Amazon's Bedrock AgentCore Payments lets AI agents pay with stablecoins under preset limits. But refund and dispute gaps threaten to drive up merchant costs and shift liability across online commerce.
AI shopping agents are now placing real orders and moving stablecoins, but the industry faces a problem that payment rails alone cannot patch. When a buyer wants their money back, the rules get murky. Amazon's Bedrock AgentCore Payments, built with Coinbase and Stripe, lets AI agents find services, authenticate, and pay using stablecoins or x402 tokens within set limits. As these agents gain more transactional muscle, the question of who eats the loss on a refund remains wide open. Merchants are watching their risk climb.
Once an AI agent finishes a purchase, the payment itself is binary: funds move or they do not. But fulfillment-whether the buyer actually gets what they paid for-remains a judgment call. The Reserve Bank of Australia's October 6 payments consultation flagged that merchants, payment providers, and issuers see no clear liability rules when an agent acts outside a user's intent. In the RBA's review, 75 organizations warned that without clear agent instruction records and the ability to spot agent-initiated transactions, no one can reliably assign blame for losses. Some cited reports of an extra 4% fee for AI-assisted purchases. Many warned that agent-driven commerce could push merchant costs higher, especially if networks cannot prove an agent followed customer instructions.
In the current landscape, there are no industry-wide mechanisms for refunds or chargebacks in account-to-account payments, leaving merchants exposed to unresolved disputes.
Edgars Nemse, CEO of the GenLayer Foundation, says the real choke point is not payment but dispute resolution. He points out that every dispute system depends on the friction of filing complaints-friction that AI is quickly erasing. The Consumer Financial Protection Bureau (CFPB) logged 6.6 million complaints in 2025, double the year before, and warned that large language models and autonomous software could swamp complaint systems with duplicate filings. A Nature Human Behaviour study found LLM use boosted the chance of favorable relief at the CFPB by 6.9 percentage points, though most complaints still target credit reporting. Mastercard and Datos expect 324 million chargebacks worldwide by 2028. A 5% rise would add $2.1 billion in operational costs, and a 15% jump would tack on $6.2 billion, using a U.S. benchmark of $128 per chargeback. Nemse notes that every one of those cases needs human review, and queues are already stretched thin.
Major platforms are not waiting for consensus. Amazon recently blocked Meta's Muse shopping agent, citing unauthorized access and policy breaches. Nemse sees this as a fight over who controls the customer interface-and the data that comes with it. He expects Google to block outside agents and launch its own. For smaller merchants, the dynamic shifts: AI agents hunt for the best solution, not the highest ad bidder. Shopify now lets browser-based AI shopping agents check out, but without a neutral, credible dispute process, buyers may still stick with trusted platforms like Amazon when recourse is unclear.
Discovery is only part of the puzzle. Platforms now own both the interface and the judge's seat in disputes. Agents can take over the interface, but agentic commerce needs a scalable, neutral dispute process to move beyond API calls. Without clear recourse and liability, merchants may cap agent purchases or hike fees to cover risk. The Reserve Bank of Australia noted that for account-to-account payments, there are no industry-wide refund or chargeback mechanisms like those in card systems. Consultation participants backed the idea that liability should fall to the party best able to prevent or control risk, without piling on smaller providers.
The Reserve Bank of Australia has not yet introduced mandatory rules for AI-initiated payments, instead favoring voluntary industry standards and ongoing market monitoring. Regulatory priorities for AI-driven payment systems are expected to be set by the end of 2026.
Several models are surfacing to close the dispute gap. Card networks lean on existing chargeback rails, but may struggle to judge agent intent and fulfillment. On-chain escrow and adjudication can enforce programmatic refunds, but cannot force off-chain merchants to return funds. GenLayer proposes a system where validators run AI models to reach consensus, with decisions enforced on-chain and open to appeal. This setup aims to resolve common cases in about 30 minutes and escalated cases in three hours, but must deter bogus disputes with fees, bonds, or reputation penalties. Appeals add time and cost, and on-chain verdicts only cover escrowed funds, not traditional card refunds.
If industry standards take hold-combining verifiable mandates, merchant evidence, and escrow that filters disputes before escalation-AI agents could safely transact with more counterparties. Otherwise, the cost and hassle of resolving disputes may push merchants to limit agent-driven commerce or steer buyers back to established platforms. As seen in recent developments around on-chain recovery and treasury claims, the technical rails for AI payments are moving ahead, but the legal and operational rails for refunds and recourse are lagging.
Mastercard's 2026 U.S. merchant benchmark puts each chargeback at $128 in internal and third-party fees, not counting lost goods or services. With 324 million chargebacks projected globally by 2028, even a small uptick in AI-driven disputes could add billions in operational costs. The Reserve Bank of Australia plans to set regulatory priorities by the end of 2026, but for now, most stakeholders back industry standards and monitoring over immediate intervention.
AI agents are making it easier to trigger disputes, but the infrastructure to resolve them at scale is falling behind. Until credible, neutral, and scalable dispute processes exist, the ceiling for agentic commerce will be set by how much loss buyers are willing to accept without recourse.
AI-driven payment systems are forcing a reckoning over liability, recourse, and operational cost in digital commerce. As technical rails for agent payments mature, the lack of robust dispute and refund mechanisms threatens to shift risk and expense onto merchants, especially those outside major platforms.
Dispute resolution in digital payments is fundamentally different from the deterministic nature of blockchain settlement. A blockchain can prove a payment was made, but cannot automatically decide if a service was delivered as promised or if a refund is justified. This gap between technical finality and subjective fulfillment sits at the heart of the current liability debate.
