AI Payments infrastructure reached a new threshold as Visa, Mastercard, American Express, and Stripe all went live with AI agent cards across four of the world’s largest payment networks on August 13, 2026, marking the first time autonomous software can hold, authorize, and dispute its own spending credentials without a human co-signer on every transaction.
Key Takeaways
Visa, Mastercard, American Express, and Stripe went live with AI agent cards on August 13, 2026
AI agent cards use dynamic rules that evaluate context at the moment of each transaction rather than fixed limits
Stripe’s pricing is flat-rate at 2.9% plus $0.30 per transaction with no agent-specific surcharge
None of the four networks has published a policy transferring liability to an agent’s operator or model provider
Visa, Mastercard, American Express, and Stripe have each built distinct implementations. A CoinGecko guide published on August 13 compares their fee structures, reusability rules, and merchant approval logic in detail.
How AI Payments Through Agent Cards Actually Work
AI agent cards are virtual payment credentials issued specifically to software agents rather than to humans or corporate entities.
The card is provisioned to an agent’s runtime environment, and the agent presents it at checkout, just as a browser autofills a saved card.
Unlike traditional virtual cards, which carry spending limits set once at issuance, AI Payments through agent cards use dynamic rules that evaluate context at the moment of each transaction, checking whether the purchase fits the agent’s current task, who authorized the session, and whether the merchant category is permitted. This matters because an AI agent booking travel, ordering cloud compute, or purchasing API credits may make dozens of micro-decisions per hour.
Requiring human approval on each one defeats the purpose of automation.
Visa, Mastercard, Amex, And Stripe Do Not Build AI Payments The Same Way
Visa’s implementation works through its Intelligent Commerce program. An agent authenticates with a token rather than a card number, and Visa’s network evaluates each charge against a set of agent-defined spending rules before settlement.
Merchants never see raw card data.
Mastercard’s approach, built under its Agent Pay initiative, routes agent transactions through a separate credentialing layer that can be toggled on or off per session. A merchant receives a cryptographic signal confirming the agent has authorization from a named human principal, without that principal’s card number being transmitted.
American Express uses what it calls agent-ready accounts.
These are standard Amex corporate accounts extended with an API surface. A developer registers an agent identity, and Amex issues a virtual card number scoped to that agent’s session.
Disputes and refunds flow back to the human account holder, not to the agent.
Stripe’s model is the most developer-native. Its agent toolkit lets engineers embed spending credentials directly into an agent’s tool-calling stack.
The agent calls a “pay” function, and Stripe handles authorization, logging, and reconciliation. Stripe also offers a reusability flag: a credential can be single-use or session-persistent, depending on how the developer configures the integration.
How AI Payments Became A Battleground In 2026
The commercial case for AI Payments did not exist at scale until large language models moved from chat interfaces into task-completion frameworks.
That shift accelerated in late 2025 as enterprises deployed agents to handle procurement, software licensing, and logistics.
The problem that emerged quickly was payment friction: an agent authorized to book a server for a compute job would halt and wait for a human to approve the charge, killing the efficiency gain the agent was meant to deliver. Payment networks saw the bottleneck and moved fast.
Visa’s Intelligent Commerce program launched in early 2026, and Mastercard’s Agent Pay followed within weeks. Both described their systems as infrastructure for the agentic economy, a phrase that has since become standard in fintech product announcements.
Stripe entered later but with a deeper developer hook.
Its agent toolkit integrates directly with popular agent frameworks, meaning a developer building on LangChain, AutoGen, or similar platforms can add AI Payments capability in a few lines of code.
Also Read: MetaMask Agent Wallet, the Dangerous AI Revolution in Crypto
The Fee And Reusability Gap Between Networks
The four networks differ most sharply on two commercial variables: transaction fees and credential reusability. Visa and Mastercard charge fees consistent with their existing virtual card rails, typically 1.5% to 2.9% per transaction depending on merchant type.
Agent transactions are not yet priced differently from human-initiated ones, though both networks have signaled tiered pricing is under review.
American Express fees sit at the high end of that range but come with tighter fraud protection, and an agent-initiated charge that fails dispute review gets refunded faster under Amex’s framework because the corporate account structure simplifies chargeback routing. Stripe’s pricing is flat-rate at 2.9% plus $0.30 per transaction for standard accounts, with no agent-specific surcharge.
The reusability advantage is Stripe’s clearest differentiator in the AI Payments space. A developer can issue a credential that persists across an entire agent work session, meaning a multi-step task generates one authorization event rather than dozens.
Single-use credentials, used by default in Visa’s and Mastercard’s current implementations, reduce fraud exposure but multiply authorization overhead.
For an agent making 50 purchases in an hour, that overhead is measurable.
What Comes After The First Generation Of AI Payments Infrastructure
When an AI agent card is used fraudulently, or when an agent makes a spending error, the human whose account backs the credential absorbs the loss under current frameworks. None of the four networks has published a policy that transfers liability to the agent’s operator or to the model provider.
That gap will close under regulatory pressure: the EU’s Payment Services Directive already covers virtual card products, and regulators in the UK and US are watching agent-initiated spending closely as agentic AI deployments scale.
The deeper shift is structural. AI Payments are one component of a broader stack that includes agent identity, credentialing, and audit trails.
As agents become persistent rather than session-scoped, the card infrastructure needs to track spending across days, not minutes.
Mastercard has indicated its Agent Pay roadmap includes persistent identity tokens, while Stripe has not commented on multi-session credential plans. With AI agent deployments scaling from thousands to potentially millions, Mastercard’s and Stripe’s competing bets on persistent versus session-scoped credentials may determine which network becomes the default AI Payments rail embedded in popular development frameworks, a structural position that compounds with every new agent deployment.
Read Next: The BIP-110 Bitcoin Chain Split Just Happened, The Rebels Got Stranded