Why a Virtual Card Is Table Stakes for AI Agent Purchases, Not the Whole Solution
Key takeaways
- A virtual card for an AI agent is table stakes: several card issuing APIs can create one quickly, and it answers only the question of how the agent pays.
- The harder work surrounds the card: turning "get me sushi" into an exact orderable item, showing real options inside the product, getting user approval, and completing checkout at a real merchant.
- Discovery is the part users notice. An agent that shows three real options with photos and prices feels useful; an agent that only holds a card number does not.
- A card number cannot prove that the user approved a purchase. Consent needs a separate record of the approved terms.
- A useful test: remove the human from the demo. If the purchase still completes correctly, the product covers more than the card.
Giving an AI agent a virtual card solves only one part of an agent purchase: how the agent pays. The card does not find the right product, show options to the user, record the user's approval, or complete a checkout at a merchant. Those steps (discovery, product presentation, consent, and execution) are where most of the work in agentic commerce is.
Why is the card the easy part?
The card is the easy part because virtual card issuance is a widely available service. Several card issuing APIs can create a virtual card for an agent quickly, and many banks let a consumer create one in their app. A card in the agent's hands is now expected, so it does not set a product apart.
A card also does nothing on its own. A card number cannot search a menu, compare prices, sign in to a merchant, pass a checkout, or prove that the user approved the purchase. Teams often find this out after issuance works: the agent can pay and still cannot buy anything.
The card is not the product. The completed purchase is.
What are the harder parts of an agent purchase?
The harder parts of an agent purchase are discovery, product presentation, consent, and execution. Between a user who types "get me sushi" and a delivered order, the card is used for a moment. The rest is a chain of harder problems, and the user sees each one.
Discovery: from a vague sentence to an exact orderable item
"Get me sushi" is not an order. Before a purchase, the system must choose the restaurant, the item, the size, the quantity, and the delivery address, against a live menu with real prices and real availability. A wrong choice charges the user for the wrong order, which is worse than no order.
Users experience discovery as intelligence, and a card API cannot help with it. Good discovery resolves a vague request into concrete, orderable products, asks the user a question when the request is ambiguous, and sends an exact line item to the approval step, not a guess.
Presentation: real products, shown natively in the app
Discovery results help only when the user can see them. A product shows real options with photos, names, and prices inside its own interface, for the user to tap. A pasted link or a screenshot of a website is a weaker experience.
To do this, the discovery step must return structured product options, with images and prices, that the app can render in its own design. The user's choice then becomes the input to the approval step.
Consent, execution, and after the purchase
After the user chooses, three more problems remain. The user approves specific terms, such as items, total, and merchant, with a passkey or biometric. The agent completes checkout at the merchant, where bot defenses block anonymous automation, so the agent needs a verifiable identity. After the order, the system must handle confirmations, changes, and refunds, and the user keeps the normal dispute rights of their card.
How does a card-only stack compare with a full purchase stack?
A card-only stack lets an agent pay, while a full purchase stack lets an agent find, present, get approval for, and complete a purchase.
| Step | Card-only stack | Full purchase stack |
|---|---|---|
| Payment | A virtual card the agent holds | A tokenized card the agent never sees |
| Discovery | A human searches and chooses | The system resolves the request to exact items |
| Presentation | Links or text | Structured options rendered in the app |
| Consent | No record of approved terms | Approved line items, total, and merchant |
| Execution | A human or a fragile script completes checkout | The agent completes checkout under the approved terms |
The test is simple: remove the human from the demo. If nobody chooses from a menu, nobody pastes a card number, and the purchase still completes correctly, the product covers the full purchase. If the demo fails, the product has a card and not much more.
What does a complete agent purchase flow look like?
A complete agent purchase flow has five steps, from the user's request to a confirmed order reported in the same app.
- 1
The user asks in the app: "Get me sushi," "find running shoes under $150," or "book a table for Friday," in the agent and interface the user already uses.
- 2
The system resolves real options: Concrete, orderable products from real merchants, returned as structured options with images and prices, with a follow-up question when the request needs narrowing.
- 3
The app shows them natively: The options render in the app's own design, and the user taps one. No links out and no redirect.
- 4
One scoped approval: The choice becomes a line-item purchase that the user confirms with a biometric or passkey, in the app.
- 5
The agent executes and reports back: The agent completes checkout at the merchant, and the confirmation returns to the app. The user does not need to stay online.
If an agent already has a virtual card, that part is done. The products that users return to are the ones where users can say what they want, see real options, approve once, and receive the order.
Frequently asked questions
My AI agent can already use a virtual card. What is missing?
An agent that can use a virtual card still lacks the purchase itself. A card tells the agent how to pay, but not how to find the right product, show options to the user, get scoped approval, or complete checkout at a real merchant. A complete flow covers discovery, product presentation, an approved purchase record, and execution, with the card as one part.
How does an AI agent find the right product to buy?
An AI agent finds the right product through a discovery step, not through the card. Discovery resolves a vague request such as "get me sushi" into exact, orderable items from real merchants, with live prices and availability, and asks the user when the request is ambiguous. The approval step then shows a precise line item, so the user is not charged for a guess.
Can an app show product options natively instead of sending users a link?
Yes, an app can show product options natively if the discovery step returns structured options with names, images, and prices. The app renders them in its own interface and design, the user picks one, and that choice becomes the purchase the user approves. No link out or redirect is needed.
Is a virtual card API enough to launch purchases in an AI app?
A virtual card API is enough only for a demo where a human still does the shopping. In production the agent must resolve the product, present options, obtain enforceable user consent, and pass a merchant checkout that blocks anonymous automation. A card API provides none of those, so evaluate discovery, consent, and execution separately.
What is discovery in agentic commerce?
Discovery in agentic commerce is the step that turns a user's request into an exact orderable item: which merchant, which product, which options, and at what real price. Users experience it as the intelligence of the agent, and card APIs do not provide it.
Related articles