Blog

What it means when an AI assistant shops and pays for you

How agentic commerce works in plain terms: an assistant acting on your instructions, the tokens and limits it pays with, and what a cardholder still controls.

For years an AI assistant could help you think about a purchase: compare two blenders, summarize reviews, draft a return request. The next step is already being built and tested. You tell an assistant what you want, with a budget and conditions, and it finds the item, fills the cart and pays. That arrangement has a name, agentic commerce, and it raises questions that sound new and are mostly old ones in new clothes: who authorized this charge, how much can it be, what happens if the item is wrong and which of my rights still apply. This explainer describes how such a payment would work, what controls a cardholder can set and which questions are still open. It is written as of November 2026, and the field is early and changing. Product names, availability, networks' rules and the legal treatment of agent-made purchases are moving, so treat this as a map of the ideas and check each product's terms.

Three levels of "the assistant shops for me"

Not everything that is called agentic commerce involves an assistant spending your money. There are three levels, and which one you are using decides what can go wrong.

  1. The assistant recommends. It researches, compares and gives you a link. You check out yourself. Nothing about your payment changes, and the risks are the ordinary ones of reading a recommendation: a stale price, a biased review, a lookalike store.
  2. The assistant prepares and you approve. It fills the cart, enters the shipping details and stops at the final button, which you press. You see the item, the price and the total before anything is paid. The payment is yours in the plain sense.
  3. The assistant pays within limits you set. You give it a standing instruction, such as "reorder dog food when the bag is low, up to this amount," and it completes the purchase without asking each time. This is the level that raises the new questions, because a payment happens that you did not individually confirm.

Most of this article concerns the third level, with the second as the safer neighbor. The practical advice is to start at the second and move to the third only where a mistake would be small.

How an agent could pay without holding your card number

A person typing a card number into a website is already a design with problems: the number is copied into a form, stored by the merchant and reused. Handing the same number to a program that acts on its own would multiply the problems. The designs being described for agent payments avoid it by giving the agent something narrower than your card. The ideas should be familiar from other pages on this site.

A stand-in credential. Instead of your real number, the agent receives a token or a virtual number issued for its use. tokenization is the general idea: the merchant and the agent handle a substitute that works only under set conditions, and the real number stays with the issuer and the network. how a virtual card number works, and when it beats your real number explains the same pattern for a virtual card number that you create yourself.

Conditions attached to the credential. A credential for an agent can carry rules: an upper amount for one purchase, a total for a period, a list of merchants or categories, a time window and whether it works once or repeatedly. The issuer or the wallet checks the rules when a charge arrives, and a charge outside them is declined. This is a spending limit enforced by the payment system and not by the agent's good behavior, which is the important difference.

A record of what you agreed to. Some of the described designs have the user approve a specific instruction, sometimes called a mandate, with the conditions written down, and attach that record to the transaction so the merchant and the issuer can see that a person authorized an agent within limits. Whether and how a given product does this varies, and the record's legal weight is not settled.

A signal to the merchant. A merchant may be told that an agent, and not a person at a browser, is making the purchase. That lets it apply its own checks. It also means that a payment from an agent can look unusual to a fraud system built for people, which is a reason a legitimate agent purchase could be declined in the early period.

None of this is a description of a specific product. It is the common shape of the proposals, so that you can ask whether a product you are offered works this way.

A walk-through: one purchase, start to finish

A labeled hypothetical, with invented figures. Rin tells an assistant: "Reorder the dog food when I am down to a week's supply, as long as the price is under $60 with shipping." Here is how the purchase could run.

  1. Setup. Rin approves the instruction in the assistant's app and picks which card or credential it may use. The app creates a merchant-locked credential limited to $65 per purchase and $65 a month, valid for six months. The $65 leaves $5 of room over the $60 limit, which is 8.3 percent (5 divided by 60), enough for a small price change and not for a different product.
  2. Trigger. The assistant sees from a note Rin gave it that supply is low. It searches, finds the same food at two stores and compares prices including shipping: $57.40 and $61.90.
  3. Decision. The second price is above $60, so it chooses the first. It builds the cart and checks that the item is the same product and size.
  4. Payment. It pays with the credential. The merchant authorizes $57.40. The credential's rules allow it.
  5. Record. The assistant posts a note to Rin with the merchant, item, total and order number, and the issuer's app shows the charge.
  6. Follow-up. Rin reads the note, glances at the charge, and the order arrives.

The arithmetic of the limit shows what the system protects. If the assistant made a mistake and chose a larger product at $64, the charge would still pass. If it chose a product at $90, the charge would be declined. The worst case on this credential is $65 for any one purchase, and also $65 in total each month, however many times the assistant tries. Without the cap, the worst case would be the card's credit limit.

What can go wrong

The risks fall into a few groups, and most can be reduced by the limits above.

  • The wrong item. The assistant misreads a product, picks the wrong size or buys a lookalike. This is the most likely failure, and the cost is a return. If an AI shopping agent buys the wrong item, who ends up paying? looks at who bears it.
  • Instructions hidden in a page. Because an assistant reads text on web pages, a page can contain text written to manipulate it, such as a line that says to buy a different item. This kind of attack is a recognized concern for assistants that browse and act. The defense is narrow permissions, a merchant list and confirmation for anything unusual.
  • A fake store. An assistant can be pointed to a store that looks real. A merchant allow-list and a single-use credential reduce the exposure.
  • A price that changes between decision and payment. A cap handles this, and an authorization hold placed at order time for a different amount than the final charge can briefly reduce your available credit.
  • A subscription or add-on. An assistant may accept a checkbox it should have declined. A credential that works once, with a cap close to the price, makes an unwanted renewal fail.
  • Too much information shared. An assistant may enter your address and phone number into forms. Know what data it is allowed to use.
  • A stolen login. If someone takes over the account you use to control the assistant, they can give it instructions. This is card-not-present fraud of a new kind, and the lock on that account is the same as for any other: a strong password and two-step verification. card account security in one sitting: alerts, locks and strong logins covers it.

