The challenge for AI agents involves determining payment responsibility when automation malfunctions

As AI agents gain authority to execute financial transactions, payment systems must resolve a fundamental gap between what users authorize and what they actually want. The emerging standards and protocols for agent payments reveal how much work lies ahead in translating human intent into machine-readable rules.

• Mastercard forecasts one in ten people will routinely use AI agents to shop and pay by 2030.
• Google’s Agent Payments Protocol uses digitally signed records called mandates to connect user instructions with proposed purchases.
• The x402 payment standard enables agents to make small, programmatic payments to services for data without requiring user approval for each transaction.

• 2030 Year when one in ten people expected to use AI agents for shopping regularly
• $0.02 Cost per lookup in illustrated example of small repeated agent payments
• $10.00 Total spent if agent repeats 500 lookups at two cents each instead of five
• $200 Spending ceiling example preventing five-hundred-dollar bookings in hotel reservation scenario

As payment companies develop infrastructure for AI agents, they confront a problem that extends far beyond technical architecture: when autonomous software executes a transaction on behalf of a user, who bears responsibility if the outcome falls short of expectations? A hotel booking agent might follow instructions precisely yet book a windowless room with non-refundable terms and hidden breakfast costs, leaving the traveler stuck with a purchase they would never have authorized had they reviewed it themselves. This gap between permission and satisfaction is becoming central to how payment systems must be designed.

The challenge is amplified by the speed and scale at which agents operate. Unlike traditional e-commerce where humans make deliberate purchasing decisions, agents may execute dozens of transactions in seconds while gathering information or completing complex tasks. Payment systems must therefore embed authorization, verification, and accountability mechanisms into the agent’s operating logic rather than relying on post-purchase review. The infrastructure emerging today will shape whether agent-driven commerce becomes a trusted feature of daily life or a source of consumer frustration and fraud.

Google’s mandates connect user intent to agent actions

Google’s Agent Payments Protocol addresses part of the authorization problem through digitally signed records called mandates. These documents establish a verifiable chain linking a user’s stated instructions to specific proposed purchases and payments. Rather than relying on chat logs or user recollection, a mandate creates cryptographic evidence of what the user approved.

For the hotel example, a mandate could record the maximum total price and the exact room the user authorized. If an AI agent books a different room or exceeds the budget, the signed record provides all parties with objective proof of what permission covered. Users can either approve specific shopping carts in advance or authorize purchases under predefined conditions, shifting the verification burden from humans arguing about what was said to systems checking whether actions matched signed instructions.

Card networks and crypto standards offer different trade-offs

Mastercard’s Agent Pay framework incorporates authorization, authentication, and granular controls over what an assistant is permitted to do, operating alongside the established dispute and refund processes card networks already provide. Card tokens replace sensitive credentials with identifiers, allowing agents to transact while keeping payment details secure. This approach leverages decades of consumer protections and chargeback mechanisms that buyers have come to expect.

Cryptocurrency stablecoins and payment protocols like x402 take a different approach. Dollar stablecoins are transferable assets designed to track the dollar’s value, while x402 enables agents to make small, programmatic payments to services without requiring users to open accounts or approve every transaction individually. A travel assistant might pay cents to check real-time room availability across multiple sources as part of its booking logic. The critical constraint in both systems is how tightly permissions are configured; a well-designed crypto wallet can enforce stricter spending limits than a poorly configured card service, but cards provide established dispute processes that basic token transfers do not.

Small payments create control and tracking challenges

When an AI agent can pay for information or services in cents, budget control becomes essential.

Consider an agent checking a data service at two cents per lookup. If the agent queries five times, it spends ten cents. If a bug causes it to repeat the task 500 times, the total climbs to ten dollars. Per-payment limits alone cannot prevent this; the system must track cumulative spending across the entire task and enforce total budgets and attempt limits. Repeated failed attempts must be distinguished from completed purchases to block duplicate charges. Receipts must connect spending to the underlying task so users can see whether their money purchased useful information or funded an infinite loop.

Under x402’s exact-payment scheme, executed transfers are irreversible, though sellers can issue refunds by sending money back in new transfers. The batch-settlement variant holds funds in escrow under specified release conditions, and refunds depend on the seller’s policy and any contractual arrangements. Basic token transfers have no card-style dispute process built in, so the surrounding service must provide its own complaint and repayment mechanism.

Dispute resolution requires evidence beyond Payment receipts

A successful payment proves money moved, but resolving a purchase requires proof of what was actually bought.

The hotel booking scenario illustrates this gap. A payment receipt confirms the transaction completed, but the customer needs evidence of the advertised room specifications, the cancellation terms, and the instruction that authorized the booking. The payment network can establish that funds transferred, but resolving a dispute may require coordination with the hotel, the booking platform, the payment provider, and the user. The Consumer Financial Protection Bureau defines when credit card issuers may reverse charges, but those processes do not guarantee refunds for every disappointing purchase, including ones delegated to an agent.

User experience determines adoption and trust

An agent that saves time only if it does not require hours of log review to justify its choices. Supervising the agent’s decisions like a junior purchasing department would defeat the purpose of automation, especially when every mistake requires reconstructing a chain of software decisions and contractual terms. Usability will ultimately determine whether consumers embrace agent-driven transactions or maintain direct control over payments.

Practical controls reflect the cost of failure. Refundable reservations could proceed automatically, while non-refundable bookings require explicit confirmation. Spending ceilings must include fees, not just base prices. Permissions should expire when the task ends. Receipts should identify merchants and explain purchases in terms customers recognize, so disputes begin with understanding what was bought rather than decoding payment addresses.

The commercial incentives embedding the agent’s choices must also be visible. A hotel recommender can rank options by user preferences, by payments from sellers, or by some combination. Automating the purchase makes those incentives invisible if the user never sees rejected alternatives. Buyers need to understand whether sellers can pay for placement and how those payments shape the options presented, alongside disclosure of any payments the agent makes to data services during its search.

Neither card networks nor stablecoins nor mandates nor dispute processes alone solve the core problem: translating an ordinary human request into precise, machine-executable instructions that preserve the user’s intent across a chain of decisions no single human will review. The question ahead is whether payment systems can incorporate all these pieces simultaneously while remaining simple enough that users will actually delegate authority to agents rather than continuing to book hotels themselves.

BlockWest is a news publication. Nothing here is investment advice. Read our disclaimer and editorial policy.