Bloomberg ran a headline last week: "Stablecoin Firms Bet Big on AI Agent Payments That Barely Exist." The same week, a16z published a thesis arguing stablecoins will own B2B agent payments. American Banker ran an op-ed saying card rails physically cannot handle AI-driven transaction volume.

Crypto Twitter is declaring victory. Fintech Twitter is pushing back. Everyone is picking a side.

The debate is wrong. "Agent payments" are at least three different kinds of transactions. Cards win one of them. Stablecoins win another. The third is a toss-up. The question that actually matters is who routes each transaction to the right rail.

Cards Work for Consumer Purchases. That Won't Change Soon.

A number worth grounding the entire conversation around: global e-commerce did $6.88 trillion in 2025. Nearly all of it ran on card rails.

Coinbase's x402, the most mature stablecoin payment protocol for agents, processed $24 million in the past 30 days. That's 0.00035% of card commerce. Stablecoin agent payments are early. The market they're targeting barely exists yet.

When your AI agent orders you dinner, it'll charge your Visa. Every restaurant on DoorDash accepts Visa. If the agent orders the wrong thing, you can dispute the charge and get your money back. Your bank's fraud team is watching for suspicious activity. Reg E and PSD2 give you legal protections. You've had a Visa since you were 18 and you trust the system.

None of that exists for stablecoins. USDC has no chargeback mechanism. There's no "dispute this transaction" button on a blockchain settlement. The GENIUS Act might create a regulatory framework for stablecoins eventually, but it isn't law yet. And 99% of merchants don't accept USDC.

IBM and Worldpay found that 45% of consumers are comfortable letting an AI agent shop for them. That comfort is built on an assumption: their card's fraud protections still apply. The agent is a proxy for the cardholder, not a new financial system. Switch the rail to USDC and you're asking people to trust the AI and trust a payment method they've never used. Two trust barriers instead of one.

For any transaction where a human is the end customer and a traditional merchant is the seller (groceries, flights, clothes, subscriptions, restaurant orders) cards are the obvious starting rail. Visa TAP gives agents cryptographic authentication. Mastercard Agent Pay issues scoped tokens with spending limits and merchant restrictions. Stripe's SPTs create single-purpose payment tokens for each agent transaction. The infrastructure is shipping now.

Cards for consumer purchases is the right answer. Full stop.

Where Cards Break: The Fee Math

Cards work when your agent buys a $150 flight. They fall apart when your agent buys a $0.05 API call.

Stripe charges 2.2% plus $0.30 per transaction. On a $150 flight, that's $3.60. About 2.4%. Fine.

On a $0.05 API call, that's $0.30 plus a fraction of a cent in percentage fees. You're paying $0.30 to move a nickel. A 602% fee.

Scale it up. An AI agent making 200 micro-purchases per day at $0.05 each spends $10. Card processing fees on those 200 transactions: $60.20. Six times more in fees than the total value of everything purchased.

This is not fixable with volume discounts. The $0.30 fixed fee is structural. It exists because card authorization bounces between four parties (issuer, network, acquirer, merchant) and each hop has a cost floor. You can negotiate the percentage down. You cannot negotiate away the architecture.

Speed is the second problem. Cards settle in T+2: two business days. Visa and Mastercard are pushing toward faster settlement, but the four-party model has latency in its design. A few seconds of authorization delay is fine for a human at checkout. For an agent executing 1,000 operations per hour, two days of settlement delay means capital locked across thousands of unsettled transactions simultaneously. The accumulation creates drag on every downstream workflow.

American Banker put it directly in January: card rails are "pull" payments designed for human speed. When agents transact at machine speed, the rails become a bottleneck.

Then there's raw volume. One agent making 1,000 transactions a day is a rounding error for the networks. A million agents each making 1,000 transactions a day is a billion new daily authorizations hitting infrastructure designed for a few hundred million human-initiated ones.

Three Types of Agent Transactions

The conversation clears up when you stop treating agent payments as one category.

Type 1: Agent-to-Merchant. Your agent buys you something from a store. Groceries, plane tickets, headphones. The buyer is a human (through an agent). The seller is a traditional merchant. Transaction values range from $10 to $2,000. Frequency: a few times a day.

Cards win here. Merchants accept them. Consumers get fraud protection and dispute rights. Regulators have decades of precedent. Visa TAP, Mastercard Agent Pay, and Stripe SPTs are already live and handling these transactions. There is no reason to replace a working system with a payment rail that merchants don't accept and consumers don't trust.

Type 2: Agent-to-Agent. Software paying software. An agent buying API calls from another agent's service. An agent paying for compute time or data access or inference credits. No human on either side of the transaction. Values from $0.001 to $1.00. Frequency: hundreds or thousands per day.

