Visa and Artemis are drawing a sharper boundary around the emerging economy of autonomous AI agents. Their joint research argues that cards remain well suited to consumer-sized purchases, while stablecoins may become the natural rail for software buying data, computing power and application access in tiny increments. The surprising divide is not between old finance and crypto, but between human-scale and machine-scale commerce. Traditional payment systems were designed for occasional purchases with recognizable customers and merchants. Agents may instead generate thousands of transactions rapidly, exposing cost, identity and accountability assumptions that were never built for autonomous commercial activity at scale.
Stablecoins Fit the Economics of Machine-Native Payments
The economics become clearest at the smallest ticket sizes. Visa and Artemis separate macro-commerce, such as booking travel or managing subscriptions, from micro-commerce involving frequent payments between software systems. Fixed card charges can exceed the value of a sub-dollar transaction, while newer blockchains can settle payments for fractions of a cent. Stablecoins therefore appear strongest where conventional payments become mathematically inefficient. An agent purchasing one API response or a brief unit of computing capacity does not need a familiar checkout screen. It needs programmable money, predictable finality and a cost structure that remains viable when volume rises dramatically over time.
Live protocol data suggests that this activity has moved beyond theory, although scale remains uneven. Coinbase-incubated x402 processed about $15 million in adjusted volume across 109.6 million transactions after launching in May 2025. Stripe and Tempo’s newer Machine Payments Protocol settled roughly $25,000 through about 115,000 transactions during its first weeks. The transaction counts are enormous relative to the money moving through them. Average payments were fractions of a cent, demonstrating why machine commerce behaves differently from retail shopping. Yet these figures also reveal immaturity: experimental protocols are processing intense activity without producing conventional payment volumes or established commercial economics.
Trust Infrastructure Remains the Harder Problem
Artemis’ data cannot resolve the harder institutional questions surrounding delegated spending. When an agent buys the wrong service, follows a malicious prompt or pays another agent through a complex chain, responsibility becomes difficult to assign. The consumer, platform, model developer and merchant may each control only part of the outcome. Faster settlement does not answer who carries liability when autonomous execution fails. Chargebacks, evidence standards and dispute windows were designed for human-paced transactions linked to identifiable orders. Reversing thousands of interconnected payments could be difficult and uncertain, particularly when stablecoin transfers settle rapidly and may cross jurisdictions without consistent protections.
Visa expects convergence rather than a winner-take-all contest. Cards can preserve authorization, fraud management and merchant acceptance for larger proxy purchases, while stablecoins can provide machine-native settlement for microscopic transfers. Hybrid workflows may use both within assignments, with an agent paying for data onchain before completing a customer purchase by card. The future payment stack may combine familiar trust systems with unfamiliar transaction frequency. Still, interoperability, identity, permissions, audit trails and accountability must mature before autonomous commerce reaches dependable scale. Stablecoins solve an important economic constraint, but Artemis’ findings show that cheap payments alone cannot create a trustworthy machine economy.
