A conversational navigation pattern designed for a complex B2B e-commerce platform, helping users move from intention to destination without needing to understand the platform's structure first.
The platform served a B2B e-commerce environment where different user roles coexisted: buyers placing recurring orders on a regular cycle, operations and logistics staff chasing specific deliveries or documents, and others managing account-level tasks. All of them landed on the same homepage.
Traditional navigation worked for users in a browsing mode. It worked less well for users who arrived with a specific goal and had to remember where the platform placed it.
The starting direction was well considered: surface the most-used actions at the top of the homepage, reduce the distance to common destinations. A set of task cards, each labelled in first-person language, sitting below a welcome header. Clear, direct, and easy to validate.
While working within that direction, a question came to mind: what happens as the platform grows and the number of roles, tasks, and destinations expands? A card grid works well when the set of actions is stable. Over time it risks becoming another layer of navigation to scan through. That observation opened up a different angle worth exploring alongside the original proposal.
The request came from a real observation: users were navigating around looking for things that should have been one step away. A previous order. A specific delivery reference. A page they had used before but could not locate again.
In a B2B context, this friction is particularly expensive. A buyer repeating a weekly order does not want to browse. A logistics manager chasing a delivery status does not want to scan a dashboard. Both users know exactly what they need. The interface was making them work to get there.
At the same time, users are increasingly comfortable with prompt-based interfaces. Not because those interfaces are AI, but because typing what you need and getting a direct result has become a familiar interaction model. I was not designing for a trend. I was designing for a friction pattern that the trend made more visible.
My hypothesis was that the homepage hero could work as a task entry point rather than a dashboard preview. Instead of presenting options, the interface could ask a direct question: what would you like to do today?
A buyer types a product name or order reference and lands directly. A logistics user types a delivery number and gets there without navigating through the buyer's workflow. Same entry point, different destinations, no hierarchy imposed on either role.
The visual treatment followed from that logic. A large conversational field, no box, no label, sitting as the primary element on the page. The only interaction cue is a typewriter animation cycling through example queries. Motion becomes the affordance. The field reads as active and inviting without a traditional input wrapper. Below the field, chips surface previous searches and recent tasks, reducing the cost of returning to something already in progress.
This was not designed as a chatbot or an AI assistant. For the MVP, I proposed a lightweight intent-matching layer: predictable, controlled, and straightforward to validate.
The system recognises common task language, product names, order references, and key platform destinations, then suggests the most relevant path. The results are deterministic. No generation, no ambiguity, no risk of misleading a user who needs precision.
That was a deliberate product decision. Reliability mattered more than flexibility at this stage. A B2B user acting on behalf of a company cannot afford to be routed to the wrong place.
Any fixed layout for a shared homepage in a multi-role environment forces a prioritisation decision. The conversational field avoids that problem by letting each user's language do the routing. The interface does not need to know which role is logged in. It responds to intent.
It also has no ceiling. As the platform adds destinations, products, or order types, the matching layer extends without changing what the user sees.
The structure was designed to be scalable to a more complex matching keywords logic or AI even. The same entry point, suggestion layer, and intent-to-destination mapping could be extended to handle more flexible or contextual queries.