Something new is going to start showing up in chargeback reports, and most merchants will not see it coming.
AI agents are buying things now. Not browsing. Buying. A customer tells an assistant to reorder, to compare, to subscribe, and the assistant completes the checkout. The plumbing for this is published and in production: an open protocol for agentic checkout and a delegated payment specification that lets an agent present a payment credential on a customer's behalf [1][2], with the large commerce platforms wiring into it [3].
Here is the part that matters if you take cards.
The merchant stays the merchant of record
In every framework published so far, the business selling the goods remains the merchant of record. The agent platform does not become the merchant, does not become the seller, and does not take on the settlement relationship.
Meanwhile the card networks have moved on identification, not on liability. Visa's agentic work defines how a trusted agent is recognised and how an agent initiated transaction is signalled through the authorization message [4]. That is a real and useful step, because without it an agent order looks like a bot to your fraud stack. But recognising a transaction type is not the same as writing a new dispute right for it, and no network has published one that shifts the loss away from the merchant.
Which means that when a customer says I never told it to buy that, the case gets argued under reason codes written for a human sitting at a checkout.
Your evidence file just got thinner
Now think about what you normally win a dispute with.
- The IP address the order came from.
- The device fingerprint.
- The browsing session, the time on page, the add to cart sequence.
- A delivery confirmation to an address the cardholder has used before.
- Login history tying the order to the account holder.
Half of that gets thin when the buyer was never looking at a screen. The IP belongs to an agent platform's infrastructure. There is no device, no session, no dwell time. The behavioural signals your processor's fraud tools were built to read are simply not generated.
For a business already being watched on its dispute ratio, that is not a technology story. It is an underwriting story. Ratios move first. Reserves follow. If you want the arithmetic on why your bank reacts before you hit any published threshold, we wrote that up in what is my chargeback limit, and who actually sets it.
What has to replace it
The merchants who come through this clean will be the ones logging something different. Three fields, captured at the time of the order, not reconstructed six weeks later when the case arrives:
- Which agent transacted, and on whose behalf. The agent identity and the customer it was acting for, from whatever signal your checkout receives. This is what the trusted agent work is designed to give you, so capture it rather than discarding it.
- What the customer actually authorized it to do. The scope and limits of the mandate: the spend ceiling, the category, the standing instruction, the duration. The delegated payment specifications carry this idea, and it is the closest thing to a signature that an agent order has.
- Which cart it approved. The specific item set, price and total at the moment of approval, so a later claim that the agent bought the wrong thing can be answered with what the agent was shown.
That record is what replaces the device fingerprint. It is not a speculative requirement. It is the only material you will have to work with.
Ask your processor, in writing: can an agent initiated order be identified in my transaction data, and is the agent signal retained where I can pull it for a dispute response? A surprising number of stacks receive that signal and drop it. Finding out in January is cheap. Finding out during a dispute is not.
What not to do
Blocking agent traffic is the obvious reflex and the wrong one for most merchants, because it blocks real customers with real intent and real money. Agent originated orders tend to be high intent: somebody explicitly told the assistant to buy.
The other wrong move is assuming your fraud tooling will sort it out. Rules tuned on human behavioural signals will either decline good agent orders or wave through bad ones, because the signals they score are absent. That is a tuning project, not a setting.
Nobody has to solve this this week. But if you sell subscriptions, digital goods, or anything that bills on repeat, it is worth asking your processor one question before the next busy season.
What does my evidence file look like when the buyer is not a person?
FR