Retailers Face New Challenge: Marketing to AI Shopping Agents

A fundamental shift is underway in retail: artificial intelligence agents are emerging as a distinct customer class that researches products, evaluates options, and increasingly completes purchases with minimal human oversight.

When Wharton professor Kartik Hosanagar recently searched for a pomodoro timer using ChatGPT's Atlas browser, he received product recommendations complete with images, prices, and purchasing details—many from Walmart—without ever visiting a retailer's website. The experience illustrated how AI-mediated commerce is moving from concept to reality.

Walmart's partnerships with OpenAI around Agentic Commerce Protocol (ACP) and with Google for Universal Commerce Protocol (UCP) signal major retailers are preparing for a world where AI systems, not humans, conduct product research and make buying decisions. These protocols enable retailers to share product catalogs with AI platforms and allow agents to execute transactions.

Two modes of AI shopping

Current AI-assisted shopping operates in two distinct modes. In assistant mode, users describe their needs and the AI researches options, but humans evaluate and complete purchases. In agent mode, users provide parameters—"Order running shoes with ankle support, size 10, under $100"—and the AI selects and purchases a product.

Fully autonomous purchasing remains challenging. When OpenAI launched Instant Checkout in September 2025, allowing purchases directly within ChatGPT, conversion rates ran three times lower than traditional click-throughs to retailer sites. OpenAI discontinued the feature five months later. The more viable near-term scenario involves conditional automation: users select products and instruct agents to complete purchases when specific conditions are met, such as price drops.

Even in "human present" mode where users approve final purchases, the agent performs the evaluation and selection. This means AI systems, not humans, become the primary audience for product information and brand messaging.

The technology enabling agent commerce

Three infrastructure layers make autonomous shopping possible. Protocol layers like UCP and ACP allow agents to communicate across platforms, preventing fragmentation where each merchant requires custom integration. Commerce layers provide machine-accessible interfaces for querying inventory and executing transactions—capabilities Shopify, Etsy, and Salesforce are already building. Governance and payment layers must verify user authorization and establish liability frameworks, though industrial-grade systems for this don't yet exist.

Why decades of marketing science may not transfer

Retailers face a knowledge gap: marketing science has spent decades understanding how biological neural networks (human brains) respond to pricing psychology, social proof, scarcity signals, and visual design. Artificial neural networks may respond to entirely different decision factors.

Marketing organizations built around human psychology must now consider that AI agents have different biases, framing effects, and decision rules. The $19.99 pricing that converts humans may mean nothing to an algorithm. Product page layouts optimized for human attention go unseen when agents make selections.

Why it matters

The emergence of AI agents as customers represents a system-level market restructuring, not simply a new distribution channel. Retailers who treat it as the latter risk becoming back-end fulfillment operations for platforms that control product discovery and customer relationships. Amazon is betting on owning the shopping agent through its Rufus assistant. Walmart is hedging by making its catalog accessible to external AI agents while keeping checkout within its own environment—a strategy that preserves the customer relationship even when discovery happens elsewhere. Most marketing organizations remain structured around human decision-making. Companies that wait for consumer behavior to fully shift before adapting may find the balance of power has already moved.

These details were first reported by Kartik Hosanagar, co-director of Wharton's Human-AI Research Center, writing in Harvard Business Review.