AI can tell a user which coffee shop nearby is the best. It still may not be able to pay for the order on its own.
Atul Khekade, co-founder of XDC Network, used that example at a late-July event in New York to describe what he sees as the missing layer in today’s AI stack: transactions. More than 100 bankers, technology partners and investors were in the room. His point was blunt. AI still does not have a proper transaction layer.
That gap is now drawing attention from both traditional payment networks and crypto-native builders. Visa launched Intelligent Commerce Connect on April 9. Mastercard introduced Agent Pay for Machines on June 10. On the onchain side, Coinbase-incubated x402 has already been running for a year, while XDC said it connected Anthropic’s Claude to blockchain payments in July and showed a live demo in which an AI agent ordered and paid for a coffee.
Visa and Mastercard are taking different routes
Visa is trying to solve distribution with a single integration layer that works across multiple standards.
According to the report, Intelligent Commerce Connect is designed to be network-agnostic, protocol-agnostic and custodian-agnostic. A merchant that integrates once can accept payments initiated through four agent protocols at the same time: Trusted Agent Protocol, Machine Payments Protocol, Agentic Commerce Protocol and Universal Commerce Protocol. It supports both Visa and non-Visa cards. The product is now in a pilot phase with seven partners, including Aldar, Amazon Web Services, Highnote, Mesh and Payabli, with a broader rollout planned during 2026.
Mastercard is taking a sharper angle. Agent Pay for Machines is not centered on AI buying things for humans. It is aimed at AI buying from AI, with an emphasis on high-frequency, very small-value payments executed at machine speed. Its settlement rails span cards, accounts and stablecoins.
The first partner group includes more than 30 companies. The list cuts across legacy payments and blockchain infrastructure, including Stripe, Adyen, Global Payments, Ant International, Coinbase, Alchemy, Polygon and Solana Foundation.
Mastercard Chief Product Officer Jorn Lambert described the category this way: machine payments let services buy and sell between agents at a completely different scale, with very large volume, very small payment sizes and very high speed. The system also introduces what it calls Verifiable Intent, effectively an identity credential for an agent that carries authorization rules so a merchant can see who approved a payment.
Crypto teams have already shown working demos
Khekade argued that card networks and banks cannot move that quickly. In his words, “banks and card networks can’t roll out this infrastructure overnight.”
XDC AI’s setup combines three parts. First is x402, the machine-payments standard open-sourced by Coinbase in 2025, which handles charging. Second is Circle’s USDC for fee-free settlement. Third is Anthropic’s Model Context Protocol, which connects AI assistants to services. The report said Circle enabled native USDC on XDC in August 2025.
XDC also pointed to institutional participation. Deutsche Telekom subsidiary MMS joined as a standby node in 2024. NTT DOCOMO GLOBAL became an institutional masternode validator in July this year. A joint venture tied to SBI Holdings is responsible for Asia-Pacific trade finance.
The report also raised a discrepancy. Khekade referred to “close to more than 200 major institutional nodes,” but that does not line up with public information. Publicly announced institutional validators number only in the 20s, and the network’s active masternode cap is 108.
The rails exist, but usage is still small
There is now enough activity to show that agent payments are more than a theory. The scale, though, remains limited.
A joint study by Visa and blockchain data platform Artemis found that x402 and MPP processed more than 110 million transactions over the past 12 months, totaling $136 million. That works out to an average of $1.24 per transaction. The data points to a real micropayments use case. It also shows how small the category still is when measured against card networks that clear trillions of dollars a year.
Other data points in the report add to the caution.
- Gartner said in June last year that more than 40% of agentic AI projects would be canceled by the end of 2027 because of uncontrolled costs, unclear business value and weak risk controls. The same research estimated that out of thousands of vendors claiming to build agentic AI, only about 130 actually fit the description.
- OpenAI and Stripe launched an online checkout product in fall 2025. By March this year, about 30 Shopify merchants had gone live, and the effort was later scaled back.
- The International Monetary Fund said in April this year that there is a structural mismatch: payment systems run on deterministic logic, while agentic AI is probabilistic.
- The UK Financial Conduct Authority added agent payments to its AI Live Testing program in April. Singapore’s central bank released a proposed set of safeguards in July requiring a checkpoint between an AI decision and the actual payment.
Identity and accountability remain the central problem
Several people cited in the report reduced the issue to trust.
t54 Labs CEO Chandler Fang put it plainly. When an AI agent shows up with $10 and an order, the merchant’s first question is, “Who are you? Who do you represent?” Concordium Chief Growth Officer Varun Kabra gave a more concrete example: if an agent books a flight on someone’s behalf, the airline still cannot verify which human behind the transaction is ultimately responsible.
Animoca Brands Executive Chairman Yat Siu pointed to another constraint on the other side of the payment flow. If an agent has a wallet, it is already acting as an autonomous economic participant. Banks, however, are unlikely to open accounts for AI software.
Two payment models are emerging
As outlined in the report, Visa’s Intelligent Commerce Connect functions as a protocol-agnostic integration layer, allowing merchants to accept payments from four agent protocols through one setup. Mastercard’s Agent Pay for Machines is built around high-frequency micropayments between agents and supports three settlement tracks: cards, accounts and stablecoins.
Khekade also said Fortune 500 companies would formally go live within the next two to three months, though he did not name any of them.