Foreign media reports that AI's proxy payment system is moving from the demonstration phase into real business scenarios. The focus is not on allowing the software to freely use funds, but rather on enabling users or enterprises to set clear permissions first, after which the proxy will complete the payments within defined limits.

Grant permissions first before execution.

This type of model relies on "delegated authorization." Users do not hand over full control of their bank cards or accounts directly to AI agents; instead, they set limits on expenditures, eligible merchants, permitted categories of consumption, as well as time and geographical restrictions first.

For example, a user can request an agent to book a direct flight to Paris for next month, with a budget limit of no more than $600. The agent then searches for options, filters the results, and completes the payment automatically when a suitable one is found.

Visa indicates that its systems for proxy businesses can set expenditure limits and additional approval thresholds, while also identifying proxy identities and monitoring fraudulent transactions. MasterCard's Verifiable Intent adopts a similar approach, which is to record what operations the user has authorized and under what conditions the authorization takes effect.

Bank cards and stablecoins go hand in hand

The article argues that AI agents will not rely solely on a single payment channel. Whether to use bank cards, bank accounts, or stablecoins will depend on the type of merchant and the amount of the transaction.

Visa is extending the bank card system to machine payments through the open standard Machine Payments Protocol, which is driven by Stripe and Tempo. According to this design, agents can receive payment requests, complete authorization in a programmed manner, and settle transactions through bank cards, stablecoins, or other supported methods. MasterCard's Agent Pay for Machines also supports settlement via bank cards, accounts, and stablecoins.

This difference is particularly evident in different payment scenarios. Hotel bookings for a few hundred dollars are suitable for the bank card network, but if an agent only needs to pay $0.002 for a single API request, the cost structure of traditional card payments is no longer appropriate. Based on this, the article argues that stablecoins and on-chain payment protocols are more suitable for handling high-frequency, small-amount machine transactions.

Micro-payment and Liability Allocation

The AI proxy will continuously call API, cloud computing power, database, and model inference services. Rather than subscribing to each service individually, the proxy is more likely to pay per use, initiating payments only when resources are truly needed.

The x402 protocol launched by Coinbase is designed around this model. It utilizes the HTTP 402 "payment required" status code to allow the server to directly make a payment request in the response. Upon receiving this request, the proxy automatically sends stablecoins before obtaining the required resources. The typical process described in the article is as follows: the proxy requests data, the server requests a payment of 0.01 US dollars, and after the proxy makes the payment, the server then returns the content.

However, the article argues that the most difficult issue is not transfer itself, but rather the attribution of responsibility. If an agent misinterprets instructions, exceeds the limit in consumption, or is induced to make a payment by a malicious website, it is still necessary to have clearer rules to determine whether the responsibility should fall on the user, the AI provider, the payment network, or the merchant.

Focusing on this issue, payment institutions are placing emphasis on identity verification, authorization proof, and auditing capabilities. In the future, for a proxy payment, it may not only be necessary to prove who made the payment, but also to specify which agent executed it, who authorized it, what restrictions were applied, and whether the transaction exceeded the authorized scope.

Tests have already been conducted in India and Europe.

The article mentions that India is studying a framework that would allow AI agents to complete small payments through UPI, without the need for individual confirmation for each transaction. According to the data in the article, UPI processed 24.51 billion payments in August 2026, with a total amount of about 314 billion US dollars, making it an important test scenario for agent payments.

In terms of actual implementation, Santander Bank and Mastercard completed an end-to-end real payment in a regulated banking environment in March 2026, which was executed on behalf of AI. In June 2026, Worldline, ING and Mastercard completed another transaction in a production environment in Europe.

The article argues that the real change is not AI "having funds," but rather individuals and businesses beginning to entrust limited, auditable payment authority to software. Once these control measures are proven reliable, the payment process may gradually shift from manual user operations to backend steps within automated tasks.