AI agents that shop on behalf of consumers are movingfrom concept to reality. Visa and Mastercard are giving AI agents their own credit cards with spending authorization. PayPal is building agent-friendly APIs. These are production-ready systems that will enable autonomous shopping agents to execute thousands of transactions per minute. And some retailers are getting on board: Walmart and Amazon have both announced plans to launch AI-powered shopping agents.

Missing from the conversation, however, is who pays the bill when these agents make mistakes or are manipulated. Say an AI agent buys the wrong thing, or the product isn’t to the customer’s satisfaction. Who bears responsibility? The consumer who authorized the agent? The company that built it? The merchant who accepted the transaction? The payment processor that enabled it?

Current liability frameworks don’t address these scenarios. Without taking steps now, retailers could face a liability gap that makes them responsible for losses they never anticipated.

**The Fundamental Challenges of Agentic Commerce **

**When AI “chooses,” who is at fault? **Current return policies assume a human buyer made the decision. With AI intermediaries, retailers must clarify whether agent-driven purchases are covered under standard return and refund policies. If a consumer asks their agent to “order running shoes under $150” and the agent selects shoes that don’t fit the consumer’s actual needs, it’s unclear who made the purchasing decision.

**The product metadata problem **AI agents rely heavily on structured product descriptions, specifications and metadata to inform their buying decisions. That means that when product information is incomplete or inaccurate, AI agents are more likely to make purchases that don’t meet consumer needs.

Retailers may see significant increases in returns due to AI misinterpretations of product specifications. A clothing item described as “fitted” might be interpreted differently by an AI agent than a human shopper would. Electronics with compatibility requirements become particularly problematic when AI agents lack contextual understanding of technical specifications.

**The friendly fraud amplification **Chargeback fraud already costs merchants an estimated $50 billion annually. AI purchases will amplify this problem. Consumers may be able to claim “the agent made a mistake” or “bought something I didn’t want.” This creates a new category of dispute that existing chargeback protections don’t adequately address.

The challenge extends beyond simple returns. Suppose an AI agent learns consumer preferences over time and makes a purchase based on that information. Whether these purchases are considered “authorized” purchases is a matter of debate and requires much more clarity for retailers, consumers, payment processors, banks and the agent providers.

Revisit Policy Frameworks

Before enabling agent-based shopping, merchants should demand clarity about who the “purchaser” is in AI-mediated transactions. This determination drives policies touching chargebacks, returns, warranties and liability disputes. Some merchants may try to shift accountability to AI providers rather than accepting it themselves.

Retailers should make sure to negotiate those liability terms now and set clear contractual terms about responsibility for agent-driven purchases before widespread adoption begins. This should apply to return policies as well. For example, when agents make purchases based on flawed product recommendations or incomplete data, merchants should have a clear understanding of whether that responsibility will fall to the payment provider, the consumer, the AI company making the purchase or the retailer itself.

Retailers should also ensure payment processors make detailed transaction records and prompt logs readily available that correspond to each agentic purchase. This data is essential for resolving disputes and defending against claims of unauthorized AI purchases.

Competitive Differentiation

Retailers that offer generous return policies, even for agentic purchase decisions, could boost consumer trust — and share of wallets — in the near term. However, they could potentially expose themselves to systemic abuse if they adopt overly lenient return policies. The balance between maximizing customer trust while minimizing risk will determine which merchants succeed in the agent commerce era.

The reality is that there is no “one size fits all” solution. Retailers will need to consider factors such as the lifetime value of their shopper community, the category of products they sell (luxury versus mass-produced goods versus digital services), and perhaps even which products they want to make agent purchasable.

With the convenience of agentic shopping combined with trusted merchant policies, consumer adoption will begin to accelerate. Merchants that recognize this reality and build appropriate defenses will thrive, as agentic commerce presents a genuine opportunity to grow. For those who are unprepared, significant risk awaits. The window to prepare, however, is closing quickly.

Kris Nagel is the CEO of Sift, an AI-powered fraud platform that protects brands like Poshmark, Yelp, DoorDash and Everlane. Previously, Nagel was COO at identity and access management company Ping Identity (NYSE: PING), where he took the company public in 2019. You can connect with him on LinkedIn.