The model supported text-only responses, with no charts or visual explanations. This meant structure, hierarchy, and language had to carry the full burden of clarity.

​

We treated formatting and information design as primary UX tools rather than enhancements.

​

That makes it feel like craft, not limitation.

Unlike traditional product work, I was not designing static content. I was defining constraints around what the system could say, how it should say it, and where it should stop.

​

This required:

​

• Defining tone boundaries
• Establishing structural templates
• Anticipating edge cases
• Aligning closely with engineering and risk teams

​

The work shifted from crafting screens to shaping system behavior.

Instead of a separate page, AI was embedded within the existing search tab structure.

​

Users could move between Overview, Products & Offers, and AI-supported deep dive without leaving the search context.

​

This preserved continuity and encouraged broader exploration. AI remained scoped to the query rather than becoming the primary surface.

Because the model output was text-only, we relied heavily on formatting and hierarchy, such as:

​

• Clear headings
• Bullet points and spacing
• Short, structured sections

​

We tested multiple summary structures and found that short labeled sections with clear spacing improved scanability compared to dense paragraphs. Structured formatting reduced perceived complexity even when content length remained similar.

​

We treated structure not just as readability, but as a trust mechanism.

AI latency introduces uncertainty, especially in financial contexts.

​

Instead of a generic spinner, we designed loading states that:

​

• Reflected the user’s query
• Used non-linear motion to signal active generation
• Reinforced that the system was responding in the moment

​

Loading became part of the interaction design, not an afterthought.