Experts from Visa and the blockchain analytics company Artemis have presented a detailed analysis of the future of payments for AI agents. The key conclusion: traditional cards and stablecoins will occupy different but complementary niches in this ecosystem.

The study's authors identified two main categories of transactions conducted by AI agents. The first is purchases made on behalf of a human: booking travel, managing subscriptions, ordering goods. These operations are inherently similar to regular online commerce and can be efficiently processed through payment cards. The second category consists of frequent but extremely small transfers between programs, costing less than $1. This involves paying for individual requests to online services, access to data, computing power, or the use of software tools.

It is here, according to the analysts, that the main advantage of stablecoins lies. The fixed fees charged by card infrastructure make such micropayments economically unviable. At the same time, transaction fees on the Base and Solana networks amount to fractions of a cent, paving the way for mass adoption by AI agents.

The report examines in detail two open payment protocols: x402, whose activity is concentrated on Base, Solana, and Polygon, and the Machine Payment Protocol (MPP). The latter supports settlements both through blockchains and through traditional payment systems. According to Artemis data as of April 21, after filtering out fake and test transactions, approximately $15 million passed through x402 across 109.6 million transactions. MPP's figures are significantly more modest: roughly $25,000 and 115,000 transfers. For comparison, the raw "dirty" data for x402 amounted to $135.7 million and 178.3 million operations, highlighting the importance of data filtering when analyzing real activity.

The authors suggest that a hybrid model will emerge in the future: cards will be used for authorization and large purchases on behalf of humans, while stablecoins will be used for fast and cheap settlements between programs. Both payment methods could be applied at different stages of executing a single task.

The main unresolved problem: trust and responsibility

Despite technological progress, the key barrier remains the issue of trust. It is still unclear who should bear responsibility for an erroneous or unauthorized purchase made by an AI agent: its owner, the platform operator, the model developer, or the seller. Existing payment reversal mechanisms (chargebacks) are designed for human actions and are not equipped to handle thousands of automatic transactions per hour. An additional risk is posed by malicious instructions capable of redirecting a payment or forcing the agent to make an unplanned purchase.

My expert commentary: The market for payments for AI agents is in its infancy, and Artemis's data shows how important it is to critically evaluate sensational figures. The real transaction volume, cleaned of "junk," is 15 times lower than Bloomberg's initial estimates. However, it is stablecoins, with their minimal fees and programmatic controllability, that appear to be the most natural tool for this niche. The issues of responsibility and security are not a technical but a legal and institutional challenge, without the resolution of which mass adoption will be difficult.