Stablecoins win here, and the gap is wide. The fee math makes cards impossible. The speed math makes cards impractical. The transaction type doesn't need human-oriented protections. There's no consumer to protect. If an agent pays for a service and doesn't receive it, the agent doesn't file a chargeback. It stops calling that API.

Coinbase's x402 protocol was built for exactly this. An agent requests a resource. The server responds with HTTP 402 and a price in USDC. The agent signs a payment, attaches it to the request header, and the transaction settles in under a second on Base. No checkout page. No card number. No four-party hop. The x402 ecosystem has processed 15 million transactions total and handles 156,000 weekly, growing 492% year over year. Small compared to cards. But this is the only category where stablecoin agent payments have real product-market fit today.

Type 3: Business-to-Business. An agent negotiating vendor contracts. Placing wholesale orders. Settling invoices. Purchasing enterprise SaaS seats on behalf of a company. Values from $500 to $500,000. Low frequency, but each transaction is complex: approval chains, multi-party settlement, sometimes cross-border.

This one is genuinely contested. a16z's Sam Broner made the stablecoin case in February with a piece called "Tourists in the Bazaar." His argument: think about a physical bazaar. Tourists use cash or card for one-off purchases. The locals (the restaurant owner buying from the butcher, the tailor paying the weaver) use credit, relationships, and standing accounts. They don't pay per-transaction tourist rates.

AI agents doing B2B procurement will look like locals. They'll build recurring vendor relationships. They'll negotiate terms. The first transaction might use a card. The hundredth transaction with the same vendor will use whatever rail is cheapest and fastest. For cross-border settlement with no FX fees and instant finality, that's stablecoins.

But the picture is messier than the thesis suggests. Enterprise procurement has approval chains, audit requirements, and regulatory compliance that stablecoin infrastructure hasn't fully solved. A CFO needs every transaction attributable, every approval documented, every settlement reconciled. Card networks and ACH have decades of tooling built for this. Stablecoin audit trails exist on-chain, but the enterprise tooling to parse and present them to a finance team is still early.

The honest answer for B2B: cards and ACH for now. Stablecoins increasingly over the next two to three years as enterprise tooling catches up.

The Convergence Is Already Happening

The clearest signal: look at what the companies are actually building.

Stripe, the company that built its business on card processing, raised $500 million through Paradigm to build Tempo, a blockchain dedicated to stablecoin payments. In February they integrated USDC payments on Base for AI agents via x402. The card-first company is building stablecoin rails.

Mastercard is simultaneously pushing Agent Pay (card-based agentic tokens) and investing in stablecoin infrastructure. They aren't picking a side because they can see the same split.

Coinbase's Erik Reppel predicts that by the end of this year, most people won't even know they're using crypto. An agent's balance goes down five dollars. The payment settles instantly on a stablecoin rail behind the scenes. The user just sees "payment confirmed." Virtual cards on the frontend. Stablecoin settlement on the backend.

That hybrid is probably where everything lands. The user doesn't care about the rail. The merchant doesn't care about the rail. The agent definitely doesn't care about the rail. The infrastructure picks the best one per transaction.

What This Means If You're Building

If you're building an AI agent that makes purchases, think in transaction types.

Your agent ordering a user's groceries needs card rails. Visa TAP or Mastercard Agent Pay for authentication. A scoped token with spending limits. The consumer protection framework that comes with cards. Building this on stablecoins would mean your users can't dispute charges and your merchants need to accept USDC. Neither is realistic today.

Your agent buying API calls from other services needs stablecoin rails. x402 or similar. Running those transactions on cards means paying 602% fees on a $0.05 purchase. The economics don't work at any scale.

Your agent doing enterprise procurement needs flexibility. Cards for quick purchases under approval thresholds. Stablecoins or ACH for large vendor settlements. An audit trail that satisfies your customer's finance team regardless of which rail settled the payment.

The worst architectural choice is picking one rail and building everything around it. An agent that can only pay with cards will burn money on micropayments. An agent that can only pay with stablecoins will be rejected by every merchant on Shopify.

Where Prava Fits

Prava's payment infrastructure supports both rails and routes between them. When an AI agent makes a purchase through Prava, the system picks the optimal rail based on the transaction: cards for consumer merchants, stablecoins for agent-to-agent micro-transactions, ACH for established business relationships.

We're working with open-source agent frameworks, teams building on OpenClaw, LangChain, CrewAI, and shipping SDKs and tooling so developers don't have to build payment rail selection themselves. One API call. The routing is handled.

The stablecoin maximalists are wrong about consumer commerce. Card maximalists are wrong about machine-to-machine. Both rails exist for good reasons. The infrastructure that wins is the one that uses both.