Sardine named a Leader in The Forrester Wave™: Financial Crime Management Solutions, Q3 2026

Learn More
Fraudology

When to Block AI Shopping Agents, and Who Pays When You Don’t

52 min

Welcome back to Fraudology.

After some great interviews, I wanted to share what keeps coming up in my conversations with merchants and fraud leaders since MFA. Agentic commerce is at the top of that list, and it is the main topic today. For the first time, merchants are seeing AI agent orders in their data, and some are already dealing with AI shopping agent chargebacks.

Banks, airlines, and rental car companies have started to take public stances on AI shopping agents. Some have said no, one has said yes, and many merchants are somewhere in the middle without enough information to decide. I talk through what each group is weighing, why the chargeback rules matter more than the headlines, and why you cannot block what you cannot identify.

I also share the Black Friday and Cyber Monday data I have been reading, along with what I am hearing about 3D Secure, account takeover, and the job market. If you are trying to decide when to block AI shopping agents, start here.

What you’ll hear in this episode:

  • Why banks are cutting off AI agents in online banking, from wrong payee mistakes to stolen credentials
  • Why United Airlines said no to AI agents booking travel while Hertz partnered with one
  • Why Visa says the cardholder is responsible for agentic purchases and merchants still get the chargeback
  • A real high dollar chargeback, and what compelling evidence for AI agent purchases would need to look like
  • Why fraudsters are early adopters of AI agents, and why that matters if you decide to allow them
  • Why many merchants cannot identify agentic orders, and what to ask your fraud provider about the identifier
  • How device ID and behavioral biometrics can help you spot an AI agent, and how Know Your Human adds a verification step
  • What the Accertify holiday shopping data says about earlier peaks, overnight fraud, and larger fraud orders
  • A short update on 3D Secure strategy, the account takeover spike, and the fraud job market

You should listen to this episode if you:

  • Own a decision about whether your site welcomes, limits, or blocks AI shopping agents
  • Work in fraud at a merchant and have been asked about agentic commerce by a growth or sales team
  • Handle chargebacks and want to understand why agent purchases default to the merchant
  • Support merchants as a fraud provider, processor, or consultant and field questions about agent traffic
  • Want data to share with leadership about the holiday shopping season
Episode notes

Banks said no first

Several big banks are blocking AI agents and have taken a clear stance. JPMorgan Chase has said it will not allow agents to perform online banking transactions for its customers, and I believe Bank of America has said the same. I think that is a big deal. Agents can make mistakes when instructions are vague. They can send money to the wrong payee, change the payee information, or withdraw more than the customer intended. Fraud is part of the picture as well, including credential theft that uses an agent as cover. A criminal who steals credentials and claims to be an agent can access an account and move money. Banks can identify agents using device ID and behavioral biometrics together, and I won’t share the specific signals on a public platform. The bigger point is that banks do not plan to reimburse customers for agent mistakes. Their message is simple. If one gets through, it is on you.

United said no and Hertz said yes

Merchants are taking public stances as well. United Airlines said recently that it does not allow agents to book travel on its platform. I don’t blame them. Booking travel is one of the main use cases for agentic commerce, which makes AI agent booking travel fraud and honest mistakes equally easy to imagine. When something goes wrong, the merchant is on the hook by default, because the chargeback rules have not been updated for agent purchases. Air travel also has very little profit margin. Every fraudulent order can take 20 to 30 good orders to make up. Call center volume is another factor. I plan to ask their fraud director some questions this week. Meanwhile, Hertz announced a partnership with a company that hosts AI agents and plans to encourage them. I do not know whether their fraud leader was consulted. If so, I think that leader lost the debate, and I worry about how they will defend their chargebacks.

Visa says cardholder, the chargeback says merchant

Visa has said that cardholders will be responsible for agentic purchases. In practice, the chargeback still lands on the merchant. Visa and the issuing bank have no way to know that an agent made the purchase, while the merchant might. Even when a merchant proves an agent was involved, the Visa and Mastercard rules do not say that proof moves liability to the cardholder. I know several merchants who have received chargebacks on AI agent purchases already, and many are high dollar. Merchant liability for AI agent purchases is the default for now. One merchant received a chargeback for about $4,000, which was roughly 10 times their average order value. They could not tell an agent was involved, though the shopper’s data belonged to a real person. The chargeback documentation showed the cardholder saying an agent made the purchase. The merchant argued that this was not their problem. The issuer replied that it was a card not present transaction and the cardholder did not make it himself, which makes it fraud. Without a change in the rules for compelling evidence, merchants carry the loss.

