Visa and data analytics firm Artemis have published a joint report saying AI agent payments have moved beyond proof-of-concept and into commercial use cases that can support real transaction flow. Drawing on on-chain data from the x402 and MPP protocols, the report says the sector has processed more than 100 million transactions this year.
AI agent payments extend beyond simply paying on a user’s behalf
The report divides AI agent commerce into two categories: macro commerce and micro commerce.
In macro commerce, AI agents carry out consumer tasks for people, such as booking flights or managing subscriptions. Payment amounts in those cases remain within ranges familiar to consumers, so existing credit card rails can still handle them. Micro commerce is different. It involves high-frequency, low-value payments between software systems, usually below $1 and often only a few cents.
The report says MPP, a machine payments protocol launched in mid-March 2026 and built by Stripe and Tempo with contributions from Visa, processed about 115,000 transactions in its first weeks and settled about $25,000.
x402, which launched earlier in May 2025, posted much larger numbers. According to the report, it handled about 109.6 million transactions, with adjusted transaction volume of roughly $15 million, mainly across Base, Solana and Polygon.
Why micro-payments need a different rail
The report uses a simple comparison to explain the problem. Average payment size on the two protocols is only a few cents. On traditional card rails, the fixed fee on each transaction can exceed the payment itself, which makes the model uneconomic.
Visa’s data team says two shifts happened at the same time. AI agents with budget management capabilities created steady payment demand, while a new generation of blockchains pushed settlement costs down to fractions of a cent. Together, those changes made payments between $0.01 and $1 workable for the first time.
The report also says the x402 Foundation, established in July last year, already has 40 members, including AWS, Google, Visa, Mastercard, Stripe, Coinbase and Circle. That, in the report’s view, shows the industry is moving faster to align on standards.
Credit cards and stablecoins are not in a zero-sum contest
The report’s conclusion is direct: the future will not be a case of credit cards replacing stablecoins or stablecoins replacing credit cards. Credit cards are better suited to macro-level agent spending, while stablecoins fit machine-to-machine micro-payments more naturally. In live usage, both can appear at different stages of the same task.
The report also says the payment logic applies in Taiwan. With e-commerce transactions there led by mobile payments, stablecoins could show a clearer cost edge if AI agents begin handling small API calls and cloud computing charges on their own.
It adds that Taiwan’s banking sector is actively testing stablecoin settlement. If AI agents later start making automatic payments for logistics, inventory management or small cross-border procurement, a dual-rail model combining cards and stablecoins could become the most likely path to deployment.
Liability remains the main unresolved issue
The report also identifies accountability as the biggest blind spot in AI agent payments.
If an agent buys the wrong item or a malicious prompt redirects its budget, it is still unclear who should bear responsibility: the user who assigned the task, the platform operating the agent, the model provider, or the merchant receiving the funds.
According to the report, current refund rules and dispute resolution frameworks were built for human behavior. If agents transact thousands of times per hour and funds move across multiple layers of agents, tracing an erroneous payment becomes much harder.