AI agents are moving closer to shopping independently: they can search for services, authenticate, and pay with stablecoins under preset limits. But a payment system can confirm that money changed hands without deciding whether the goods, service or result met the buyer’s expectations. That unresolved question could determine how widely autonomous shopping develops and who ultimately bears the cost of mistakes.
Payment is becoming easier than recourse
Amazon’s Bedrock AgentCore Payments, developed with Coinbase and Stripe, is designed to let AI agents discover paid services and complete transactions using stablecoins and the x402 payment system. Spending limits can be set in advance, allowing an agent to act without asking for approval at every step.
The difficulty begins after settlement. If an agent follows its instructions and pays a merchant, but the buyer says the service was incomplete or unsuitable, existing payment rails do not necessarily provide a neutral way to assess the outcome. According to CryptoSlate, Edgars Nemse, chief executive of the GenLayer Foundation, described payment as relatively straightforward because it is clear whether funds moved. Determining whether a promised result was delivered is more subjective.
The Reserve Bank of Australia highlighted this issue in an Oct. 6 summary of its payments consultation. Written submissions from 75 stakeholders, including merchants, payment providers and issuers, said liability is unclear when an agent acts beyond the authority granted by its customer. They also warned that networks may find it difficult to establish whether an agent properly followed the buyer’s instructions.
The RBA said adoption remains at an early stage and evidence of harm is limited. It plans to announce regulatory priorities by the end of 2026. Some submissions nevertheless pointed to a possible additional charge of 4% for purchases assisted by AI.
More automated disputes could increase costs
AI may also make complaints and disputes much easier to file. Nemse argued that conventional dispute systems rely partly on friction: customers often abandon a claim because it takes time and effort. An AI agent could prepare and submit a complaint at little cost, potentially changing the balance between consumers and merchants.
The source material cites Consumer Financial Protection Bureau data showing that complaints doubled to 6.6 million in 2025. The regulator warned that large language models and autonomous software could generate duplicative submissions. A study in Nature Human Behaviour estimated that using an LLM increased the probability of favorable relief at the CFPB by 6.9 percentage points. The cited CFPB figures mainly concern credit reporting, rather than autonomous shopping disputes.
Mastercard and Datos projected 324 million chargebacks worldwide by 2028. Mastercard’s 2026 US merchant benchmark put internal and third-party costs at $128 per chargeback, excluding the value of lost goods or services. As an illustration, a 5% increase on the projected volume would mean 16.2 million additional chargebacks and roughly $2.1 billion in operational costs. A 15% increase would mean 48.6 million more chargebacks and about $6.2 billion in costs.
Those figures are illustrative applications of the US benchmark to global volume, not a forecast of AI-driven disputes. The underlying concern is that merchants may need to spend more time and money answering claims while agents can submit them almost instantly.
Platforms, merchants and the missing judge
The dispute question is also tied to control over online shopping. Amazon blocked Meta’s Muse shopping agent, citing unauthorized access and its own policies. Nemse interpreted the conflict as competition over the customer interface, data and advertising space, rather than uncertainty about whether Meta’s agent could technically make a purchase.
Large platforms can build their own shopping agents and retain control over both transactions and disputes. Smaller merchants may benefit from agents that search for the seller best able to solve a customer’s problem, rather than simply selecting the business that paid most for advertising. Shopify has moved toward allowing browser-based AI shopping agents into checkout.
However, discovery alone does not solve the problem. If a merchant and buyer disagree about delivery, quality or performance, a trusted and neutral process is still needed. A survey of 1,011 US consumers by CI&T found that 27% were comfortable with fully AI-driven shopping. Nemse said the eventual scale of agentic commerce could be limited by the loss consumers are willing to accept when they have no effective recourse.
Visa Intelligent Commerce, Mastercard Agent Pay and Google’s AP2 focus on authorization, including signed mandates, tokenized credentials, spending controls and agent identity. These tools can help prove what a customer instructed an agent to do, but they do not by themselves determine whether the merchant fulfilled the agreement.
One proposed approach is to use validators running AI models to assess evidence such as receipts, tracking information and task specifications. GenLayer says its design can resolve ordinary cases in about 30 minutes and fully escalated cases in roughly three hours, with decisions enforced on-chain and subject to appeal. Fees, bonds or reputation penalties could be used to discourage claims that are cheap to file but costly to investigate.
That type of process would apply to escrowed funds, while an ordinary card refund may remain outside the reach of an on-chain verdict. If standards emerge around verifiable mandates, merchant evidence and dispute screening, agents could eventually transact beyond familiar platforms. If not, merchants may raise prices, restrict autonomous purchases or direct customers back to dominant marketplaces.
Moneyiar market picture
Moneyiar data at 11:08-11:09 on Oct. 9, 2026, showed mixed performance among the listed assets:
- Cronos (CRO): $0.060841228, down 2.69% in 24 hours and 12.5% over seven days.
- Stable (STABLE): $0.0253673074, down 3.23% in 24 hours and 10.18% over seven days.
- Monad (MON): $0.0244295781, down 3.97% in 24 hours and 27.92% over seven days.
- BAS (BAS): $0.0227231127, up 0.31% in 24 hours and down 4.33% over seven days.
These market figures describe price movements at the stated times and do not determine how AI payment or dispute systems will develop.
