UX Case Study — Enterprise AI Platform

AIDEX — Document Intelligence, Redesigned.

Transforming a fragmented, manual document platform into a unified AI agent suite serving legal, compliance, and knowledge management teams.

Role

Lead UX Designer

Duration

14 Weeks

Platform

Web App · Desktop

Team

PM · 2 Engineers · SME

Year

2024

0%

Reduction in document processing time

0

AI agents shipped across 6 categories

0%

Task completion rate in Round 2 testing

0.0

Average satisfaction score out of 5

01 — Problem Statement

The wrong tool for the job.

Legal and compliance teams were spending the majority of their working day doing tasks that were repetitive, error-prone, and cognitively exhausting — not because the work was complex, but because the tools forced them to work manually at every step.

"I always second-guess whether I caught everything. There's no way to know if I missed a PII item unless I re-read the whole document."

— Senior Legal Analyst, User Interview #4

01

Fragmented Tooling

Teams switched between 4–6 separate applications for summarization, classification, PII detection, and review. No shared history, no output continuity, no single source of truth.

02

No Trust in Output

Manually produced document summaries had inconsistent quality. There was no confidence scoring, no source attribution, no way to audit or verify AI output. Users couldn't trust what they were seeing.

03

Zero Automation

Every document required a human to initiate, configure, run, and check each step independently. No task chaining, no workflow reuse, no bulk processing. Every Monday was the same manual cycle.

Problem

Legacy Excel-based workflows — fragmented, manual, and error-prone

02 — Research & Insights

What the data actually said.

8 stakeholder interviews, a 2-week diary study, and competitive analysis across 6 enterprise document tools surfaced the real picture behind the daily workflows.

Method

Stakeholder Interviews

8 interviews across legal analysts, compliance officers, and document operations managers. Sessions were 60 minutes each with a structured protocol and 20-minute open exploration.

Method

Diary Study

14-day diary study where participants logged every document-related task, the tool used, time spent, and emotional state. Generated 340+ task entries across 8 participants.

Method

Competitive Analysis

Evaluated 6 competing platforms: Kira Systems, Luminance, ContractPodAi, Harvey AI, Relativity, and Casetext. Mapped capability gaps, UX patterns, and trust mechanisms across each.

"We've built elaborate Excel trackers, shared Google Docs, and Slack approval threads just to manage what should be one workflow. Every workaround is a product gap."

— Document Operations Manager, Diary Study Week 2

Key Pain Points

Pain Point

Context Switching

Users switched tools an average of 11 times per document review cycle. Each switch cost 3–7 minutes of reorientation time.

Pain Point

Output Anxiety

6 of 8 users expressed anxiety about whether they had caught everything. Manual review had no systematic completeness check.

Pain Point

Repetitive Cycles

70% of recurring workflows were identical week-to-week. No mechanism to save, reuse, or automate repeated task sequences.

Research

03 — Target Users & Personas

Who we designed for.

Three distinct user archetypes emerged from research. Tap each to expand.

LA

Ananya — Legal Analyst

Mid-level · Legal & Compliance Team · 4 years experience

Processes 15–25 contracts per week. Needs speed and accuracy above all else. Gets anxious about missing PII or misclassifying a document type.

Fears missing PIITool overloadManual repetitionFast auditable outputSingle workspace

Tap to expand →

CO

Rohan — Compliance Officer

Senior · Risk & Compliance · 8 years experience

Reviews analyst outputs before sign-off. Needs to trust the review process, not redo it. Manages a team of 5 analysts.

Can't verify AI outputNo audit trailConstant re-reviewConfidence scoresTeam visibility

Tap to expand →

DM

Priya — Document Ops Manager

Lead · Document Operations · 6 years experience

Owns the weekly document processing cycle. Needs to automate repetitive batch jobs without relying on engineering.

Manual batch cyclesNo workflow reuseEngineering dependencyReusable automationsBatch processing

Tap to expand →

04 — Information Architecture

Building the mental model.

I rebuilt the IA from scratch around three primary pillars: what AIDEX can do, how to automate it, and where work lives.

AIDEX Platform

Home

Dashboard · KPIs

Agents

Catalog · Config

Workflows

Templates · Builder

Workspace

Projects · Docs

Guide

Help · FAQs

05 — Design Process

From research to pixels.

A 14-week process across 5 distinct phases, each building directly on the last. No phase was skipped, no shortcut was taken.

01

Week 1–2

Discovery & Research

8 stakeholder interviews, 2-week diary study, competitive analysis across 6 platforms. Synthesized 340+ diary entries into 12 core insight clusters.

02

Week 3–4

Information Architecture & User Flows

Rebuilt navigation model from scratch. Mapped task flows for all 12 agent types. Validated IA with 3 participants via card sorting. Established the Configure → Process → Results pattern.

03

Week 5–7

Wireframing & Ideation

Low-fidelity wireframes for all 13 primary screens. Explored 3 approaches for the workflow builder. Node canvas with template-first entry won.

04

Week 8–11

High-Fidelity Design & Design System

Built 40+ component library in Figma. Designed all 12 agent flows, workflow builder, workspace, and AI assistant. 100% component coverage for engineering handoff.

05

Week 12–14

Usability Testing & Iteration

Two rounds of testing with 5 participants each. Task completion rate improved from 60% to 94% between rounds. Final handoff delivered Week 14.

