AI agents moved from recommending to executing. They compare products, select services and coordinate tasks on a user’s behalf, and every autonomous purchase depends on a payment system that can recognize the agent, enforce its spending authority and settle the transaction.
What emerged over the past year is not one payment layer but two, divided almost exactly along the consumer-versus-machine line.
Card networks took consumer checkout, because that is where fraud tooling, merchant relationships and consumer trust already live. Stablecoins took machine-to-machine, because a card network was never designed to move a fraction of a cent between two pieces of software.
The numbers make the division visible. Visa’s own research puts the x402 protocol at roughly $15 million in adjusted volume across 109.6 million transactions since launching in May 2025, with an average payment of a fraction of a cent. Those are not people buying headphones.
How AI Agents Become Buyers
An agent becomes a buyer when a user grants it a goal, a budget, and authority to act within defined limits. It evaluates offers, chooses one, and initiates payment without approval at every step.
A typical agentic transaction runs through four stages:
- Interpret the request. Identify the product or service, budget, and restrictions.
- Select an offer. Compare price, availability, and relevant terms.
- Verify authority. Check the purchase against the agent’s spending limit and approved conditions.
- Execute payment. Settle the transaction and record the authorisation.
Crypto payment infrastructure already supports structured payment requests. A merchant can accept payments with crypto links, with ChangeNOW providing an example of a flow connecting a defined payment request to crypto settlement. An agent can fold that instruction into a larger purchasing workflow.
Merchants need machine-readable product information, pricing, and availability. Payment systems need to identify the agent and verify its authority. Visa’s Trusted Agent Protocol, launched in October 2025 in collaboration with Cloudflare, is one example of payment infrastructure being adapted for this.
Why the Payment Layer Had to Change
Once an agent can initiate a purchase, the payment system has to decide whether the transaction is permitted. That creates a control layer between the agent’s decision and the movement of funds.
Establishing Who Is Acting
The system needs to know who is transacting and what that agent may do. Limits can apply to amount, merchant category, payment method or time period.
Both card networks converged on cryptographic proof that a cardholder authorised the agent, while differing on when consent is captured. Mastercard’s Agent Pay, launched in April 2025, uses Agentic Tokens carrying mandates defined once at provisioning, validated by the issuer against each subsequent purchase. Visa extended its token service with agent credentials plus a signed-intent payload accompanying each authorisation.
The route matters too. An agent may use cards, bank transfers or blockchain settlement, each with different conditions. The same infrastructure businesses use to accept crypto payments can support agent transactions where it also handles settlement and accounting.
Security and Accountability
Valid credentials cannot prevent every automated payment risk. A compromised agent still holds access, and misleading product information can steer its decision. Monitoring therefore has to weigh transaction behaviour alongside authorisation.
That matters more as attackers automate. AI-driven credential attacks compress the window between compromise and use to something no manual review catches, and an agent credential with standing spend authority is a valuable target.
Records matter equally. When a payment is disputed, businesses need to see what the user permitted, what the agent did, and how settlement was completed. Approval gates and defined limits belong in the agent architecture before deployment rather than after the first incident.
Where Crypto Changes the Economics
Agentic commerce generates frequent, low-value, cross-border payments between software systems. That makes per-transaction cost decisive, particularly when an agent pays for a resource used once.
Stablecoins as a Settlement Asset
Stablecoins give agents a digital payment asset pegged to a reference currency such as the US dollar. That suits services priced in fiat while avoiding the volatility of assets like Bitcoin.
They also work across markets. An agent buying data, software access or compute can settle in a common asset without both parties sharing a local currency.
Card networks noticed. Mastercard agreed in March 2026 to acquire stablecoin platform BVNK for up to $1.8 billion, and Visa has extended settlement across multiple chains.
Paying for Digital Resources on Demand
The x402 protocol demonstrates the model. Incubated by Coinbase and Cloudflare, it reactivates the long-dormant HTTP 402 “Payment Required” status code: an agent hitting a paid API receives a 402, settles the charge, and retries the request, with no account signup or card entry.
This suits API calls, data queries, and compute. Rather than packaging every service into a subscription, providers charge for individual usage.
The protocol moved from experiment to infrastructure this year. Coinbase contributed it to the Linux Foundation, and on 14 July 2026 the x402 Foundation launched with 40 member organisations, including premier members Visa, Mastercard, Stripe, Google, AWS and Coinbase. Amazon integrated x402 into its Bedrock AgentCore Payments service, where settlement completes in roughly 200 milliseconds on Base for a fraction of a cent.
