Give AI Agents a Prepaid Card, Not Your Credit Card

Opening

An interesting service popped up recently. It’s a startup called AgentCard, and what it does is pretty straightforward: it issues prepaid Visa cards built specifically for AI agents. You load whatever amount you want, hand the card number to your AI agent, and the agent handles the actual online payment itself.

Your gut reaction might be, “Wait, you’re giving a card to an AI?” But that instinct is actually the whole point of the service. Instead of handing over your main card wholesale, each card is capped at a fixed amount — so you can say, in effect, “You get up to ₩50,000 for this task, and no more.”

Honestly, at this point it feels like the solution arrived before the problem did. But the mere existence of this service says something worth paying attention to. Today I want to use this one prepaid card as a thread to pull on, and look at the bigger picture of AI agent payment infrastructure.

What It Takes to Give an Agent “Payment Authority”

AgentCard’s mechanics are simple. You sign up through a CLI¹​, create a card, and load it with a dollar amount. In return, you get an actual Visa card number, CVV, and expiration date. Hand that information to an AI agent, and it can pay anywhere Visa is accepted. It also supports MCP²​, so it plugs directly into tools like Claude Code.

The core design principle here is isolation. Each card can never spend more than what’s been loaded onto it. You can issue Agent A a $5 card and Agent B a $50 card. Even if an agent tries to make a payment somewhere it shouldn’t, the damage is capped at that card’s balance.

This is actually a pattern that’s already well established in corporate expense management. Corporate card services like Brex and Ramp built exactly this structure — issue a virtual card per team member, isolate the spending limit, track expenses in real time. AgentCard just swaps the “person” for an “agent.” Loading a fixed budget onto a marketing team’s ad account is a similar concept.

But there are real limits to this right now. Very few workflows today are trusted enough for an AI agent to make independent payment decisions on its own. In a survey conducted by Harvard Business Review in December 2025, only 6% of organizations said they fully trusted AI to handle a business process from start to finish without oversight. 94% still limit agents to low-risk tasks that can be supervised.

Big Tech Is Playing a Different Game

If AgentCard’s approach is “give agents limited authority via prepaid cards,” the giants of payment infrastructure — Visa, Mastercard, and Stripe — are building something far more structural.

Visa announced an initiative called Visa Intelligent Commerce in April 2025 — infrastructure that lets AI agents search for products, make recommendations, and complete payments directly on top of the Visa network. In October 2025, it followed up with an open framework called the Trusted Agent Protocol, a standard that lets merchants tell the difference between “an AI agent with genuine purchase intent” and “a malicious bot.” Visa expects millions of consumers to complete purchases through AI agents by the 2026 year-end shopping season. More than 100 partners worldwide are currently participating, and hundreds of agent-driven transactions have already gone through in live environments.

Mastercard is making a similar move under the name Agent Pay. Its core concept is the Agentic Token — essentially applying the tokenization technology used in mobile payments to agentic commerce. Each agent carries its own unique cryptographic credential, which it uses to initiate transactions. Mastercard has partnered with Microsoft, Cloudflare, Google, and PayPal on this.

And then there’s the move that really matters. In September 2025, Stripe and OpenAI jointly announced an open standard called the Agentic Commerce Protocol (ACP). The Instant Checkout feature inside ChatGPT — letting you discover a product and complete the purchase on the spot — runs on top of this. It launched first with Etsy sellers in the US, and is now expanding to Shopify’s 1 million-plus merchants. At the heart of it is a new payment primitive³​ Stripe built called the Shared Payment Token (SPT), which lets an agent complete a payment without ever seeing the user’s actual payment method.

Three Approaches, One Question

If I had to sum up what’s happening right now, three approaches to agent payment infrastructure are competing for dominance.

First, the prepaid-card model (AgentCard and similar services). This uses the existing Visa/Mastercard network as-is, but limits risk by isolating the amount available. The upside is that it works right now — anywhere Visa is accepted, the card works. The downside is that fundamentally, you’re still handing a card number over to an agent, which raises PCI compliance⁴​ issues and can trip bot-blocking mechanisms like CAPTCHAs or two-factor authentication.

Second, the platform-native model (Visa Intelligent Commerce, Mastercard Agent Pay, Stripe’s ACP). This redesigns the payment network itself to be agent-friendly — building new tokenized credentials, agent authentication protocols, and trust frameworks with merchants from scratch. It’s the most fundamentally sound direction, but it takes time, because the whole ecosystem has to move together.

Third, the protocol-based model — things like Google’s AP2 or Coinbase’s x402 — which tries to bypass the existing card networks entirely and enable direct payments between AI agents. x402 in particular repurposes the HTTP status code 402 (“Payment Required”) to handle micropayments between machines. For more on this, check out my earlier newsletter, <A World Where AI Agents Pay for Things — and Korea’s Door Is Closing Before It Even Opens>.

All three approaches ultimately need to answer the same question: who bears legal liability for an AI’s economic actions? If an agent makes a bad payment, who has the authority to reverse it? What procedure applies when disputes arise? Do consumer protection laws even apply? Technology is moving fast, but no one has a clear answer to this question yet.

