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

Amazon’s Bedrock AgentCore Payments, developed in partnership with Coinbase and Stripe, allows AI agents to locate paid services, handle authentication, and complete transactions using stablecoins and x402 under predefined spending limits.

Artificial intelligence agents are securing digital wallets even before merchants establish a legal framework, and shopping assistants have already begun placing orders across the open web.

The greater complication arises once a transaction clears—specifically when a buyer’s agent pays accurately, but the purchaser subsequently requests a refund. The Reserve Bank of Australia (RBA) officially highlighted this regulatory gap on Oct. 6, and Edgars Nemse, CEO of the GenLayer Foundation, stated that this issue ultimately caps what autonomous agents are able to purchase.

Nemse explained to CryptoSlate that executing a payment “is the easy part, because it’s deterministic: the money moved, or it didn’t,” whereas “the outcome isn’t.” Determining whether a service was rendered as promised remains a subjective evaluation.

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 resolution mechanism operates on the unstated premise that filing a grievance is arduous enough that most individuals simply abandon the effort. AI is rapidly dismantling that barrier, with humans currently deploying software to file complaints on their behalf.

Complaints submitted to the Consumer Financial Protection Bureau doubled to 6.6 million in 2025, prompting the regulator to issue warnings that large language models (LLMs) and autonomous applications could overwhelm complaint infrastructure with duplicate filings.

A study published in Nature Human Behaviour estimates that utilizing LLMs increases the likelihood of favorable relief at the CFPB by 6.9 percentage points.

While these statistics describe general complaint channels, the CFPB dataset predominantly reflects credit reporting issues.

Projections from Mastercard and Datos for 2025 estimate 324 million chargebacks globally by 2028. Mastercard’s 2026 U.S. merchant benchmark places the cost at $128 per chargeback—accounting for internal expenses and third-party fees, but excluding the lost merchandise or services. Applying this U.S. baseline to the worldwide volume hypothetically, a 5% increase would introduce 16.2 million additional chargebacks and approximately $2.1 billion in operational expenditures, while a 15% increase would yield 48.6 million additional cases and roughly $6.2 billion in costs.

Nemse pointed out that every human-staffed support queue handling these disputes, “Amazon’s included,” is “already bending” under the strain.

परिदृश्य	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 concerning its payments consultation incorporated written contributions from 75 market participants. Merchants, payment service providers, and card issuers indicated that current chargeback frameworks leave liability ambiguous when an autonomous agent operates outside its authorized scope.

They warned that agentic commerce could drive up merchant expenditures and that payment networks may find it difficult to verify whether an agent faithfully adhered to consumer directives. One participant cited reports indicating an extra 4% surcharge applied to AI-assisted purchases.

Submissions characterized industry adoption as preliminary and concrete evidence of consumer harm as minimal, leaning generally toward establishing voluntary industry standards and ongoing monitoring. The RBA intends to release its definitive regulatory priorities prior to the close of 2026.

Because a consumer’s agent can file a dispute with virtually no financial overhead, while merchants must respond by compiling evidence such as fulfillment records and processor workflows, Nemse anticipates that agents will “dispute far more often, because disputing costs them nothing.”

Platforms keep the judge

Amazon blocked Meta’s Muse shopping agent, citing unauthorized access and corporate policy violations—a move Nemse interprets as a strategic battle over user interfaces.

He noted that Amazon “has no doubt Meta’s agent can buy something,” but prefers to retain control of the interface because operating merely as an API backend for another firm’s agent would surrender valuable customer relationships, proprietary data, and lucrative advertising real estate.

Google confronts a similar dilemma, and in his estimation, “they’ll block outside agents and ship their own.”

In contrast, independent merchants occupy a different position, since “an agent searches for whoever solves the problem best, not whoever bought the ad.” Shopify has recently taken steps to welcome browser-based AI shopping agents into its checkout flows.

Nemse asserted that product discovery covers only half the challenge, as closed platforms control both “the interface and the judge.” While independent agents can conquer the former, the latter requires a dependable, neutral adjudication mechanism. He added:

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

A survey conducted by CI&T involving 1,011 U.S. consumers revealed that 27% feel comfortable trusting automated shopping tools completely. Nemse estimates the upper limit of agentic commerce will be dictated by “the loss they’ll accept with no recourse.”

Given that API requests cost only fractions of a cent, agents routinely pay for them autonomously. However, when dealing with complex labor, insurance claims, or financial refunds, “nobody lets an agent commit” funds unless verified consumer recourse exists and liability is clearly assigned. Nemse emphasized that “better payment rails don’t move that ceiling.”

Who judges the machines

Google’s AP2, Mastercard Agent Pay, and Visa Intelligent Commerce focus heavily on transaction authorization via signed mandates, tokenized credentials, spending caps, and verified agent identities. While a mandate verifies what the buyer requested, it leaves fulfillment evaluation unresolved.

Nemse’s proposed solution involves distributed network validators executing AI models to reach consensus regarding task outcomes, with decisions enforced on-chain and made accessible for appeal—an architecture currently being developed by his GenLayer Foundation.

मॉडल	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

According to GenLayer, routine cases can reach final resolution within roughly 30 minutes, whereas fully escalated scenarios require approximately three hours.

To prevent frivolous arbitrations when filing is cost-free for automated systems, platforms must implement financial fees, security bonds, or reputation penalties, allowing validators to objectively assess submitted evidence including digital receipts, tracking logs, and task definitions.

While structured appeals guard against erroneous model outputs, they introduce additional time and expense. Furthermore, an on-chain verdict can automatically govern escrowed capital, whereas traditional credit card refunds processed by standard merchants remain outside direct blockchain control.

Where agentic commerce goes from here

If merchants and payment networks successfully establish universal standards—incorporating verifiable mandates, comprehensive merchant evidence records, and escrow systems that filter out unfounded disputes before they escalate into chargebacks—AI agents can safely transition from basic API queries into complex service transactions with unfamiliar counterparties. Consequently, smaller independent merchants would secure the same consumer protections currently enjoyed by dominant closed platforms.

Conversely, if dispute filings remain inexpensive to initiate yet costly to resolve, merchants will likely increase service fees, restrict autonomous agent purchases, or direct consumers back toward established, trusted platforms.

?अक्सर पूछे जाने वाले प्रश्न

01What are AI shopping agents?

AI shopping agents are autonomous software programs designed to browse the internet, discover services, authenticate identities, and complete financial transactions on behalf of users.

02Why are transaction disputes difficult for AI agents?

While transferring money is a straightforward and deterministic process, determining whether a delivered product or service meets expectations requires subjective judgment that traditional automated systems struggle to evaluate.

03What is a chargeback?

A chargeback is a demand by a credit card-issuing bank for a merchant to make good on the loss on a fraudulent or disputed transaction.

04How do platforms handle autonomous shopping agents?

Major platforms like Amazon often block outside AI agents to protect their proprietary user interfaces, customer data relationships, and advertising revenue.