The Brands Winning Agentic Commerce Are Rewriting Catalogs Around Customer Questions

Retail sites are racing to ship AI shopping agents before proving the agents can actually help shoppers buy. The demos look polished. The investor decks read clean. But underneath, the product data either reflects how customers think or lives in catalogs a merchandising team wrote a decade ago. That gap is where agentic commerce breaks down. A brand can ship the visible agent, but without rebuilding the data layer beneath it, the result is a tool that answers confidently and wrongly. That’s worse than no tool at all.

Nagesh Koritala, a Digital Product Manager at Kohler Co., owns the conversion funnel at Kohler.com from product detail page through checkout, payments, and AI-powered shopping assistance. Before product management, he spent years in front-end development at Google, Crate and Barrel, and Verizon, building computer vision prototypes and product-recognition tools that could identify a Kohler faucet from a customer's photo years before large language models reached the boardroom. His argument is that the smartest shopping agent is no smarter than the catalog it inherits.

"I strongly recommend focusing on the foundational layer before bringing in any kind of shopping agent," Koritala says. "These tools are easy to plug in now. But the quality of the response is what really matters."

Shopping agents are only as smart as their inputs

The behavior has changed faster than the catalogs. Customers who used to compare specs across browser tabs now ask an agent and expect a single useful answer. The brands meeting that expectation are the ones whose data layers were already part of a product roadmap before the agent arrived.

Koritala starts every conversation with the same question. "Leaders who are thinking in this space definitely need to understand what outcome they're driving towards. Is it having more sales-driven, or is it more about customer satisfaction? Strongly focusing on the foundational layer is more important than bringing any kind of shopping agent," he says. Build it in the wrong order, and the agent ships as a polished interface over a catalog the system never bothered to organize.

Build against the standards customers already use

For Koritala, the practical move is to stop inventing internal taxonomies and start aligning to schemas that major search and shopping platforms have already standardized. Google's product data specification is the most public example, and the discipline of mapping attributes to it forces a brand to write product information in the format the broader commerce web expects.

That foundation pays off twice when the experience moves beyond text. Google's Universal Commerce Protocol updates point to where the interface is going, with agents pulling real-time inventory and pricing from retailer catalogs and handling multi-item carts the way a person would. Multimodal interactions, where a shopper asks a question or uploads a photo of their bathroom, rest on that same foundation. None of it works if the catalog underneath does not speak the language the agent expects.

Compatibility and relationship metadata between SKUs is the part that most legacy catalogs skip. An agent that can tell a shopper which trim kit pairs with which faucet body, which finish is in stock with which spout, and which accessories ship together is doing work that no product detail page ever asked the catalog to surface. That work now belongs upstream, in the data layer, because every downstream interaction will draw from it.

Even the emotional layer of the conversation rests on the same data discipline. "If someone is frustrated talking about the product not being received by them on time or received in a very bad condition than they expect, understanding the emotional intent of the conversation is really important. Humans have a lot of empathy, machines may or may not have, but they are doing a good job there right now," Koritala says.

The shipping weight tells you whose side the catalog is on

The clearest illustration of the internal-versus-customer-facing data gap is mundane. Koritala uses shipping weight because the example is small enough to be obvious but large enough to risk losing a sale. "I may not think about shipping weight because I feel like it's something that I will handle for you, but the consumer may think, okay, if you're leaving it in front of my door, it's so heavy, I need to be able to carry it. Giving proper information on those foundational elements is really important," he says.

That reframing applies across every attribute a legacy brand catalogs. Weight, dimensions, installation requirements, compatibility with adjacent SKUs, and included accessories. Each was originally entered for a logistics or merchandising purpose. Each now has a second job: answering the question a customer asks an agent at the moment they decide whether to buy. The brands that write their attributes for a customer get usable answers from the agent. The brands that wrote them for a warehouse get AI summaries that flatten the catalog into guesswork.

What plug-and-play actually costs

The shortcut is to skip the data work and buy an off-the-shelf agent that promises to handle everything. Koritala has watched what happens when brands take it. "I have seen examples where people moving really fast and adopting AI are convincing those shopping agents to sell the product for $1," he says. Koritala is pointing to the Chevy Tahoe dealership incident, the one where a customer talked a generic chatbot into agreeing to a one-dollar SUV. The dealership pulled the bot. The screenshots are still online.

Hallucination is the second problem stacked on top. "AI hallucinates a lot. It also confidently tells something, even if it doesn't work properly; it gives out the answer like, I'm pretty sure it works out really well," Koritala says. An agent that fabricates a specification with confidence creates a credibility problem that the brand has to clean up across every channel where the answer gets shared.

The cost of automating the wrong moment is the relationship

Allocation is the work. The same judgment that finds the right place for AI is the judgment that keeps it out of the wrong one. Koritala draws a hard line at conversations where authenticity is the entire point of the interaction. "You don't want to put AI where a patient is calling a doctor, where they have negative thoughts. You can mimic the voice of the doctor, but when the patient gets to know you did not talk to a doctor, you have spoken to an AI bot, the complete trust and the feedback that the system gave falls down," he says.

For commerce, the analog is the moment when a shopper needs human reassurance that a product will arrive intact, fit a specific space, or solve a problem the agent has not been trained to handle. The brands that route those moments to a person and use AI to handle the volume around them preserve the relationship. Koritala's own workflow reflects the same logic in reverse. "Something that is easy, rudimentary, that can be done without your overseeing capacity, let them do it, and you take the most challenging one and solve that, and then figure out the next thing," he says. Let the agent take what it can handle. Keep human judgment where the cost of an automated mistake is highest.

A wrong answer tells you what your data does not say

Koritala uses a simple test to measure whether his data layer is doing its job. A wrong answer is a hallucination, which is a technical problem to be patched. A stupid answer is the agent failing to retrieve information that the brand has, but has not connected to the relational map the agent searches. The first is unavoidable in the current state of the technology. The second is entirely a function of how seriously a brand treats its first-party data and channel architecture.

The fix is the kind of slow work that no investor deck rewards, teaching the agent what the brand actually knows. "Shopping agents are really helpful for people, but they have to be trained. Do not think it's something I'll buy in a market and plug it in. Just dedicate some time, give it some love, and they will do great jobs. It's like a new intern. They have a lot of capacity, but they don't have knowledge," Koritala says. Treat the agent like an intern, and the customer experience holds up over time. Treat it like a vending machine, and the customer ends up with whatever fell out of the slot.