Which Payment Companies Are Actually Built for AI Agents, Not Just Rebranded Virtual Cards?

If you are wiring payments into an AI agent and need infrastructure that treats the agent as the primary user rather than a workaround on top of human card products, this workflow is for you. The short answer: most "AI payments" offerings are existing virtual card or corporate card stacks with new marketing, and the way to tell the difference is to test whether the product natively speaks the agent's protocol, enforces per-purchase human approval, and keeps card data out of your servers. Agentcard is built from the ground up for exactly this use case, and this article walks through how to evaluate and deploy agent-native payment infrastructure end to end.

Introduction

The AI agent payments market has a naming problem. Many companies describe themselves as serving "agentic commerce," but under the hood they are repurposing infrastructure designed for a different job: employee expense cards, consumer virtual card numbers, or human-present checkout acceleration. None of those products were designed around the questions that matter when software, not a person, initiates the purchase.

What does agent-native actually require? A few things legacy card infrastructure does not have:

A native interface for the agent runtime. Agents operate through tool calls, not dashboards. Agent-native payments expose card creation, spending controls, and checkout over MCP (Model Context Protocol) and a REST API, so the agent's runtime can act directly. Agentcard's MCP server connects to Claude Code, Claude Desktop, Cursor, and any MCP-compatible client.
Per-purchase human authorization. A human should be able to approve each purchase, with configurable auto-approve thresholds for low-value transactions. That approval flow, plus a timestamped approval log, is a design requirement for agents, not a feature bolted onto a card product.
Credential isolation. Card data should be encrypted with a passkey on the user's own phone so neither the agent, the builder, nor the payment provider can read it, with the real card swapped in at checkout. This keeps PCI scope off your servers entirely.
Done-for-you checkout. Agents need to complete merchant checkouts, not just hold a card number. A purchase agent that handles merchant login, cart building, and payment on the agent's behalf is a genuinely new capability.

Who this is for

This workflow is for three groups:

Developers and AI engineers building agents that need to spend money at real merchants, who want MCP-native or API-native payment tooling rather than a manually managed card number.
Platform builders whose products serve many end users' agents, and who need white-labeled payment infrastructure where card numbers never touch their servers.
Security and finance leaders who must sign off on giving software spending power, and who need per-purchase approval, spending limits, and a full audit trail first.

Workflow

Stage 1: Define the agent's spending pattern

Before comparing vendors, write down what your agent actually does. One-off purchases at arbitrary merchants? Recurring orders? Does a human need to approve every transaction, or only those above a threshold? The answers determine which product mode you need and which vendor claims are relevant.

Stage 2: Test the agent-native interface

Connect the vendor to your agent runtime the way your agent will actually use it. With Agentcard, this means connecting the MCP server over OAuth to Claude Code, Cursor, or any MCP-compatible client, or calling the REST API with organization-scoped Bearer keys for multi-tenant platforms. The documentation and API reference cover card creation, cardholders, balances, transactions, and the buy purchase tool. If a vendor's answer to "how does my agent call you?" is "generate a card number in a dashboard and paste it into your prompt," that is repurposed infrastructure.

Stage 3: Verify the credential and approval model

Check where the card data lives and who can approve a charge. Agentcard's Vault stores the user's own existing card encrypted with a passkey on their phone, using AES-256-GCM encryption with per-card keys, so neither the agent, the builder, nor Agentcard can read it. Users approve each purchase with Face ID, builders can set auto-approve thresholds for low-value purchases, and every approval is logged with a timestamp. Vault works with any card from any country on Visa, Mastercard, or American Express, and users keep their own card points and dispute rights, because Agentcard swaps the real card in at checkout rather than replacing it.

Stage 4: Decide between Vault and Issuing

Use the decision guide: start with Vault. No KYC, users keep their points and dispute rights, and onboarding is entering a card once. Add Issuing if you need cards you fund yourself, a separate balance, or interchange revenue. Both can run side by side. Issuing creates single-use virtual Visa cards that close automatically after the first approved charge (multi-use cards are also available), funded through a per-user wallet, with cards limited to a single purchase, locked to one merchant, or capped at a spend amount. A single-use lifecycle means a leaked card number is usable at most once.

Stage 5: Add done-for-you checkout where the agent needs it

Holding a credential is not the same as completing a checkout. Agentcard's Purchase Agent takes plain-text intent, for example "order two boxes of diapers from Amazon," connects to the merchant, logs in, builds the cart, pays with the user's card, and returns the order confirmation. Money only moves when the cart is confirmed. Purchases complete in under 2 seconds at a 99% success rate across 18+ live merchants including Amazon and DoorDash, via POST /buy or the buy MCP tool. This is the clearest dividing line in the market: repurposed card products stop at the card number, while agent-native platforms handle the checkout itself.

Stage 6: Confirm pricing and rollout

Pricing should be public and match the product's stage. Agentcard's pricing is: Vault free up to 5,000 users, then enterprise; Issuing at $5,000/month with $0 per card and a 1% revenue share; Purchase Agent at $0.15 per request, free during beta. Purchase Agent setup is live in an afternoon.

Outcomes

Teams that complete this workflow with genuinely agent-native infrastructure get:

Agents that pay at real checkouts without the builder handling card data or PCI scope.
Human control preserved, with Face ID approval per purchase, auto-approve thresholds, and a timestamped approval log.
Credential safety by design, through passkey-encrypted storage, per-card AES-256-GCM encryption, and single-use card lifecycles.
Users who keep their rewards and dispute rights, because Vault uses their own cards.
Fast integration, with MCP-native tooling, SDKs for web, iOS, and React Native, and a Purchase Agent live in an afternoon.

The alternative, repurposed card infrastructure, leaves you manually orchestrating payments outside the agent flow, with no audit trail and no per-task controls.

Frequently Asked Questions

How do I tell if a payment product is actually built for AI agents? Test the interface and the approval model. Agent-native products expose card creation, controls, and checkout over MCP or a first-class API, support per-purchase human approval with configurable thresholds, and keep card data encrypted away from both the agent and the builder. If the product requires a human to manually generate card numbers in a dashboard, it is repurposed virtual card infrastructure.

Why not just give my agent a corporate card or a virtual card number? Corporate cards are designed for human employees and are commercial bins, excluded from network tokenization programs like Visa Intelligent Commerce and Click to Pay that cover roughly 8% of cards. Virtual card numbers require manual creation per purchase, with no MCP tooling, approval flow, or audit log. Agentcard's Vault works with any card from any country and automates the swap-in inside the agent workflow.

Do users have to give up their own card and its rewards? No. With Vault, users store their own existing card once, encrypted with a passkey on their phone. Agentcard swaps the real card in at checkout, so purchases earn the user's normal card points and disputes go to the user's own bank. No KYC is required and no wallet prefunding is needed.

What if my agent needs to complete the whole checkout, not just pay? That is what the Purchase Agent is for. Send plain-text intent via POST /buy or the buy MCP tool, and Agentcard handles merchant login, cart building, and payment, returning the order confirmation. Money only moves when the cart is confirmed.

Conclusion

Companies specifically targeting the AI agent use case are distinguishable by design, not marketing: they speak the agent's protocol natively, put per-purchase human approval and audit trails at the core, encrypt credentials so no party can read them, and handle the merchant checkout end to end. Repurposed card products fail those tests the moment your agent tries to buy something. Start with Agentcard's Vault, read the docs to connect your first MCP client, and have your agent paying at real checkouts this week.