A competing machine rail also emerged. The Machine Payments Protocol, built by Stripe and Tempo with Visa contributions, went live in mid-March 2026 and settled around $25,000 across roughly 115,000 transactions in its first weeks. Much newer than x402, though its two-party design and multi-rail plans make it worth watching.
What Autonomous Commerce Looks Like in Practice
Three categories have formed, and the settlement rail differs across them.
Consumer Commerce
The clearest application is an agent completing a purchase after the customer sets requirements.
Visa documented a reference flow in which Skyfire powered a Consumer Reports product-recommendation agent that executed a Bose headphones purchase through browser automation in closed beta. OpenAI’s Instant Checkout, live since September 2025 and built on a protocol co-developed with Stripe, settles over card rails using a shared payment token scoped to one merchant and cart, starting with Etsy and extending across more than a million Shopify merchants.
Checkout becomes part of the agent’s workflow. The customer still owns the purchase decision while software handles the steps between selection and completion.
Business Procurement
Corporate spending offers a wider set of routine transactions: recurring purchases, software renewals, invoices and expenses following established rules.
Card networks are building toward this through their agentic programmes, and the appeal is straightforward. Where vendor, amount and conditions are already known, the manual handling adds little.
Machine-to-Machine Commerce
The distinct model appears when software is both buyer and consumer.
This is where stablecoin rails dominate, and the volume data shows why. An agent paying for compute, a data feed, or an API call at fractions of a cent, thousands of times an hour, operates at a price point and frequency cards were never built for.
A service becomes a buyer of another service, with payment occurring inside the interaction between two systems rather than at a human-facing checkout. Agent traffic of this kind also breaks assumptions older network monitoring was built around, so the payment layer is not the only thing needing rethinking.
What Could Slow Adoption
Capability is not the constraint. Liability, cost, and willingness to delegate are.
Accountability Is Still Unsettled
An autonomous transaction can involve the customer, the agent provider, and the merchant. When a purchase produces a dispute, an unsuitable order or a loss, responsibility can be hard to assign.
The imbalance is visible in what shipped. Card networks and crypto rails released protocol after protocol through 2025 and 2026. Comparatively little arrived on dispute rules, refunds, and consumer protection, and those will shape how businesses actually use agent-led transactions.
Automation Has to Pay for Itself
Agentic commerce requires integration, maintenance, and oversight. The economics vary by task.
Automating thousands of routine transactions can justify the investment. A process taking a person a few minutes usually cannot.
Delegation Has a Natural Limit
Some decisions delegate more comfortably than others. A business may let an agent reorder supplies while keeping major purchases under human control. Consumers draw similar lines between routine spending and expensive or unfamiliar purchases.
Adoption will range from narrow routine tasks toward more complex ones, and liability, cost and comfort will decide where that boundary settles.
When Commerce Becomes a Software Process
Agentic commerce changes where commercial activity happens. A purchase becomes one step inside a larger software process, sitting alongside product selection, service access, procurement or resource management.
That gives payment a different role. It becomes an action software initiates while completing a task, with people remaining responsible for the goals and commercial policies shaping those tasks.
McKinsey has estimated agentic commerce could drive $3 trillion to $5 trillion in global sales by 2030. Treat that as a projection rather than a measurement, though the direction is not seriously disputed.
FAQs
Q. What are AI agent payments?
Transactions initiated by software acting on a user’s or business’s instructions, within defined authority.
Q. How can AI agents pay for products and services?
Through cards, bank transfers, or blockchain settlement, depending on the merchant and transaction requirements.
Q. Why did agentic commerce split into two rails?
Cards carry consumer purchases because fraud tooling and merchant trust already exist there. Stablecoins carry machine-to-machine payments because card economics do not work at fractions of a cent.
Q. Why are stablecoins relevant here?
They provide a digital payment asset pegged to a reference currency, useful for cross-border transactions and services priced in fiat.
Q. What is x402?
An open protocol reviving the HTTP 402 status code so an agent can pay for a web resource and retry its request. Contributed to the Linux Foundation, with a governing foundation launched in July 2026.
Q. What is machine-to-machine commerce?
Commerce between software systems, where one pays another for resources such as data, compute or software access.
What could limit adoption?
Unresolved liability, implementation cost, and how much decision-making users will delegate.
Related: AI Academic Agents Can Read the Papers You Can’t
| Disclaimer: This article is for informational purposes only and does not constitute financial, legal or investment advice. AI agents and autonomous payment systems involve technical, financial and regulatory risks. Readers should assess these risks and verify relevant information before using any related service. |
