AI Agent Spending Controls: Limits, Approvals, and Scoped Intents
Key takeaways
- The strongest control on an AI agent's spending is a per-purchase approval of specific terms: the items, the merchant, and a price ceiling.
- An amount cap alone is weak, because it says how much the agent can spend but not on what, where, or when.
- Merchant scope, expiry, and revocation each close a different gap, and they work best together rather than as substitutes.
- Industry protocols such as Google's Agent Payments Protocol (AP2) and the Agentic Commerce Protocol (ACP) express these controls as signed or scoped permissions rather than open card access.
- An agent that never holds the raw card number cannot spend outside the controls by reusing the card somewhere else.
To control what an AI agent spends, make the agent ask before each purchase and have the person approve specific terms: what to buy, from which merchant, and the most it may cost. Add an expiry so the approval does not last forever. An amount cap is a useful outer limit, but on its own it lets the agent spend up to the cap on anything, anywhere.
What kinds of spending controls exist for AI agents?
AI agent spending controls fall into five types: per-purchase approval, amount ceilings, merchant scope, expiry, and revocation. Each control answers a different question about a purchase, and most real setups combine several of them.
- Per-purchase approval: A person sees the purchase and confirms it before the agent buys. This answers "did a human agree to this specific purchase?"
- Amount ceiling: A maximum the purchase may cost, including fees. This answers "how much?" but nothing else.
- Merchant scope: The purchase may happen only at a named merchant. This answers "where?"
- Expiry: The approval is valid only for a set time. This answers "until when?" and stops an old approval from being used months later.
- Revocation: The person can withdraw permission before it is used. This answers "what if I change my mind?"
How do the controls compare?
An approved line-item intent limits the most, and an amount cap alone limits the least.
| Control | What it limits | What it leaves open | Best used as |
|---|---|---|---|
| Amount cap only | The total per charge or per period | What is bought, where, and when | An outer boundary |
| Merchant scope | Which merchant can be charged | Which items and how much | A filter on top of other controls |
| Expiry | How long a permission is valid | Everything inside that window | A default on every approval |
| Revocation | Future use after the person withdraws | Anything already executed | A safety exit |
| Approved line-item intent | Items, quantities, merchant, ceiling, and fulfillment details | Very little, by design | The primary control for each purchase |
Why is a spending cap alone not enough?
A spending cap alone is not enough because it describes a ceiling, not a decision. If an agent has a $200 limit, every purchase under $200 looks the same to the cap: the right pair of shoes, the wrong pair, or something the person never asked for.
- It cannot see the basket: A substitution or an added item passes the cap as long as the total stays under it.
- It cannot see the merchant: Unless the cap is tied to a merchant, the agent can spend at any store that accepts the card.
- It invites spending to the limit: A cap set high enough for the largest expected purchase is also a permission to spend that amount on a mistake.
- It is weak evidence: If a charge is questioned later, "it was under the cap" does not show that the person approved that purchase.
A cap tells you how much an agent can lose. An approval tells you what the person actually agreed to.
How does an approved intent bound what the agent buys?
An approved intent bounds what the agent buys by naming the merchant, the items, the quantities, the price ceiling, and fulfillment details such as an address, before any money moves. The person approves that plan, and the system that executes the purchase checks the checkout against it.
- 1
The agent drafts the purchase: The request becomes concrete products at a named merchant, with a total and a ceiling.
- 2
The person approves the terms: The person confirms the approval with a strong method such as a passkey or a device biometric. An agent can draft an intent, but only a person should be able to approve it.
- 3
The system executes inside the terms: Checkout runs against the approved merchant and items. A well-designed system fails the order rather than exceed the approved ceiling.
- 4
Material changes go back to the person: If the price, the item, or availability changes in a way the approval does not cover, the purchase does not silently proceed. The person is asked again.
The approval should also be scoped in time. An intent that states how long the authorization is valid prevents a forgotten approval from staying usable indefinitely.
Industry protocols follow the same idea. Google's Agent Payments Protocol (AP2) records the person's instructions as cryptographically signed mandates, so a later dispute can refer to what the person actually authorized. The Agentic Commerce Protocol (ACP) from OpenAI and Stripe describes a delegated payment token that is limited to one merchant, a maximum amount, and an expiry time.
Should the agent ever see the card number?
No, an AI agent should not see the card number. The safer design keeps the card details in a tokenized vault and gives the agent only a label and the last four digits, so the agent acts on an approval and holds no credential it could reuse somewhere else.
The card networks support this design with agent-specific tokens. Visa Intelligent Commerce and Mastercard Agent Pay each issue tokenized credentials for agent transactions, and both use strong cardholder authentication, such as passkeys or device biometrics, to tie a purchase to the cardholder. When the charge runs on the person's own card, a purchase the person did not approve can be disputed through their bank under the usual card network rules.
This page explains how spending controls work. It is not legal advice. Dispute and liability outcomes depend on card network rules, issuer policy, and the agreements between the parties.
How should a builder combine these controls?
A builder should make the per-purchase approval the primary control and use the other four controls to narrow it.
- 1
Require an approval for each purchase: Make the line-item approval the main control, not a setting the person turns on later.
- 2
Put a ceiling on every approval: Include fees, so the total the person sees is the most that can be charged.
- 3
Name the merchant: An approval for a specific store is easier to check and harder to misuse than one for "any store".
- 4
Set a short expiry: Match the expiry to the task. A dinner order needs minutes; a purchase that waits on a restock may need days.
- 5
Show the person what happened: Report the order number and the amount charged, so the person can compare the result with what they approved.
What mistakes weaken agent spending controls?
The most common mistakes are standing approvals, ceilings that leave out fees, approvals with no expiry, card numbers given to the agent, and approvals that are not recorded.
- One standing approval for all purchases: A single "yes" that covers every future purchase is a cap with extra steps. Each purchase needs its own approval.
- A ceiling without fees: Taxes, delivery, and service fees can push a charge past the amount the person had in mind.
- No expiry: An approval that never ends turns into an open permission that nobody remembers to withdraw.
- Giving the agent the card number: Once an agent holds the card details, every other control depends on the agent choosing to respect it.
- No record of the approval: Without a logged approval, a later question about a charge becomes a matter of memory instead of evidence.
Frequently asked questions
How do I stop an AI agent from overspending?
Require the person to approve each purchase with a price ceiling that includes fees, and use a payment system that fails the order rather than exceed that ceiling. The agent can draft a purchase, but only a person should be able to approve it.
Is a spending limit on a card enough to control an AI agent?
No, a card spending limit is not enough on its own. A limit controls how much an AI agent can spend, but not what it buys or where. An approved line-item intent names the merchant, items, and ceiling, so a wrong item or an unexpected merchant falls outside the approval.
Can I limit an AI agent to one merchant?
Yes, an AI agent can be limited to one merchant if the approval or the payment token names that merchant. The Agentic Commerce Protocol, for example, describes delegated payment tokens that are scoped to a single merchant.
What happens if the price changes after I approve an AI agent purchase?
If the new price is still within the approved ceiling, the purchase stays within the approved terms. If a change is material and the approval does not cover it, a well-designed system does not let the purchase silently proceed, and it asks the person again.
Should an AI agent approval expire?
Yes, an AI agent approval should expire. An approval that never expires can be used long after the person has forgotten it. An expiry on each approval limits the window in which it can be used.
Should an AI agent get my card number?
No, an AI agent should not get your card number. A safer design stores the card in a tokenized vault, shows the agent only a label and the last four digits, and runs the charge on your own card through network tokens such as those from Visa Intelligent Commerce or Mastercard Agent Pay.
Related articles