AI Agents Can Now Pay for Your Shopping, but Who Gets Your Money Back?

The Automation of Commerce Disputes and the Collapse of Process Friction

In the evolving landscape of autonomous digital transactions, processing payments represents the simplest step of the pipeline. According to Nemse in comments to CryptoSlate, payment “is the easy part, because it’s deterministic: the money moved, or it didn’t,” while “the outcome isn’t.” The subjective challenge of establishing whether contracted work or a purchased product was delivered as promised remains an interpretive judgment call—one that legacy retail and payment infrastructure is increasingly ill-equipped to handle.

Historically, dispute workflows have relied on the practical friction of administrative effort to deter trivial claims. 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. The downstream effects of this shift are already visible in consumer protection mechanisms. Complaints submitted to the Consumer Financial Protection Bureau doubled to 6.6 million in 2025, prompting the regulator to warn that LLMs and autonomous software can flood complaint systems with duplicative submissions. Furthermore, a study published in Nature Human Behaviour estimates that LLM use raises the probability of favorable relief at the CFPB by 6.9 percentage points, highlighting the growing power of automated consumer advocacy in domains like credit reporting.

Rising Merchant Costs and Emerging Liability Questions

The acceleration of agent-driven submissions threatens significant economic consequences for commercial dispute resolution. A 2025 outlook by Mastercard and Datos projected 324 million chargebacks worldwide by 2028. Based on Mastercard’s 2026 US merchant benchmark of $128 per chargeback—which accounts for internal expenses and third-party fees, excluding the cost of lost goods or services—even marginal increases in chargeback rates create severe overhead. Illustratively applying that US benchmark to worldwide totals, 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. As Nemse noted, the dispute queues handling these claims rely on human personnel, “Amazon’s included,” and they are “already bending.”

Financial regulators are actively documenting these vulnerabilities. The Reserve Bank of Australia released an Oct. 6 summary of its payments consultation incorporating feedback from 75 stakeholders across merchant, payment service provider, and issuer groups. Participants warned that prevailing chargeback regulations fail to clarify liability when an autonomous agent acts outside its user-granted authority. Stakeholders cautioned that agentic commerce could drive up merchant costs—citing reports of an additional 4% charge for AI-assisted purchases—while card networks may struggle to distinguish whether an AI agent faithfully executed consumer instructions. While submissions characterized current adoption as nascent and documented harms as limited, the RBA intends to outline regulatory priorities by the end of 2026. Because an autonomous consumer agent can initiate actions with virtually zero resource expenditure while sellers must assemble detailed evidence portfolios, Nemse expects agents to “dispute far more often, because disputing costs them nothing.”

Interface Control and the Search for Neutral Arbitration

Beyond administrative processing costs, agentic transactions have triggered major ecosystem battles over consumer interfaces. When Amazon blocked Meta’s Muse shopping agent over unauthorized access and internal policy adherence, Nemse framed the defensive measure as a struggle over interface ownership. He observed that Amazon “has no doubt Meta’s agent can buy something,” but maintains restrictions because becoming an API that another company’s agent consumes would hand over the customer relationship, data, and advertising real estate worth billions. In his assessment, Google faces the same problem, and in his view “they’ll block outside agents and ship their own.”

Conversely, independent sellers and smaller merchants face different dynamics because “an agent searches for whoever solves the problem best, not whoever bought the ad.” While platforms like Shopify have initiated efforts to welcome browser-based AI shopping agents into checkout funnels, discoverability solves only half the structural challenge. Because centralized marketplaces currently control both the interface and adjudicative processes, Nemse emphasized the necessity of neutral dispute frameworks, remarking: “Without it, your agent finds the small merchant, and you still go back to Amazon.”

Decentralized Consensus and On-Chain Adjudication

Consumer readiness for end-to-end agentic transactions remains bounded by structural safeguards. A CI&T survey encompassing 1,011 US consumers indicated that only 27% felt comfortable with full AI shopping. Nemse characterizes the boundary for agentic commerce as “the loss they’ll accept with no recourse.” While micro-transactions and API calls function easily because they cost mere cents, larger commitments such as service agreements, insurance claims, or refunds face strict barriers where “nobody lets an agent commit” without definitive recourse and liability frameworks, noting that “better payment rails don’t move that ceiling.”

Industry-led protocols—such as Google’s AP2, Mastercard Agent Pay, and Visa Intelligent Commerce—concentrate primarily on the authorization layer through agent identity verification, signed mandates, tokenized credentials, and configurable spending controls. Yet while mandates authenticate user instructions, they do not resolve post-purchase delivery disputes. To address the outcome evaluation problem, Nemse’s GenLayer Foundation is developing a consensus model where network validators operate AI models to judge submitted evidence—such as task specifications, tracking data, and receipts—enforcing decisions directly on-chain alongside an appeals process.

According to GenLayer, typical disputes can reach finality in approximately 30 minutes, with fully escalated appeals resolving in roughly three hours. To prevent bad-faith or trivial submissions when agent execution costs remain negligible, the system uses economic disincentives including fees, bonds, or reputation penalties. However, while on-chain verdicts can govern funds locked in escrow, traditional merchant card refunds continue to exist outside smart contract execution rails.

Why This Matters

The progression of agentic commerce hinges directly on bridging the gap between authorization and outcome settlement. If commercial networks and merchant ecosystems develop universal standards that combine verifiable mandates, robust transaction evidence, and escrow safeguards to resolve conflicts prior to chargebacks, AI agents can reliably facilitate trade across unfamiliar merchants and independent web properties. Conversely, if dispute generation remains practically free for automated agents while defense costs remain steep for sellers, merchants will inevitably enact defensive friction—imposing surcharge fees, blocking autonomous buying tools, and further cementing consumer reliance on dominant platform walled gardens.

Frequently Asked Questions

Why are AI agents increasing dispute and chargeback risks?

Traditional dispute frameworks depend on human friction; filing disputes requires time and effort, deterring small or marginal claims. AI agents make submitting disputes virtually instantaneous and free for consumers, enabling automated mass filings that can overwhelm merchant support teams and dramatically increase operational fees.

Why are dominant platforms restricting third-party shopping agents?

Major platforms like Amazon maintain extensive value through their direct consumer interfaces, user behavioral data, and advertising businesses. Permitting third-party agents to process purchases directly turns platforms into commoditized back-end APIs, jeopardizing their customer relationships and monetizable ad space.

How do proposed validator-based dispute networks function?

Rather than relying on centralized manual support staff or rigid binary payment signals, validator-based dispute models—such as the one being developed by GenLayer—use AI models operated by network validators to assess evidence, reach consensus on whether contractual expectations were fulfilled, and enforce outcomes programmatically on-chain.