Most merchants are somewhere in the middle

Most merchants are somewhere between United and Hertz. Some can identify agent transactions, and when they do, they tell me most are fraudulent agentic orders, which makes sense because fraudsters are early adopters. Others do not know how much agent traffic they have. Every growth team is excited about a new channel, and I understand why. I think a measured approach is wise while the chargeback rules lag. Whether a company ends up more like United or more like Hertz depends on who makes the decision and whether liability changes. If the fraud manager is consulted, the answer may change. My guess is that agentic commerce may be great for revenue, but after chargebacks, the profit will be small. We don’t have enough data to know yet. If a merchant can prove an AI agent bought the item at the shopper’s direction, more merchants will feel comfortable allowing it.

You cannot block what you cannot identify

I know a retailer who ran test purchases through every AI platform they could find. Then they worked with their fraud provider to compare every signal against a regular purchase. Of all the data the provider ingested, none identified the agent. Many retailers use Stripe and Shopify, and I don’t think Stripe has the infrastructure to flag it. Shopify might, but I don’t know whether they would share it. My advice is to ask your fraud provider exactly where the identifier for AI commerce lives. If it isn’t there, you may need another tool in your stack that strengthens device ID and behavioral biometrics. Some vendors can detect agents very quickly, and I’m happy to share a list.

Know Your Human and a way to verify the person

Sardine, the sponsor of this podcast, announced a product last week called Know Your Human, or KYH. We have KYC for customers and KYB for businesses, and now this idea applies to the person behind an agent. The merchant sends a one time passcode through a link to the cardholder. When they click, the merchant can capture the device, IP, how they hold the phone, and other telemetry and behavioral biometrics in a quick ping. The merchant gets back a yes, no, or maybe. It does not solve intent or spending limits, though other protocols do. It is also protocol agnostic, which matters if your customer chooses a competitor’s platform. If a merchant has proof that the cardholder approved the order, I think issuers could eventually be pushed to accept it when representing chargebacks. I’m planning an anonymous webinar for merchants in January or February. My fear is that Visa will not change its rules until next year, since updates come in October and April.

Holiday data and a few other things I am hearing

I also spent time on holiday data. Accertify analyzed about 576 million transactions across five cyber week seasons, worth $73.1 billion. Volume grew fastest outside Black Friday and Cyber Monday. The Tuesday and Wednesday before Thanksgiving grew 61% and 66% from 2021 to 2025, compared with 39% and 37% for Black Friday and Cyber Monday. Between 11 p.m. and 3 a.m. Central, about 8% of transactions carried more than 14% of attempted fraud. The average attempted fraud order was about 1.9 to 2.3 times larger than a legitimate one. Their holiday data this year shows average order value and dollar volume both up sharply. I am also hearing that account takeover is still spiking, and I believe AI makes credential stuffing fast and systematic. On the job market, I know it is hard. Hang in there, update your AI skills, and be kind to yourself if you were laid off.

Key takeaways
  • Banks are blocking AI agents in online banking because mistakes and credential theft are likely and they do not plan to reimburse customers.
  • Visa has said the cardholder is responsible for agentic purchases, but the chargeback still goes to the merchant by default.
  • Merchants cannot yet use proof that an agent made a purchase to shift liability, because the chargeback rules do not cover it.
  • A high dollar agentic chargeback can be many times a merchant’s average order value, which makes the exposure larger than the volume suggests.
  • Fraudsters are early adopters of AI agents, and merchants who can identify agent orders often find most of them are fraudulent.
  • Airlines run on thin profit margins, and one fraudulent order can take 20 to 30 good orders to make up.
  • You cannot block what you cannot identify, so ask your fraud provider exactly where the AI agent identifier lives.
  • Device ID and behavioral biometrics can help flag agents, and a verification link such as Know Your Human can confirm the person behind an order.
  • A measured approach protects margin while the rules catch up, and the fraud manager should be part of the decision.
  • Holiday data shows earlier peaks, overnight fraud, and larger fraud orders, which makes it useful for planning and for talking to leadership.
Final takeaway

Deciding when to block AI shopping agents is not a yes or no question for every merchant. It is a question about who pays, what you can identify, and who in your company makes the call. Visa can say the cardholder is responsible, but until the chargeback rules say how a merchant proves it, the loss usually stays with the merchant. If you can’t identify an agent order, you can’t decide whether to approve, review, or block it. Start by asking your fraud provider where the signal is. Bring your fraud team into the conversation before the growth plan is final. And plan to be measured, because we do not have enough data yet to know whether agentic commerce is a growth channel or a margin problem.

Connect with Karisse Hendrick | LinkedIn
Host of the Fraudology Podcast
Award-Winning Cyberfraud Expert
Ecommerce Fraud Prevention Consultant
Startup Advisor, Keynote Speaker, and
Consultant to Fortune 500 merchants