Commerce has always been agentic, with in-store sales associates available to guide and assist customers as needed. Agentic Commerce is an extension of this in-store experience. While often defined as “AI search,” a feature that relies on generative AI to better understand user queries and deliver natural language responses, agentic is much more than this.
In our Agentic AI Strategy Masterclass, Agentic Commerce Unpacked, we define agentic commerce as a system of technologies that combine search, generative AI, and agentic capabilities. The underlying components of an agentic commerce system work together to interpret intent, choose next actions, and respect business constraints like inventory and pricing. They orchestrate and reunify:
- Search
- Conversation
- Reasoning
- Rendering
- Business Optimization
It’s these combined characteristic that form the Intent Box, a utility with agentic capabilities rolled into a familiar search UI.
Watch Coveo’s Agentic AI Strategy Masterclass, Agentic Commerce Unpacked, to see live demos of conversational product discovery, agentic merchandising, and AI copilots in action.
The Limits of the Traditional Search Box
With traditional search, when Jake types a query into the tried-and-true search box, the system searches for relevant content and retrieves it based on how an algorithm ranks it. This works well when Jake’s intent is clear, but the experience can go awry if Jake types, “I need shirts for gardening in hot weather.” Traditional search has a hard time with ambiguous queries like this, so Jake might type this into a separate front end – a chatbot.
When search and chat each handle a slice of the discovery experience, but don’t work together, the experience is fragmented and frustrating for users. Jake might get relevant results in search and a reasonable answer in a chatbot window, but the two don’t share context. That leaves him with the task of bridging the gap.
Search alone is limited because it removes the human reasoning layer and replaces it with query retrieval. This creates dead ends like zero results, too many results, incorrect results, and open-ended questions that have nowhere to go. These kinds of frustrating search experiences represent a significant share of sessions where purchase intent exists, but the experience can’t meet it.
Static search UI states compound this. The interface doesn’t adapt to the shopper because it’s simply not designed to do that. Everyone gets the same layout, no matter where they are on their shopping journey. The Intent Box is built to adapt to limitless scenarios. It interprets a shopper’s goals and adjusts both results and UI in real time.
Treating chat and search as two separate systems (e.g., two “brains”) creates two separate experiences. But pivoting exclusively to chatbots doesn’t solve the problem of intent. What retailers and customers both want is a unified experience. Routing between chat and search is ultimately a design compromise because:
Why Chatbots Are Not the Answer
- Context does not persist across channels
- Ranking logic and conversational logic are disconnected
- Most conversational overlays do not reason within constraints
One UI that can manage a single shopper’s varying needs and understand context is the ecommerce version of an in-person sales associate. It brings retrieval and reasoning together in a single place and interacts with shoppers like Jake intuitively versus forcing him to navigate between windows, tools, and experiences.
What Is an Intent Box?
The Intent Box is not a mere chatbot, but it’s also not a search box with generative AI stitched into it. It’s a UI that interprets what a shopper needs and dynamically renders the most appropriate experience — enabling what’s increasingly called conversational product discovery: the ability to find, refine, and evaluate products through natural dialogue rather than keyword queries alone. It incorporates several different capabilities including:
- A single entry-point for discovery: It provides one place for search, guidance, comparison, and follow-up questions
- It’s context-aware: It remembers what’s been returned in the session, so when Jake says “show me the red ones under $50,” it knows what “ones” refers to
- It can interpret layered intent: It goes beyond keyword matching and reasons about the goal behind the query
- Orchestrates multiple retrieval tools: Discovery agents work in the background, calling the right tools in the right sequence to return the most relevant result
- Decides how to render the experience: The UI is dynamic, waiting for the system to tell it what to display based on what the shopper needs
The intent box is an environment where search becomes one layout among many. The layout adapts to intent rather than defaulting to a ranked list of results every time. Here’s how that plays out from a shopper’s perspective:
- Transactional → optimized product results for shoppers who know what they want
- Exploratory → product education and broader assortment for shoppers still figuring it out
- Comparative → structured side-by-side views for shoppers weighing options
- Bundle-building → multi-product assembly for complex or high-consideration purchases
- Ambiguous → clarifying questions that help the system narrow intent before returning results
Discovery Layouts (The Critical Differentiator)
The Intent Box moves the search experience beyond a static, one-size-fits-all results page. Often, an endless grid of products—the kind of experience we’ve come to expect from traditional search—isn’t helpful in the context of a user’s query.
A chatbot can switch the experience to a conversational one, which can be helpful with ambiguous queries like Jake’s, but lacks the context of search relevance and a grounding in real-life things like inventory and availability. An agentic system uses discovery layouts to reshape the search interface (and thus the entire experience).
Discovery layouts are configurable UI recipes that let the system decide, in real time, what works for shoppers like Jake. The system can dynamically choose the most effective visual framework based on context, then populate that framework with the most relevant content based on five primary discovery layouts.