Controls a cardholder should look for

If a product offers to pay for you, ask whether it has these controls. If it does not, that is a reason to stay at the second level.

  1. A per-purchase cap and a period cap, set by you.
  2. A merchant or category allow-list.
  3. Confirmation above a threshold, so that an unusual or large purchase asks for your approval.
  4. A single-use or merchant-locked credential, not your real card.
  5. An expiry and a revoke button that works immediately.
  6. A log of what the assistant did, with the merchant, amount and order details, which you can read.
  7. Alerts on every charge through your issuer, separate from the assistant's own notes.
  8. A clear statement of what happens on a return or a dispute.

A labeled limits exercise helps. Say your usual weekly grocery order is $95 and an assistant will place it. A cap per order of 120 percent of the usual is $114, which you might round to $115. At four orders a month, the cap allows 4 times $115, or $460, while the expected spending is 4 times $95, or $380. The headroom is $80, or about 21 percent of expected spending (80 divided by 380). If that feels like too much, lower the monthly cap to $420 and accept that a holiday week may need a manual order.

What happens when something goes wrong

This is the least settled part. The rules that protect cardholders were written before such agents existed, and the way they apply is still being worked out. Four points can be said with confidence.

Your card agreement still governs. The issuer's terms, including its dispute process, apply to charges on your account whoever or whatever initiated them. Federal law caps a cardholder's liability for unauthorized use of a credit card at $50, and a written billing-error notice must reach the issuer within 60 days of the statement showing the charge. Those rules are the baseline.

"Unauthorized" is the open question. An unauthorized charge is one you did not authorize. If you instructed an agent to make purchases within limits, a purchase inside those limits may be treated as authorized by you, even if you did not like the result, which would put it closer to a dispute with the merchant than to unauthorized use. A purchase outside the limits you set, or made after you revoked the agent, is a stronger case for being unauthorized. How issuers and courts will treat these cases is not settled, and this site's research did not find a card rule written specifically for AI agents. Do not assume a protection exists because the purchase was automated. scam payment or unauthorized charge: why your protections can differ shows how the line between unauthorized and authorized-but-regretted is drawn for other situations, and the $50 limit on unauthorized charges: what it covers and what it does not explains the $50 rule.

The merchant's own return policy applies. A wrong item is first a matter between you and the seller, and the agent is no defense for either of you.

Speed still matters. Read the alert, check the log and call the issuer using the number on the back of the card if anything looks wrong. If the issuer does not resolve a real problem, the CFPB takes complaints about card issuers.

Returns, refunds and the paper trail

Buying is the easy half. A purchase that goes wrong ends in a return, a refund or a dispute, and each of those needs facts that an assistant should leave behind.

Returns. A return follows the merchant's policy, and the assistant may or may not be able to start one. Before you rely on an agent for anything you might send back, ask whether it can find the return window, print the label and track the refund, or whether those steps fall to you. A refund goes back to the credential used. If that credential was single-use and has expired, the same issue that affects ordinary virtual numbers applies, and virtual card mistakes with free trials, returns and hotel stays describes it.

Records. For each purchase, you want the order number, the item and price as the agent saw them, the time, the merchant's name and the credential used. A good product shows these in a log without being asked. If yours does not, save the confirmation emails in one folder. A dispute is much easier to win or settle when the facts are in one place.

Statements. Read the card statement as you would any other. A charge made by an agent appears like any other line, sometimes under a payment processor's name. Match each line to the assistant's log in the first months, because mismatches show where the setup needs tightening. how to use a virtual card number for online shopping, step by step shows the record-keeping habits that apply to any online purchase made with a stand-in number.

Cancelling. If you stop using an assistant, revoke its payment credential, check for standing instructions, such as reorders or price alerts that trigger purchases, and look for any subscription it started on your behalf. Revoking in the assistant's settings is not always the same as closing the credential at the issuer, so check both.

Questions to ask before turning an agent loose

  1. What can the agent spend, per purchase and per period, and who enforces the limit, the agent or the payment system?
  2. Which credential does it use: my real card, a token or a virtual number?
  3. Can I see a log and revoke it immediately?
  4. What does the product's agreement say about wrong orders, refunds and disputes?
  5. Which data does the agent store and share with merchants?
  6. What happens if the company changes the product or shuts it down?
  7. Would I be comfortable if the agent made its worst allowed purchase today?

The last question is the test. If the worst allowed purchase is a nuisance, the setup is reasonable. If it is a hardship, lower the limits or keep a person in the loop.

Key takeaways

  • Agentic commerce ranges from an assistant that recommends, to one that prepares a cart for your approval, to one that pays on its own within limits. The risk grows at each step, so start with approval.
  • The designs being described give an agent a stand-in credential with rules attached, such as a cap per purchase, a period cap, a merchant list and an expiry, enforced by the payment system and not by the agent.
  • In the example, a $65 cap on a $60 purchase leaves about 8.3 percent of room, and a $460 monthly cap on $380 of expected orders leaves about 21 percent. The cap sets the worst case.
  • Your card agreement, the $50 cap on unauthorized use and the 60-day billing-error notice still apply, but whether an agent's purchase inside your limits counts as unauthorized is unsettled, so do not assume a protection.
  • Look for per-purchase caps, a merchant list, confirmation above a threshold, a log, a revoke button and alerts on your own card account, and treat any product without them as a reason to stay in the approval step.

Terms used in this article

Latest on this topic