N^o 1Context & Product Reality

Bridging the "Trust Gap" in Autonomous Finance

ChatFin is an enterprise AI agent used by FP&A and Treasury teams to automate high-stakes financial operations. While the backend was powerful, the product faced a critical bottleneck: The "Black Box" problem.

Users were hesitant to authorize the agent for high-value actions because:

• Ambiguity: AI responses lacked a clear audit trail.
• UX Debt: Fragmented workflows led to "hallucination anxiety" during data queries.
• Scale Risk: The system lacked a standard interaction model for multi-step agentic tasks.

N^o 2My Role & Engagement Model

AI-Driven Product Designer (Consulting)

• Joined post-MVP, pre-scale
• Partnered with CTO, frontend engineering, and leadership
• Audited UX vs UI vs AI behavior
• Prioritized improvements aligned to deployment roadmap

Key constraint:

Incremental changes in a live system — no rewrites, only high-leverage fixes.

N^o 3Core Problem:

UX in Engineering-Led AI Products

Despite a powerful backend, the early experience hit a cognitive ceiling. The AI was functionally correct but contextually poor, leading to:

• Unstructured Outputs: Dense, unformatted paragraphs forced users to "mine" for data.
• Mapping Gaps: Poor translation between financial jargon and interactive UI components.
• Audit Friction: High effort required to verify AI claims against source ledgers.
• Decision Paralysis: Users spent 60+ seconds reading text instead of scanning charts.

This is the AI UX problem: Great intelligence, but weak situational comprehension.

N^o 4Approach:

Designing for Trust, Not Just Intelligence

I focused on decision-critical AI UX, not visual polish.

Key principles:

• AI outputs must be interpretable
• Users need to understand why, not just what
• Financial objects must behave predictably
• Systems must scale without increasing cognitive load

N^o 5Initiative 1:

Platform Audit & UX Diagnosis

Conducted a full audit across:

• Information architecture
• Conversational logic
• Component usage
• Data visualization patterns
• Terminology alignment with finance workflows

Outcome:

A prioritized UX backlog mapped directly to engineering sprints.

N^o 6Initiative 2:

The Framework: Stabilizing Agentic Behavior

To reduce "hallucination anxiety," I applied Object-Oriented UX (OOUX). I worked with engineers to ensure the LLM and the UI shared a unified mental model of financial objects.

Core objects became "Actionable Entities":

• Transactions: Transitioned from static rows to actionable nodes with "Verify" and "Flag" states.
• Cash Flow Trends: Replaced 180-word summaries with high-density line charts and interactive sparklines.
• Assumptions: Turned static predictions into editable modules, allowing CFOs to "stress-test" AI forecasts in real-time.

N^o 7Initiative 3:

Designing for Accountability

For a financial agent, "accuracy" isn't enough; "provenance" is the goal. I designed specific Explainability Patterns to build user trust:

• Step-by-Step Reasoning: Users can expand the agent’s logic to see how it arrived at a financial summary.
• Confidence Markers: Visual indicators for data points with high variance or missing inputs.
• Source Referencing: Direct deep-links to the ERP or bank ledger for every claim made by the agent.

N^o 8Initiative 4:

Optimizing for Scanning, Not Reading

The final refinement stage focused on information density and hierarchy. By shifting from conversational blobs to structured data tiers, we achieved:

• Immediate Hierarchy: Primary insights (Runway, Burn Rate) moved to the top of the chat node.
• Visual Anchors: 3-month trends are now instantly parsed via line charts rather than buried in paragraphs.
• Agentic Shortcuts: Contextual buttons (e.g., "Analyze Cloud Spend") anticipate the user's next question, reducing prompt fatigue.

N^o 9Impact

Quantifiable, Production-Level Impact

By maturing the interaction model, we transitioned ChatFin from a "Black Box" into a trusted advisor:

• +24% Weekly Engagement: Users shifted from occasional queries to daily operational reliance.
• -45% Support Volume: Clearer data provenance and "Show Reasoning" patterns reduced help desk escalations.
• 80% Faster Time-to-Insight: Redesigned layouts moved users from "reading to understand" to "scanning to act."

This wasn’t just a UI refresh — it was the design of a scalable system for agentic trust.

N^o 10What This Demonstrates

This project shows my ability to:

• Design agentic workflows for real users
• Improve AI systems without retraining models
• Balance intelligence, trust, and usability
• Lead UX in production, constrained environments

Strong AI needs strong UX — or it becomes noise.