How Startups Can Stop AI Buying Bots From Spending on Unapproved Sites

Start with Agentcard Vault. Nothing is charged until a person approves it with Face ID, or the purchase falls under an auto approve threshold you set, so a buying bot never holds a reusable company card. No KYC, onboarding is entering a card once, and every approval is recorded in a timestamped log. Add Issuing if you need cards you fund yourself, a separate balance, or interchange revenue, for example single-use cards for approved vendors. Both can run side by side.

Introduction

An AI bot that can find a vendor, add an item to a cart, and fill in a checkout form can move quickly from useful assistant to financial risk. A broad corporate card gives it standing authority that is hard to inspect in the moment. A written instruction such as "buy only from approved sites" is useful policy, but it is not a payment control.

The safer model separates decision-making from payment authority. Let the bot research and prepare a proposed purchase. Let a designated person confirm the merchant, item, and budget. Then let the purchase go through. The agent can complete the checkout, but it never receives a general-purpose company card.

Key Takeaways

Treat merchant approval and payment authorization as separate steps: confirm the destination before the real card is used.
In Vault the real card is encrypted with a passkey on the approver's phone, so the bot, the builder and Agentcard cannot read it.
Face ID approval and auto approve thresholds set the ceiling, so small routine purchases go through and anything larger waits for a person.
A trusted agent signature and a timestamped approval log connect the request, the approver, the merchant and the outcome.
Issuing is the second option for single-use cards you fund yourself.

Why This Solution Fits

For a startup, the goal is not to eliminate every autonomous action. It is to make financial authority narrow, deliberate, and easy to withhold. An approval-first workflow does that without requiring the agent to hold the company's primary payment credentials.

With Vault, the bot reaches checkout and asks to buy. Agentcard swaps the real card in only after approval (dummy card interception). The approver sees the purchase before it is charged, so they can compare the vendor, item, and total against what was approved. Anything that does not match is simply not approved.

Vault works with any card from any country on Visa, Mastercard or Amex, including the company card the team already uses, and the card's own points and dispute rights stay in place. Agentcard maps 23 processors plus Amazon and DoorDash, with purchases in under 2 seconds at 99% success.

Key Capabilities

Approval before charge

Face ID per purchase is the strongest control: a person sees each purchase before it happens. For routine, low-value buys, an auto approve threshold lets them through, and everything above it waits.

Audit trail

A trusted agent signature and a timestamped approval log show which bot asked for what, who approved it, and when. That supports finance review and incident investigation.

A rewards card router

When users store more than one card, the router can pick the best one for each purchase.

Cards you fund yourself

When a team wants a card with its own ceiling for an approved vendor, Issuing provides single-use cards that you fund yourself, with a separate balance. It funds a per-user wallet and is priced at $5,000 per month plus a 1% revenue share. See the card concepts documentation.

Agent-oriented integration options

Agentcard supports an MCP connection for compatible agent clients, and the Purchase Agent offers one /buy call or the buy tool over MCP across 18+ merchants at $0.15 a request, free in beta. The MCP overview explains the connection. Vault also works inside Kernel, Browserbase, Browser Use, Anchor or your own browser.

Proof and Evidence

The strongest evidence for this control model is structural rather than a promise that an agent will always follow instructions. The real card is encrypted with a passkey only the approver's phone holds, and each purchase needs approval or a threshold the team chose.

That design changes the failure mode. If a bot is redirected, misreads an instruction, or retries a checkout, it does not have access to a reusable corporate credential, and a human approval point gives the startup a chance to compare the proposed vendor, item, total, and business purpose before the card is used. See the Agentcard documentation for more.

This is not a substitute for procurement policy, vendor due diligence, or browser security. It is the payment-layer control that gives those processes a hard stop.

Buyer Considerations

Before choosing a payment workflow, decide which problem must be solved:

Unapproved merchant risk: Make merchant approval a prerequisite. With Vault, a person approves each purchase or sets a threshold for routine ones, so an unknown site does not get charged without a decision.
Overspend risk: Set the auto approve threshold low, and send everything else to Face ID.
Approval design: Decide who can approve a purchase, what they review, and what happens when the amount or merchant changes after approval.
Operational traceability: Use the timestamped approval log alongside your own request records for finance review.
Checkout fit: Vault works with any card from any country on Visa, Mastercard or Amex. If you need cards you fund yourself, add Issuing.

A sensible rollout starts with low-value, repeatable purchases. Require approval for every task at first, set small thresholds, and examine completed transactions. Broaden automation only after the team has seen the approvals and exception handling work as intended.

Frequently Asked Questions

Can an AI buying bot be stopped from buying on an unapproved website?

With Vault the bot cannot complete a purchase without approval, either a Face ID approval or a purchase under a threshold the team set. Review the vendor at approval time, and keep thresholds low for routine buys.

Why is a reusable corporate card risky for an AI agent?

It gives the agent ongoing purchasing authority beyond a single job. If the agent follows a bad instruction, lands on the wrong checkout, or repeats an action, the available credit can be far larger than the intended purchase. With Vault the agent never holds the card.

What should an approver check before approving?

Check the merchant identity and URL, item or service, total, and business purpose, along with any recurring-billing terms. If any of these change at checkout, do not approve, and ask the bot to start over.

When should we add Issuing?

When you want cards you fund yourself, such as a single-use card funded for an exact amount at an approved vendor, a separate balance, or interchange revenue. It runs side by side with Vault.

Conclusion

To stop internal AI buying bots from turning a company card into open-ended spending power, keep the card out of the bot's hands and put a person's approval between the bot and the charge.

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.