resume

Designing Trust: How I improved adoption of an AI native platform for Regulation SMEs

A system-level behavioral redesign that replaced a rejected Appian prototype with a human-in-the-loop platform built for high-stakes regulatory work.

Role

Product Designer

Duration

4 Months

Client

Status

Shipped

Work

Contents

Team

Context

Discovery

Strategy

Design System

Redesign

Impact

Learnings

Strategy

📌Executive Summary

Highly skilled banking SMEs rejected a technically accurate AI regulation change management tool because it disrupted their workflows, felt like a black box, and lacked user agency in a high-stakes environment. I pivoted the project from a simple UI facelift into a system-level behavioral redesign, utilizing psychology principles to build user trust and keep humans meaningfully in the loop.

-12%

Error Rate

64 → 87

SUS Score

+60%

Task Speed

👥The Team & Cross-Functional Collaboration

Enterprise design is a team sport. Here is how I collaborated across product, engineering, and client stakeholders.

I led the UX strategy and end-to-end product design, collaborating daily in an agile loop with:

Lead of Engineering & AI Lead

Aligned on LLM capabilities, token parsing limitations, processing latency, and platform constraints.

Product Manager (Internal) & Product Owner (Client-Side):

Tied user experience milestones directly to business metrics, compliance mandates, and deployment roadmaps.

Regulatory SMEs (Users & Domain Experts):

Engaged directly through contextual inquiries, A/B testing, and ongoing validation loops to ground assumptions in real-world user workflows.

Co-designing and aligning on system behavior during a working session with the Client Product Owner, AI Lead, and Tech Lead.

🛑Context: Extremely high risk ($) domain

Every time the FDIC or SEC releases a new regulation document, Subject Matter Experts (SMEs) at large financial institutions undergo a grueling, manual workflow:

The cost of making mistakes is catastrophic. To put this in perspective, in 2024 alone, the SEC collected $8.2 billion in compliance-related penalties. Furthermore, governing bodies constantly alter the Code of Federal Regulations, while these financial institutions operate globally across overlapping jurisdictions.

The Proposed Solution

Our engineering team trained an AI model to automate this process. The AI broke down large documents into granular "obligations" and mapped them automatically. Instead of manual extraction, SMEs simply needed to review the AI's work and correct it where necessary.

The Problem: Technical Accuracy ≠ User Adoption

When the team ran an initial pilot with the SMEs, they refused to adopt the system. Despite the high accuracy of the underlying AI model, the system failed because:

It didn't fit into their existing mental models or workflows.

They completely mistrusted the automated output.

They felt a lack of user agency to properly correct what the AI had generated.

It felt like it was slowing them down rather than speeding them up.

I was brought onto the team to solve this. This wasn't a system capability issue; it was a human-system interaction failure.

🔍The Discovery: Diagnosing the Trust Deficit

To understand the friction, I looked past standard feature requests and analyzed recorded sessions of Regulation Change Analysts (SME stakeholders) working on cases in their current Excel-based workflows versus our pilot application.

The initial Appian pilot interface layout: Dense data tables, lack of bulk management tools, and obscured data origin reducing trust in output.

Key Research Findings

The Excel Preference: SMEs had a deeply ingrained psychological preference for Excel during the Applicability Assessment phase. Despite Excel's heavy cognitive load and friction, it provided total layout visibility and control.
The Real Success Metric: While the business focused on reducing manual hours, the SMEs judged the platform strictly on error prevention and ease of correction because the penalty for errors is too high.
Interaction Fatigue: Each document produced hundreds or thousands of obligations. Reviewing and editing them one by one on our pilot interface was a tedious, frustrating task.

🧠 The Psychological Framework

Rather than treating this as a superficial UI issue, I conducted academic research into automated workflows. I mapped the user resistance to 4 core psychological phenomena widely documented in human-automation interaction:

Automation Bias

The tendency to trust the system's output blindly over personal judgment, leading to catastrophic overlooked errors.

Complacency

A lack of cognitive vigilance that occurs naturally when automated systems appear highly accurate over long periods.

Loss of Situational Awareness

Analysts losing the "big picture" understanding of the larger regulatory landscape because they didn't do the manual extraction themselves.

Skill Degradation

The fear of losing specialized expertise over time due to a shift from active production to passive monitoring.

💡The Strategy: Adapting to User Comfort And Designing 'Deliberate Friction'

Focusing only on optimizing for speed was hurting trust and accuracy. I proposed 2 main changes in the workflow to improve that.

Workflow Augmentation

The Excel Integration

Export to familiar Excel layouts — no forced UI adoption

Deep links to exact source document sections for easier referencing

Intentional Friction Points

Human-in-the-Loop

Reviewer + Approver collaborative workflow

Mandatory "Mark as Reviewed" on every AI line item (Active Monitoring)

These suggestions caused a split between the PM and Product Owner due to the additional work being added by the intentional friction points and extra SME involved.

I suggested a A/B test to see the impact of the implementations to resolve the conflict.

A/B Test To Resolve Conflict

Speed Focus Workflow

15% error rate

~80% faster case completion

Low user confidence

Active Monitoring

>3% error rate

~50% faster case completion

Higher user confidence

Due to the high risk nature of the domain, the lower error rate was prioritized over the faster case completion rates.

🎨Design System Foundations

With the behavioral strategy validated by data, we moved forward with designing and building the entire web application platform from scratch. I established a custom design system tailored for high-density enterprise data representation.

💻The Custom Product Redesign (Deep Dive)

This section goes through some design decisions I made in the user interface to help