent to receipts, monitoring, and job specs.
Appeals shield towards dangerous mannequin outputs and add time and value, and an on-chain verdict governs escrowed funds whereas an unusual product owner’s card refund sits exterior its attain.
Where agentic commerce goes from right here
If retailers and networks choose requirements, with verifiable mandates, service provider proof information, and escrow that filters disputes earlier than they turn out to be chargebacks, AI agents can transfer from API calls into providers and unfamiliar counterparties. The lengthy tail would achieve the recourse that platforms take pleasure in at the moment.
If disputes keep low cost to file and expensive to resolve, retailers elevate charges, limit agent purchases, or ship consumers again to trusted platforms.
The publish AI agents can pay for your shopping. Who gets your money back? appeared first on CryptoSlate.
