What Should an AI Shopping Agent Do? Discovery Research Has to Answer That First When Designing AI Shopping Agents

Ying Chen

8月20日

讀畢需時 7 分鐘

A high-level synthesis of how Discovery Research can help define user's shopping intents using examples from Amazon, The Home Depot, and Sephora's LLM chatbots.

*Originally published on Signals to Solutions substack here.

Before a product team decides how to build an AI shopping agent, there’s a prior question worth starting with: what should the agent actually help someone do? That question sounds obvious, but it can be easy to skip past once you realize that your team can build almost anything with the technical capabilities now. Discovery Research is one of the tools that can keep the question in front of the team, because its job is to find where a user need starts, not just how well a solution performs once it exists.

For context, “AI agents” mentioned in this article includes AI LLM chatbots with agentic capabilities, as well as specific AI agents within AI chatbots. I am taking the adoption of AI agent features in LLM chatbots here into consideration for this high-level discussion, since retailers are just starting to embed these agentic capabilities into their chatbots at the time of this writing.

In retail specifically, that need rarely starts at the product page. Someone buying running shoes may have started with a goal to take up running. Someone shopping for lumber at The Home Depot may really be repairing a deck. Someone who eventually buys moisturizer may have started by wondering why their skin feels dry. The purchase is usually the end of a chain, not the beginning of one. So if an AI agent’s job is to help with making purchases, the useful question is what part of the user journey is the agent actually being asked to help with.

It’s actually not as easy to answer as it sounds, because user intent isn’t one thing. In this article, I will look at examples like Amazon, The Home Depot, and Sephora (which have each built AI shopping assistants) to see how at least three different questions get bundled under the word of “intent”:

what kind of help a user wants (e.g., to learn about product options first or move products into the cart)
how a user is organizing their need in their own head (e.g., to focus on finding a specific product or generating a list of products to buy)
how far along a user is in forming the need at all (e.g., need is barely formed or nearly settled).

Treating all of that as one question may be where some agent design goes wrong before it starts; a team can build confidently and still end up creating something that does not solve the user’s real pain point. From reading industry sources, I think that Amazon, The Home Depot, and Sephora each seem to be leaning on one of these more than the others, and it is worth a look if you are desiging AI agents today.

What Kind of Help a User Wants

Let’s start with investigating the kind of help someone wants. Amazon’s Rufus, later renamed Alexa for Shopping, launched in 2024 and has reportedly drawn more than 300 million users during 2025. It sits inside the shopping environment - Amazon’s website and app - that customers already use, letting them ask in plain language about finding or comparing products (instead of typing keywords into a search bar).

Amazon Rufus within the Amazon App (Picture Credit: Here)

The trouble is that a query like “best running shoes,” without a clearly specified intent or context, can mean two different things. It can mean “help me find and compare best running shoes options,” or it can mean “help me understand what to look for when I want to buy the best running shoes before I decide on anything.” From the backend, interaction data will tell a team what happens next, like a user click on a specific product after the AI returns some product options. But it probably can not tell which of those two questions the user actually walked in with, since both can end in the same click on the same product choice. That distinction may still be findable, though: through conducting Discovery Research on a real user’s conversations with the AI agent and the actions they take, product teams can identify the kind of help that the user really needs from the agent, and therefore help the teams understand the user’s intent.

This also plausibly changes what the agent should say back to the user. Perhaps asking a clarifying follow-up question to ask if a user wants to “decide what to buy now” or to “find and compare options” allows the agent better help the user. And if after a round of conversation, the agent learns that the original “best running shoes” queries turn out to skew toward “decide-what-to-buy-now” moments, its job may be to narrow fast: ask two or three constraints, then return a product choice to the user. If they skew toward “explore” more options, the agent may need to instead walk someone through a list of criteria and what differs between product options before it recommends anything. So, the same query can return two different possible intents, and backend click-through data after the fact is unlikely to settle which one to build on its own. However, Discovery Research seems better positioned to.

How a User Organizes Their Need

Next is how someone organizes their need in the first place; namely, their mental model. The Home Depot’s Magic Apron, introduced in March 2025 across the website and app, answers questions and summarizes reviews on product pages. According to their news releases, The Home Depot has also built a Material List Builder for Pro Xtra members and a Blueprint Takeoffs tool that generates material lists and quotes. So as we can see here, there are three separate tools, which is itself a small clue that no single Home Depot AI placement fully answers what customers need.

The Home Depot’s Magic Apron on Product Websites (Picture Credit: Here)

What holds those tools together is a mental model that doesn’t map directly onto a product catalog. A Home Depot customer often starts with a renovation project (such as repair the deck or redo the kitchen), and purchasing a product is just one step within a larger goal. I think that is also why a Home Depot associate’s help is very valuable to customers in physical stores, as their first question is rarely asking which specific drill the customer wants to buy, but rather a broader question about what renovation project they are working on. They would then recommend products based on the customer’s project goals.

So, conducting Discovery Research here might mean following that project logic rather than the product logic: interviewing them about where they got stuck mid-project and investigating what they would need to complete their projects. One hypothetical and plausible-sounding finding from that kind of work is: perhaps customers don’t mainly get stuck on which product to buy, but on what order to do things in (e.g., tear out the old decking before or after buying the new lumber). If something like that turned out to be true, it wouldn’t just refine Magic Apron’s answers on a product page. It could point toward a different feature entirely, such as providing a “sequencing guide” rather than a smarter product-question box. This could be a decision that testing the existing chatbot’s accuracy in an eval seems unlikely to surface on its own (since an eval mostly checks whether the thing you already built works), not whether it was the right thing to build.

How Far Along the User’s Need Already Is

Last is how far along the user need is within the user journey in the first place. Sephora’s app inside ChatGPT, launched through OpenAI’s Agentic Commerce Protocol (ACP) and currently only available in the US at the time of writing, is a bet on meeting people even earlier: while a skincare need is still forming, but before anyone has decided to shop at all. A conversation that the user has within the Sephora app in ChatGPT might start with “why does my skin feel dry,” move to “what can I do?” and only later become “which moisturizer should I buy?” Only that last turn looks like a normal retail query.

Sephora App inside ChatGPT (Picture Credit: Here)

This is also the case where Discovery Research gets a little more complicated, since Sephora doesn’t own data to ChatGPT conversations the way it owns its app. The early and half-formed conversational turns happen inside someone else’s product, and a retailer can’t shadow what it can’t see. The available options to capturing user intent in this case seem mostly indirect: UX researchers could conduct diary studies where customers log what led them to a purchase before they arrive at Sephora’s product recommendations, or gather whatever aggregate signal a platform partnership like ACP might be willing to share back. Each of these looks more like a proxy for understanding the user journey and the moment of intent formation itself than an observation of it. So I think this case is harder to research than the other two, and it could be worthwhile for retailers to consider how they could learn about their users’ behaviors in these types of situations.

What This Means for What Gets Built

These three examples aren’t meant as a complete list of what “user intent” can mean in retail, and I’d be wary of treating them as one. I believe there are likely other configurations of AI agents a different retailer or category would surface, and possibly other dimensions of intent besides the three here. However, I think what they seem to show is that intent collapses several separate questions into one word, and that a team could answer those questions wrong while still ship something that tests reasonably well in an eval.

That, at least, is part of the case for Discovery Research doing some of its work before an agent gets built. Evaluative research (including the user evals, usability tests, etc.) can tell a team whether the agent it built seems to be doing its job well. But Discovery Research is better positioned to tell the team what that job actually is in the first place. So, maybe we should start with not just “how should we build this agent,” but “what should it actually do?” And that looks like a research question before it’s a design or engineering one.