Product Education Layout
Recall that Jake needs shirts suitable for gardening in hot weather. Though he may be ready to buy, he’s uncertain and needs a solution to a problem–gardening in the heat. In this scenario, a grid of various shirts can be overwhelming. With a product education layout, the UI incorporates content that educates to help Jake learn the benefits of different fabrics like linen versus moisture-wicking synthetics. It categorizes options by breathability, UV protection, durability, thread count, and fit.
Comparison Layout
Jake’s drawn to moisture-wicking shirts with UV protection. Once the system recognizes this, a comparison layout renders a side-by-side view of shirts with these features. The agent dynamically selects the comparison columns because it reasons (from the earlier discussion) that for a gardener in the summer sun, these attributes are more relevant than things like sleeve length or the number of pockets.
Multi-Intent Layout
Like an onion, Jake’s needs are layered. He needs to stock up on gear, so his query becomes more ambigous and more complex, “I need summer gardening gear that includes overalls, a hat to protect my head and neck, and thorn-proof gloves.” The agent disambiguates these different intents within that one complex request, treating them as distinct tasks rather than one monolithic search. It dissects the request and renders separate, clearly labeled sections for each category, allowing Jake to evaluate each part of his “summer gardening kit” independently without losing the thread of his original goal
Bundle Layout
Jake’s a complicated man with complicated gardening needs. If his query is even broader, something like, “help me get ready for a summer of vegetable gardening,” the system recognizes he is looking for a complete solution, not a single item. The bundle or “kit” layout activates. It reasons across categories to assemble a cohesive package that may include a moisture-wicking shirt, a pair of ergonomic shears, and a kneeling pad. They’re rendered as a unified “Summer Starter Kit” with a total price and dynamically generated benefits explaining how each item will help Jake work more comfortably in the heat.
Conversational Refinement
After seeing the initial results, Jake can keep refining his search using natural language. If he says, “show me the green ones” or “only the ones with chest pockets,” the conversational refinement layout applies these as deterministic filters. Because the system is context-aware, it knows “the green ones” refers specifically to the gardening shirts it just showed him. The UI updates instantly, showing Jake exactly how his verbal feedback has narrowed the catalog and giving him the option to remove any filter if he wants to broaden his results again.
When Jake finally knows exactly what he wants, the system shifts into a traditional search-as-a-layout grid. Jake’s intent has become clear and the UI adapts to the type of transactional grid view that was once the default because it’s the most appropriate layout for this specific moment in Jake’s journey.
Why This Changes Performance Economics
Until now, AI in commerce has largely been about search and relevance. Machine learning predicts intent and optimizes result ranking so shoppers see products that make sense and are easier to find. Agentic commerce builds on this foundation, but adds reasoning and acts inside the real boundaries of your business.
Agents grounded directly in your catalog and live data have guardrails. Each response is shaped by constraints like availability, location, entitlements, and pricing. When a shopper gives a price limit, it interprets that as budget, checks what’s in stock, and refuses to invent options that don’t exist. The same pattern extends to B2B, where buyers work inside gated experiences, contracted assortments, and account-specific conditions.
Dead ends, overly broad result sets, and “zero results” moments are effectively missed opportunities. By continuously expanding, refining, and rewriting queries, and by using tools like commerce search and the search API, the agent can keep moving sessions forward toward products and actions that are both relevant and valid for that customer.
Why Merchandisers Should Care
Many generative and agentic systems are a black box. You send a command and get an answer, but you have no way to validate the system’s response. This is a major risk for commerce brands where you absolutely need a deterministic experience. In a black box scenario, the AI is hallucinating the UI and the logic because they are disconnected from the actual commerce engine.
With Coveo’s agentic commerce approach, merchandisers can maintain absolute control by defining the concepts and boundaries that make an interaction meaningful.
This is the purpose of discovery layouts, which allow merchandisers to specify exactly which layouts to serve. These configurations tell the agent which tools to use and how to dynamically render the UI. Because this orchestration happens server-side, it eliminates the need for constant front-end deployment cycles and brings control back to the business user by providing three important elements:
- Guardrails, not guesswork: The agent is free to reason, but it only uses the preset constraints and components you provide.
- Touch it, feel it, smell it: Before you publish, a dynamic preview lets you see exactly what will happen on the site.
- A/B test the agent: Since it is all config, you can test and iterate on these layouts just like any other part of the platform.
Agentic commerce doesn’t have to be synonymous with unpredictability. An effective agentic system should be adaptive within your boundaries, evolving the search box into an intent box that brings traditional search, conversational discovery, and reasoning together in a single dynamic UI. This is what allows digital experiences to feel as compelling as they do in a store, ultimately allowing digital and in-person models to converge and influence each other.