AI shopping assistants can't check out safely until we can trust them

AI shopping assistants are moving from helping you find products to buying them, and the systems that would make that safe are not ready. At Money20/20 USA, executives from Mastercard and Trulioo said the market is at "day zero" commercially. The big open question is who pays when an agent gets an order wrong.

What changed

People already use AI agents to compare prices and plan purchases. The next step is letting them complete the purchase. Sumant Gandhi of Trulioo said closed pilots are underway. Both speakers said the technology is moving faster than the trust systems around it, and that adoption will depend more on trust than on how capable the agents are.

The trust problem

Seckin Yilgoren of Mastercard said the industry must be able to tell whether an agent started a transaction, whether the shopper approved it, and whether the agent is hallucinating, meaning making things up. Mastercard's framework for agent trust rests on four pillars:

• Identity: who or what the agent is
• Intent: what the shopper actually asked for
• Intelligence: whether the agent is behaving sensibly
• Controls: limits on what the agent can do

Who pays when it goes wrong

Liability is unresolved. Today, a cardholder with a problem calls their card issuer. Once a shopper hands authority to an agent, that path is far less clear. Current fraud and chargeback rules were built for in-store and regular online purchases. They do not cover a case like an agent ordering 500 items when asked for one dog toy. Merchants are unsure whether to accept agent orders, and shoppers are unsure where their protection lies.

What sellers face

Merchants have long treated automated orders as suspicious and declined them. Now some may be legitimate. Personal agents also hide who the real customer is, which weakens loyalty programs and rewards. Yilgoren named three missing pieces:

• The shopper's intent
• The identity of the agent itself
• Brand preference and loyalty data

Closing that gap is the goal of "know your agent," or KYA. Gandhi said it is hard because it spans the companies building agents, the people using them and the agents. Intent is especially tricky, since it can build up slowly over a long conversation.

What to watch

Yilgoren expects agent purchases to grow ahead of the holiday season, along with fraud and scams. He wants risk warnings to reach everyone before a transaction, not after. Fraudsters are also trying to turn legitimate agents into malicious ones. Merchants should track two early signals: how often shoppers browse with agents but never buy, and how often they get orders they cannot judge. Both speakers said the aim is not to block agents but to build enough confidence for the market to grow.