Cards and stablecoins will occupy different niches in AI agent payments. This conclusion was reached by experts from Visa and blockchain analytics firm Artemis.
The authors identified two categories of such transactions:
- Purchases made on behalf of a person: booking trips, managing subscriptions, and ordering goods. These transactions are comparable to regular online shopping and can be processed through payment cards;
- Frequent transfers of less than $1 between programs. For instance, an AI agent might pay for individual requests to online services, data retrieval, computing power, or the use of software tools.
According to Visa and Artemis, fixed fees make card infrastructure unprofitable for such micropayments. In Base and Solana, transaction fees are fractions of a cent.
The report examined two open payment protocols: x402 (primarily active in Base, Solana, and Polygon) and the Machine Payment Protocol (MPP). The latter supports settlements across blockchains and traditional payment systems.
The authors proposed a hybrid model where cards are used for authorization and purchases on behalf of people, while stablecoins are used for transactions between programs. Both payment methods can be applied at different stages of a single task.
According to Artemis data as of April 21, after excluding identified fraudulent and test transactions, approximately $15 million was processed through x402 across 109.6 million transactions. For MPP, the figures were about $25,000 and 115,000 transfers, respectively.
The initial data was significantly higher: $135.7 million and 178.3 million operations through x402, as well as over $38,000 and around 184,600 transactions through MPP. The difference is attributed to the exclusion of operations that Artemis classified as fraudulent activity or testing. The figures were calculated by the company, and no independent assessment is provided in the report.
The main unresolved issue identified by the authors is trust. It remains unclear who should be responsible for erroneous or unauthorized purchases: the owner of the AI agent, the platform operator, the model developer, or the seller.
Existing refund rules are designed for human transactions. There is currently no established mechanism for disputing thousands of automatic transactions per hour. Additional risks arise from malicious instructions that could redirect payments or cause the AI agent to make unplanned purchases.
Recall that in March, an analyst from Artemis excluded fraudulent transactions from the x402 statistics. After filtering, the transaction volume of AI agents turned out to be 15 times lower than Bloomberg's initial estimate.
In June, Alchemy and Visa launched the AgentCard service for purchases through AI agents within user-defined limits.
