Foreign media report that AI agent payments are transitioning from demonstration stages to real-world business scenarios. The focus is not on allowing software to freely access funds, but rather on enabling users or businesses to set clear permissions first, after which the agent executes payments within those defined limits.
Grant permissions first, then execute.
This type of pattern relies on "delegated authorization." Users do not grant AI agents full control over their bank cards or accounts; instead, they first set spending limits, approved merchants, permitted categories of expenses, and time and geographic restrictions.
For example, a user can ask the agent to book a direct flight to Paris next month with a budget of no more than $600. The agent then searches for options, filters the results, and automatically completes the payment when suitable options are found.
Visa states that its system for agent businesses can set spending limits and additional approval thresholds while identifying agent identities and monitoring fraudulent transactions. Mastercard’s Verifiable Intent follows a similar approach, recording what actions users have authorized and under what conditions those authorizations take effect.
Bank cards and stablecoins operate in parallel
The article argues that AI agents will not rely solely on a single payment channel; the choice between credit cards, bank accounts, or stablecoins depends on the type of merchant and the transaction amount.
Visa is extending the card network to machine payments through the open standard Machine Payments Protocol, driven by Stripe and Tempo. Under this design, agents can receive payment requests, programmatically complete authorization, and settle via credit cards, stablecoins, or other supported methods. Mastercard’s Agent Pay for Machines also supports settlement via credit cards, accounts, and stablecoins.
This distinction is especially evident in different payment scenarios. A hotel booking worth hundreds of dollars is well-suited for card networks, but if an agent only needs to pay $0.002 per API request, the cost structure of traditional card payments becomes inappropriate. The article concludes that stablecoins and on-chain payment protocols are better suited for handling high-frequency, low-value machine transactions.
Micropayments and Responsibility Allocation
AI agents will continuously call APIs, cloud computing resources, databases, and model inference services. Instead of subscribing to each service individually, agents are more likely to pay per use, triggering payments only when resources are actually needed.
Coinbase's x402 protocol is designed around this model. It leverages the HTTP 402 "Payment Required" status code to enable the server to directly request payment within its response. Upon receiving the request, the proxy automatically sends stablecoins to obtain the requested resource. The typical flow described in the article is: the proxy requests data, the server requires a payment of $0.01, and after the proxy makes the payment, the server returns the content.
However, the article argues that the most difficult issue is not the transfer itself, but rather accountability. If the agent misinterprets instructions, overspends, or is tricked into making payments by a malicious website, it remains unclear whether responsibility should fall on the user, the AI provider, the payment network, or the merchant—requiring clearer guidelines.
In addressing this issue, payment institutions are focusing on identity verification, proof of authorization, and audit capabilities. In the future, an agency payment may need to not only prove who made the payment, but also identify which agent executed it, who authorized it, what restrictions applied, and whether the transaction exceeded the authorized limits.
Testing has already been conducted in India and Europe.
The article mentions that India is developing a framework to allow AI agents to make small payments via UPI without requiring individual approval for each transaction. According to the data in the article, UPI processed 24.51 billion transactions totaling approximately $314 billion in August 2026, making it a significant testing ground for agent-based payments.
In practical implementation, Santander Bank and Mastercard completed an end-to-end real payment executed by an AI agent in a regulated banking environment in March 2026. In June 2026, Worldline, ING, and Mastercard completed another live transaction in Europe.
The article argues that the real change is not AI "holding funds," but rather individuals and businesses beginning to grant software limited, auditable payment permissions. Once these controls are proven reliable, payment processes may gradually shift from manual user actions to background steps within automated tasks.