AI shopping agents are giving retailers a new challenge: How do they serve a customer when an agent handles the purchase?
Preparing for agent-led shopping requires changes to commerce systems, according to Anthony Ferry, CEO of omni-commerce intelligence platform Wayvia. He warns that merchants must prepare to serve two customers simultaneously — the human shopper and the machine agent.
While human consumers will still use visual websites to browse and discover products, he notes that retailers need accurate product data feeds and machine-accessible transaction paths so AI agents can check inventory, submit payment information, and complete purchases without clicking through web pages.
A Thoughtworks briefing cites research projecting that AI agents could drive up to 25% of global e-commerce volume by 2030. Agent-led shopping can compress the traditional multi-step consumer funnel into a single, automated interaction.
While Ferry expects people and AI agents to share the shopping process, he doubts retailers will redesign their websites entirely around bots. People and agents are trying to do different jobs, he reasoned.
“A person may want to browse, compare, and see how a product fits into his or her life. An agent wants a clean answer: Is this the right item? What is the actual price? Is it in stock? When will it arrive? Am I allowed to buy it?” he told the E-Commerce Times.
Ferry said retailers should support both human shoppers and AI agents using the same underlying information.
He cautioned that retailers should not force AI agents to click through pages built for people, as the agents may scrape, guess, or move on. “The website is not becoming obsolete. It becomes one of several doors into the retailer. It should still be the best door for a person. It just cannot remain the only door,” he explained.
Adjusting to Changing Marketing Metrics
Marketing teams have relied on conversion funnels, user session data, and behavioral tracking metrics to measure transactional success and failure. As checkout pages fade into the background for automated transactions, that changes.
Ferry noted that most retail measurement assumes that if you did not see a session, nothing happened. That is now a bad assumption.
For example, a shopper could ask an assistant to find a coffee maker under $100 that can arrive by Friday. The AI agent may compare 50 retailers and send one order. The winner sees the sale. The other 49 may see nothing.
Page views, bounce rate, and cart abandonment will still matter for people shopping on the site. But marketers also need to know whether the product qualified for the request, whether the retailer was considered, and what ultimately drove the purchase: price, delivery, a loyalty benefit, or a review.
According to Ferry, the biggest gap is intent. The agent platform may understand the shopper’s entire request while the retailer receives only a SKU and an address.
“Retailers need enough information to understand demand without turning agent shopping into surveillance. Otherwise, they will know what sold but not why,” he said.
What Clean Commerce Data for Agents Looks Like
Ferry defines clean commerce data by whether an agent can use it to complete an order, not simply by whether a product feed passes a formatting check. It needs to identify the exact product and variant, relevant compatibility or restrictions, and the price after applicable offers.
It must also indicate availability at the fulfillment location, delivery timing, and return terms. Just as important, an agent has to know when each answer was last updated. Yesterday’s inventory is not a promise you can safely make today.
That exposes how much legacy commerce technology was built for publishing pages rather than making live decisions, he added. Product content, promotions, store stock, and order management often sit in separate systems. An authenticated API may still deliver only a periodic snapshot.
“Retailers are at various stages of updating their legacy systems and databases. Some have done serious integration work; others are still stitching together feeds built for a different era,” he offered.
Ferry suggested a simple test. Can an authorized agent request a specific item, receive a reliable offer, and complete the transaction without encountering conflicting answers along the way?
“If not, the business is not ready, regardless of how modern the website looks,” he cautioned.
Autonomous systems need timely access to product data, pricing, and inventory. Shared data formats and transaction standards can help agents work across retailers.
Ferry does not expect one giant API for every merchant and category. What matters is a shared language for the basics: What product is this? What is the final price? Is it available? When will it arrive? Can it be substituted, purchased, or returned?
He expects a few common standards, with room for retailers and categories to add what they need. An agent should not need a new dictionary every time it enters a different store.
“Open banking is a useful comparison because identity and permission matter in both. Commerce is messier. A bank transfer does not usually involve size, color, local pickup, a warranty, a bundle, or a substitute product,” he added.
Model Context Protocol (MCP) can help an agent connect to a retailer’s tools, he continued, but it does not make the retailer’s data accurate or consistent.
“It is a way to open the door. The industry still needs agreement on what the agent should find when it gets inside,” he insisted.
Who Pays When an Agent Makes a Mistake?
Ferry explained that a shopping agent does not have to hallucinate a product to make an expensive mistake. A catalog error or missing shipping information can cause an agent to choose the wrong option.
“If that error is repeated at machine speed, what looked like a small data problem can become a significant headache,” he warned.
Some retailers are already addressing responsibility for agent activity in their terms. Target, for example, states that it is not responsible for the conduct of third-party AI shopping agents. Orders placed by approved agents remain subject to its standard return policy, and disputes are evaluated based on customer permissions, account protections, evidence of authorization, and applicable law.
Ferry said responsibility should follow the failure. A retailer or brand should be accountable for inaccurate product claims, price, or stock data it supplies. The company providing the agent should be accountable when its system goes beyond the customer’s instructions, such as changing a size, substituting a brand, or placing an order without the required approval.
“The merchant that accepts an order still has to deliver what it agreed to sell and deal with returns. The precise legal liability will depend on the contract, the transaction, and the consumer protection rules involved. There is not a universal answer yet,” he said.
However, from the customer’s standpoint, none of those distinctions should turn a refund into a dispute between two companies. Resolving those disputes requires a record of the original request, the information the agent used, every material substitution, and the final authorization.
“Retailers that make customers absorb every agent mistake might protect a line in their terms, but they are putting repeat business at risk,” he said.
