Contextual Chat

Designing a context-aware AI assistant for investor workflows - so teams reach actionable insight under pressure, not just more data.

AI Search B2B SaaS Fintech

View Case Study

Role

Product Designer II
Systematic · Remote

Timeline

2023 - Present

Focus

Human–AI interaction Context systems Query states Inline assistance

Speed

<30s

from 30–40 minutes

Engagement

25%

session time up

Preference

70%

preferred chat flow

Before

34%

query failure rate

The Hook

Investors didn’t lack data. They lacked speed to insight.

The existing experience rewarded users who already knew exact filters, taxonomy labels, and query structure. Everyone else kept re-running searches, reopening tabs, and trying to reconstruct context from memory. Data existed. Decision velocity didn’t.

Problem Visualization

Search workflows created friction exactly when confidence should increase.

The workflow looked functional on paper, but failed in real decision sessions:

Filter fatigue from too many rigid query controls
Zero-result dead ends with no recovery guidance
No exploratory mode for uncertain or broad intent
Context loss between sessions and repeated comparison effort

Screen recording of filter-heavy discovery workflow across companies and profiles

Research & Insights

The issue wasn’t lack of data quality. It was how users navigated uncertainty.

34% query failure

More than one in three searches led to no usable output or forced query restarts.

Broad to narrow behavior

Users started with exploratory language, then got trapped by rigid filter-first architecture.

Taxonomy fatigue

Users spent too much effort translating intent into system vocabulary before finding signal.

Persona: Associate Analyst · speed over perfection Persona: Portfolio Lead · context continuity across sessions

User journey map: broad query, system filters, refinement, poor results, and retry or drop-off

The Tension

The team had two obvious solutions. Both were incomplete.

Option A

Improve filters. Better controls, better labels, better defaults.

Option B

Add a chatbot. Let users ask for insights conversationally.

→ Filters without intelligence remain rigid. Chat without context becomes generic. Both fail in isolation.

Key Decision

We didn’t build a chatbot. We built a system that understands context.

Context had a concrete definition in the product:

Screen state: what data, filters, and entities are already active
Behavior: what patterns the user has shown in prior interactions
Intent: what the current query implies, even when phrased ambiguously

Solution Approach

Five design systems that made AI useful in live investor workflows.

Context awareness

Assistant reads current screen state so answers are tied to live workflow context, not generic summaries.

Inline assistance

Support appears where decisions happen, reducing tab-switching and preserving cognitive flow.

Smart prompts

Prompt chips guide users toward high-value actions without forcing rigid command syntax.

Memory across sessions

System remembers prior exploration to avoid repetitive setup and repeated discovery loops.

Actionable outputs

Responses produce next-step actions: compare, shortlist, export, and continue analysis.

System architecture diagram for the context-aware AI assistant

Design Evolution

From loose ideas to operational UI that teams could trust.

Early concepts

Existing workflow - fragmented, multi-step, and difficult to navigate.

Existing investor workflow - fragmented, multi-step paths across systems

Refined mid-fidelity

Explored ways to bring AI directly into the workflow and reduce navigation overhead.

Mid-fidelity exploration - research notes mapping inline assistant and layout patterns

Inline assistant Unified search & AI Less context switching Collapsible assistant

Production candidate

Final direction - a context-aware assistant embedded within the core experience.

Production candidate UI - active chat with history, structured responses, and in-workflow analysis

Interaction Details

Micro-interactions that removed friction without stealing attention.

Chat expansion behavior

Assistant opens progressively to preserve focus and avoid layout shock in dense data views.

Prompt chips + dual-mode

Users can move between classic search and AI mode without losing work-in-progress context.

Notification states

Subtle status updates communicate model progress and confidence without noisy interruptions.

Impact

Measured outcomes after launch

70%

Users preferred chat-assisted workflow

25%

Increase in session time

<30s

Time to insight

What Didn’t Work

Wall-of-text outputs were technically correct but operationally useless.
Early chat felt detached from user context and repeated obvious information.
Too much AI freedom made results feel inconsistent and hard to trust.

Learnings

AI is useless without context.
Proactive systems outperform reactive ones.
Structure builds trust in AI.

My Role

Led UX for the AI interaction model, defined the context-driven product experience, and translated ambiguous AI behavior into a predictable interaction system.

Collaboration

Worked tightly with PM and engineering on state architecture, interaction sequencing, response patterns, and launch measurement criteria.

Next Steps

Where this product can go next

Personalization layer that adapts prompt strategy by role and behavior.
Proactive alerts for key shifts in followed entities and portfolio signals.
Collaboration features for shared analyst notes and team-level context memory.

Next projects

Deal Room Structured deal execution on the company profile B2B SaaS Deal Room

EasyEat Customization redesign for multi-country ordering FoodTech Mobile UX

Let's build smarter products together.

in ✉ Resume