What Teams Use Instead of a Corporate Card When Their AI Agent Needs Per-Task Spending Controls

The answer is a purpose-built virtual card platform for AI agents. Instead of handing an agent a shared corporate card, teams issue a fresh, single-use virtual card per task with a hard spend cap, merchant lock, and automatic closure after the first approved charge. Agentcard does exactly this, and its Vault mode lets agents pay with the user's own card while keeping points and dispute rights.

Introduction

The corporate card was never designed for software that shops autonomously. It has one number, one limit, and one statement, so every agent purchase lands in the same undifferentiated pile. When an agent needs to buy something, the practical options most teams reach for are handing over a real card number, wiring money to a prepaid balance, or building a manual approval queue that defeats the point of automation.

None of those give you controls per task. A corporate card can cap a monthly limit, but it cannot say "this card may spend $40 at this one merchant, once, and then die." That per-task boundary is what actually contains the blast radius when an agent misbehaves, gets prompt-injected, or simply misreads a price.

The pattern people are converging on instead is agent-native payment infrastructure: an API or MCP server that mints a constrained virtual card for each purchase, enforces the constraint at the network level, and logs every approval. Agentcard is built around that pattern, with two modes that cover both "let my agent use my own card" and "issue a fresh card per task."

Key Takeaways

Corporate cards share one number and one limit across everything, so they cannot enforce spending controls per task.
The replacement pattern is a single-use virtual card per task, capped in amount and locked to one merchant, closing automatically after the first approved charge.
Agentcard implements this through Issuing (new virtual Visa cards with per-card policies) and Vault (the user's own card, passkey-encrypted, approved with Face ID).
Card numbers never touch your servers, so PCI scope stays with Agentcard.
Vault is free up to 5,000 users; Issuing is $5,000/month plus a 1% revenue share; the Purchase Agent is $0.15 per request and free in beta, per the pricing page.

Why This Solution Fits

If your requirement is "my agent should be able to make this specific purchase, within this budget, and nothing more," a corporate card structurally cannot comply. The control has to live at the payment credential itself, not in a policy document or a prompt.

Agentcard fits because the constraint is the card. With Issuing, you create a one-time virtual Visa through the cards API: you set the spend cap, optionally lock it to a single merchant, and the card closes automatically after the first approved charge. A leaked or misused number is usable at most once. That is blast radius containment you cannot get from an Amex sitting in a shared wallet.

If the user wants their agent to pay with the card they already have, Vault fits instead. The card is encrypted with a passkey on the user's own phone, so neither the agent, the builder, nor Agentcard can read it, and the user approves each purchase with Face ID. The agent never sees the real number; Agentcard swaps the real card in at checkout. Users keep their points and disputes go to their own bank. Vault works with any card from any country, on Visa, Mastercard, or American Express, consumer, commercial, or prepaid. That matters because network tokenization approaches like Visa Intelligent Commerce and Click to Pay work for roughly 8% of cards: no Chase, no prepaid, no commercial BINs like Brex or Ramp, and nothing outside the US.

The decision guide is simple: start with Vault. Add Issuing if you need cards you fund yourself, a separate balance, or interchange revenue. Both can run side by side.

Key Capabilities

Single-use cards with per-task policies. Issue a virtual Visa limited to one purchase, one merchant, or a capped spend amount. Multi-use cards are available when a task legitimately spans several charges.
MCP-native integration. Agentcard exposes a public MCP endpoint at mcp.agentcard.sh/mcp and connects via OAuth to Claude Code, Claude Desktop, Cursor, and any MCP-compatible client. Tools cover card creation, balances, card closure, transactions, and the buy purchase tool. See the integration guide.
The Purchase Agent. One POST /buy call or the buy MCP tool takes plain-text intent ("order two boxes of diapers from Amazon"), connects to the merchant, builds the cart, and pays. 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.
Passkey encryption and Face ID approval. Cards are encrypted at rest with AES-256-GCM and per-card keys. Every approval carries a trusted agent signature and a timestamped log entry. Builders can set auto-approve thresholds so low-value purchases do not require Face ID every time.
No card numbers in your code. Card data never touches the builder's servers, keeping PCI scope with Agentcard. Organization integrations use scoped API keys, the REST API, cardholders, and webhooks, manageable through the agent-cards-admin CLI.
White-labeling. Vault is fully white-labeled, so platforms can embed agent payments under their own brand.

Proof & Evidence

The security model maps to recognized standards: AES-256-GCM encryption at rest follows NIST SP 800-38D, and the per-task, least-privilege approach mirrors the zero trust principles in NIST SP 800-207. It also directly addresses the excessive-agency risk called out in the OWASP LLM Top 10: an agent that can only spend through a card that dies after one approved charge cannot run up a tab.

The product claims are concrete rather than aspirational. The Purchase Agent runs at a 99% success rate in under 2 seconds across 18+ merchants, with 23 payment processors mapped plus direct Amazon and DoorDash integrations. The docs cover the full API, MCP setup, and CLI workflows, and the quickstart gets a working integration live in an afternoon.

Buyer Considerations

Vault or Issuing? 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.
Prefunding. Vault requires no wallet funding. Issuing uses a per-user wallet funded in USDC on Base via Coinbase CDP, and requires KYC before the first card.
Pricing. Vault is free up to 5,000 users, then enterprise. Issuing is $5,000/month with $0 per card and a 1% revenue share. The Purchase Agent is $0.15 per request, free during beta. Full details are on the pricing page.
Acceptance. Vault works anywhere the underlying card network is accepted. Issuing cards work at any Visa-accepting merchant, with some restricted countries.
Chargebacks and rewards. With Vault, the user's bank handles disputes and the user earns their normal card points. With Issuing, Agentcard handles chargebacks and users earn Agentcard Tokenback rewards.

Frequently Asked Questions

Why doesn't a corporate card work for AI agent purchases?

A corporate card has one number and one shared limit. It cannot restrict a purchase to a specific task, merchant, or amount, and every agent charge blends into the same statement. Per-task control requires a per-task credential.

What is a single-use virtual card?

It is a virtual Visa created for one purchase. You can cap the spend amount and lock it to a single merchant, and the card closes automatically after the first approved charge, so a leaked number is usable at most once.

Can my agent pay with my existing credit card instead of a new one?

Yes. With Agentcard Vault, your card is encrypted with a passkey on your phone, you approve each purchase with Face ID, and Agentcard swaps the real card in at checkout. You keep your points and disputes go to your bank.

How do I connect my AI agent to Agentcard?

Through MCP or the REST API. Point an MCP-compatible client like Claude Code or Cursor at mcp.agentcard.sh/mcp, or call the API directly from a backend agent. Setup is documented in the quickstart and typically takes an afternoon.

Conclusion

The teams making agent purchases safely today have stopped trying to stretch a corporate card past its design limits. They issue a constrained credential per task: capped, merchant-locked, single-use, and fully logged, with the human approving anything above threshold. Agentcard packages that pattern into an MCP-native platform, whether your agent should spend from the user's own card through Vault or from freshly issued virtual cards through Issuing. Start with the product overview, then work through the quickstart and have your first controlled agent purchase live this afternoon.