An agent that orders groceries can read the list, match products, pick a delivery slot and build the cart. An agent that books travel can find the fare and fill in the passenger. Then both reach the same form: card number, expiry, CVV. Most of them stop there and hand the screen back to a person.

Agents that do finish the purchase use one of four approaches. They differ on what's behind the number that goes into the form.

What are the options when an agent reaches the payment step?

There are four in use today. The agent pays with a card saved in the customer's account at the store, or its tooling types in the customer's own card. It uses a standing virtual card with a limit, or it gets a card issued for that one purchase and capped at its amount. Each exposes a different amount of money to a mistake.

Take the saved card first. The agent logs in to the customer's account at the store and pays with whatever card is on file. It's the fastest to build because the store does the work. What's behind the number is the customer's whole card. The store's account controls are the only controls, and most stores built them for a person. The agent's maker usually adds a confirmation step, so a person approves every cart, which is safe and also means the agent can't finish anything alone.

The second option is the customer's own card, typed in. The customer gives the agent's tooling their card once. It's kept in a vault and injected into checkout pages so the agent's model never reads the digits. This fixes the "agent sees my card" problem. It leaves the bigger one: the full credit line sits behind every purchase, and the only bound on a wrong charge is whether the person read the approval prompt carefully.

The third is a standing agent card. A bank or card platform issues a virtual card for the agent with a daily, weekly or monthly limit. This is an improvement. The agent's spending is separated from everyone else's, and the limit caps the damage. The limit is per period, though, so a single wrong purchase can still take whatever's left in it. And these cards usually come with a business account at the issuer, which suits a company's own agents better than an agent acting for thousands of consumers.

The fourth is a card per purchase. The agent requests a purchase, the request is checked against the limits the customer set, and a card is assigned to that purchase alone, capped at its amount. What's behind the number is the price of the one thing being bought.

How do the four checkout options compare?

The saved card and the vaulted card both put the customer's full card behind every purchase. A standing agent card puts a period limit behind it, set by whoever created the card. A card per purchase puts one purchase's amount behind it, under limits the customer set, and it works at stores nobody integrated in advance.

Saved card at the storeCustomer's card, vaultedStanding agent cardCard per purchase
What sits behind the numberthe customer's full cardthe customer's full cardthe period limitthis purchase's amount
Who sets the boundthe store, if at allnobody, unless approval is the boundwhoever created the cardthe customer, in their own app
Works at a store you never integratedno, needs an account thereyesyesyes
Unusual purchasegoes through or needs a person every timewaits on approval every timegoes through inside the limitblocked or escalated to the customer
Record of the decisionthe store's order historythe tool's logthe card statementa tamper-evident audit log, per agent

An agent that's useful finds merchants nobody planned for. A method that needs a customer account at each store, or an integration with each one, covers the stores you thought of in advance. A card works wherever cards are accepted.

Which one should you build on?

Build on the option whose worst case you can afford. If your agent buys from one store and every cart gets a human yes, the saved card is enough, and adding more would be waste. Once you want the agent to finish some purchases alone, the question becomes how much money a wrong one can reach.

Take a meal-planning agent with a weekly grocery budget of 120 euros. On a saved card, a duplicated order is a second full charge on the customer's card, and they find out from their bank. On a standing card with a 500-euro monthly limit, it's bounded at what's left of the 500. On a card per purchase with the customer's own limits, the duplicate is checked against the weekly budget before any card exists. If it doesn't fit, it's blocked or it lands on the customer's phone with the store, the amount and the agent's name, and they decline it.

That's the design Shatale ships: a monthly budget and a cap per purchase, set by the customer in the Shatale app. A spending limit enforced at the authorization moment covers why the check runs before the card. If you build the agent, you integrate an API or the MCP server, and you don't run a card programme to do it.

What none of the four solve

No payment method checks whether the cart was right. A card capped at 84 euros will pay 84 euros for the wrong groceries. Substitutions, a wrong delivery address and a stale price are product problems, and they need product answers: a cart summary the customer reads, and an approval threshold set low enough to catch the odd ones.

Returns are the other gap. A refund goes back to the card that paid. Ask any provider what happens to a refund when the card that paid was issued for one purchase, and get the answer in writing before you ship.

FAQ

Is it safe to let an AI agent use a customer's saved card?

It's as safe as the confirmation step in front of it. A saved card has the customer's full limit behind it, so the protection is a person approving each purchase. That works while every purchase is confirmed. It stops working the day you let the agent complete purchases by itself.

Can an AI agent complete a checkout without a human?

Yes, if it has a payment credential and the checkout doesn't demand a step only a person can do, such as a code sent to a phone. Whether it should depends on the bound. With a card capped at the purchase amount and limits the customer set, routine purchases can go through alone and unusual ones wait for approval.

What's the difference between a standing agent card and a card per purchase?

A standing card has one number and a limit per day, week or month. A per-purchase card is assigned to one purchase and capped at its amount. The first bounds a period of spending. The second bounds each purchase and ties the card to one decision.


Corporate cards against agent payments covers the company-card version of the same choice. If you build agents, you can list yours in the Shatale marketplace, free while in early access.

Shatale is the control layer for AI-agent payments. Its authorization architecture is the subject of European patent application EP26194994.5 (filed; priority 28 July 2026).