Oz’s Lens

Watching all this unfold, I keep thinking back to the early 2000s and the dawn of online payments.

PayPal got a similar reaction when it first launched — “Wait, you’re typing your card number into the internet?” It’s second nature now, but back then, the trust problem was a much bigger obstacle than the technology problem. What ultimately closed that gap wasn’t the technology itself — it was institutional infrastructure: escrow systems, buyer protection policies, dispute resolution procedures.

Services like AgentCard are a classic “bridge solution.” They’re a stopgap that runs on top of existing infrastructure until the final destination — a truly agent-native payment system — is reached. Once Visa and Stripe finish building native agent payments, AgentCard’s reason for existing may simply disappear. But that “in-between time” could be longer than it looks. New payment standards always take longer to actually spread across merchants worldwide than the optimistic forecasts suggest.

What I find myself paying closer attention to is Korea’s situation. Stripe doesn’t officially operate in Korea, and PayPal’s domestic usage is limited — so when, and in what form, might global agent-payment standards like ACP or Visa Intelligent Commerce actually reach the Korean market? The one notable domestic move so far is NHN KCP joining as a partner on Google’s AP2. The Financial Services Commission is consolidating and revising its AI usage guidelines, but a concrete regulatory framework for “scenarios where an agent pays directly” hasn’t surfaced yet.

In the end, the real bottleneck for agent payments isn’t technology — it’s the design of trust. How much can you entrust to an agent? How do you undo a mistake once it’s made? Who audits that whole process? The institutions and standards that answer these questions have to come first, before the technology can move at its own pace.

Closing

For AI agents to move beyond searching and organizing information and start actually acting in the real world, they need a payment layer — and that payment layer has to look different from the card systems built for humans. Prepaid cards like AgentCard’s serve as an immediate bridge, while Visa, Mastercard, and Stripe are redesigning the underlying infrastructure from the ground up. And in between, the hardest problem of all — legal liability and consumer protection — remains unresolved.

If you want to dig deeper into this topic, I’d recommend reading through Stripe’s published Agentic Commerce Protocol spec directly. It’s a technical document, but it lays out a concrete design philosophy for exactly what steps an agent should go through when making a payment — which helps in understanding where this whole space is headed.

References & Further Reading

• AgentCard, “Prepaid Virtual Cards for AI Agents”, 2025. : The service that sparked today’s newsletter. You can check out the CLI-based issuance process and MCP integration directly.
• Stripe, “Developing an open standard for agentic commerce”, 2025.09. : Explains ACP’s design philosophy and how the Shared Payment Token works.
• OpenAI, “Buy it in ChatGPT: Instant Checkout and the Agentic Commerce Protocol”, 2025.09. : OpenAI’s account of the background behind Instant Checkout in ChatGPT and the open-sourcing of ACP.
• Visa, “Visa and Partners Complete Secure AI Transactions”, 2025.12. : Covers real completed transactions under Visa Intelligent Commerce and its global expansion plans.
• Mastercard, “Mastercard unveils Agent Pay”, 2025.04. : Mastercard’s official announcement, including the Agentic Token concept and its collaboration with Microsoft.
• Proxy, “AI Agent Payments Landscape 2026”, 2026. : A full map of the agent payments ecosystem as of 2026, covering wallets, identity layers, card issuance, and network standards.
• Crossmint, “AI Agent Payments: Traditional Payment Methods vs. New Solutions”, 2025.04. : Breaks down exactly why traditional payment methods don’t fit agents, examining the limits of stored cards, shared sessions, and virtual cards.
• CIO Korea, “How Financial Sector CIOs Are ‘Safely’ Adopting AI Agents”, 2026.02. : Includes the HBR survey finding (6% full trust in AI) and shows the realistic state of agent adoption in the financial sector.
• Portone Blog, “2026 Global Payment Trends: The Age of AI Agentic Commerce Arrives”, 2025.12. : A Korean-language rundown of competing protocols including Google AP2, ACP, and x402, including NHN KCP’s partnership with AP2.

The author, Kwangseob Ahn, is a professor of business administration at Sejong University and lead consultant at OBF (Oswarld Boutique Consulting Firm). He teaches statistics and data analysis — business data management and business analytics — while leading GTM and AI strategy consulting in the field, designing the seam between technology and business. He has published academic research on a memory architecture for AI dialogue systems (HEMA) and runs Daily Arxiv, a daily curation of global AI papers. He holds a master’s from Korea University’s Graduate School of Technology Management and a KMBA. He is the author of Homo Brainless: The People Who Outsource Their Thinking.

Footnotes

1. CLI (Command Line Interface): A way of operating a program by typing text commands instead of clicking with a mouse. Mainly used in developer tools. ↩

2. MCP (Model Context Protocol): A protocol that lets AI models communicate with external tools and services. It’s the standard AI tools like Claude and Cursor use to integrate with other services. ↩

3. Payment Primitive: The most basic building block of a payment system. Like Lego bricks, these can be combined to create a wide range of payment scenarios. ↩

4. PCI Compliance: An international security standard that every company handling credit card information must follow. It contains strict rules on how card numbers can be stored and processed. ↩