06 — Before vs After

The transformation.

The same user goal — review and process a contract — through two radically different experiences.

← drag →
Before
After
Before
After

Before

4–6 separate tools for summarization, classification, PII detection, and review

No shared workspace or project continuity between sessions

No confidence indicators or source attribution on any output

Cannot chain tasks or save a workflow for reuse

Generic grey interface with no visual hierarchy

Users described it as: "feels like a form, not a product"

Average contract review cycle: 4+ hours end-to-end

After

12 AI agents under one unified roof with consistent interaction patterns

Persistent workspace with project history, member collaboration, and run tracking

Confidence scores, source citations, and override controls on every result

Agentic workflow builder with 8 pre-built templates and unlimited custom chains

Dark sidebar + warm white content surface — designed with a clear point of view

Users described it as: "finally feels like it was built for us"

Average contract review cycle: under 25 minutes

07 — Interaction Design

One pattern. Twelve agents.

The most important design decision was establishing a single interaction grammar that all 12 agents follow. Learn one, know all twelve.

STEP 01

Configure

Upload documents via drag-and-drop zone. Select agent-specific settings. Every agent has a unique configuration panel but identical upload experience.

STEP 02

Processing

Animated progress with step-by-step status: Uploading → Parsing → AI Analysis → Generating → Finalizing. Users always know what the system is doing.

STEP 03

Results

Agent-specific output view with confidence scores, source citations, export options, and direct path to Add to Workflow or Run Again.

Interaction

Home dashboard — KPIs, recent activity feed, and project grid

08 — Challenges & Constraints

What made this hard.

Every project has constraints. The goal is never to remove them — it's to design around them without compromising the outcome.

Constraint

12 Agents, 1 Pattern

+

Every agent has unique configuration needs and unique result formats. Making all 12 agents feel consistent while still being meaningfully different required 3 rounds of pattern iteration.

Constraint

Workflow Builder Blank Canvas

+

A blank node canvas caused every user to hesitate in Round 1 testing. The solution was template-first entry — users start from a pre-built chain and modify it.

Constraint

AI Trust Communication

+

Confidence scores were misunderstood — users thought "High 92%" meant 8% was wrong. Relabeling to "AI Confidence: Very High" resolved the confusion entirely in Round 2.

09 — Usability Testing

Testing what we built.

Two rounds of moderated usability testing with 5 participants each — legal analysts and compliance officers from the target user group.

Round 1 — Week 12

60%

Task completion rate on core workflows

Workflow builder blank canvas caused hesitation in all 5 participants

Confidence scores misunderstood — "92%" read as "8% wrong"

Step indicator in agent flow not noticed by 3 of 5 users

Export options not discovered — hidden behind icon, not labeled

Round 2 — Week 13

94%

Task completion rate after iterations

Template-first entry resolved all workflow builder hesitation

Confidence label change resolved misunderstanding completely

Larger step indicator with labels visible to all 5 users

Export moved to labeled button — discovered by 5 of 5 users

"The AI suggestions were surprisingly useful. It suggested adding PII Detection before my export step — I hadn't thought to do that but it made perfect sense."

— Usability Testing Participant #3, Round 2

10 — Impact & Results

What changed.

Outcomes measured from usability testing, stakeholder review, and engineering handoff assessment.

0%

Reduction in steps required for a standard contract review cycle vs. legacy platform

0%

Task completion rate in Round 2 testing across all core document workflows

0.0

Average satisfaction score out of 5 across all tested agent flows

0%

Of participants said they would switch from their current tool to AIDEX

0%

Reduction in engineering design ambiguity at handoff due to full component coverage

0m

Average contract review cycle time, down from 4+ hours manually

11 — Developer Handoff

Built for engineers too.

The design system was built with engineering handoff as a first-class output — not an afterthought.

🎨

Design Tokens

All colors, typography, spacing, shadow, and motion values exported as named CSS variables.

🧩

40+ Components

Every component built with Auto Layout, all states, and detailed annotation notes for interaction behavior.

📐

Spacing System

Strict 4px base unit throughout. All padding, margin, and gap values are multiples of 4.

Motion Specs

All transitions documented: duration, easing curve, and trigger condition. Page transitions at 300ms.

📋

13 Page Specs

Every screen annotated with component names, interaction notes, responsive behavior, and edge case states.

🤖

AI Architecture

Agent system architecture documented with input/output specs, confidence scoring logic, and processing state triggers.

12 — Key Learnings

What this project taught me.

The insights that will change how I design every AI product going forward.

01

Consistency is a feature.

Building one interaction pattern that all 12 agents followed was the single highest-leverage design decision.

02

Trust is a design problem.

Users don't distrust AI. They distrust outputs they can't verify. Transparency is not a nice-to-have — it is the product.

03

Template-first beats blank canvas.

Every time I gave users a blank canvas, they hesitated. Every time I gave them a starting point, they moved forward.

04

Empty states are the first impression.

Every empty state in AIDEX was designed as a guided entry point with a clear next action — never a blank wall.

05

Design systems pay off immediately.

By screen 20, every new page took a fraction of the time. The system was the most valuable artifact delivered.

06

Great design removes friction.

The best enterprise product design is about designing the human-AI collaboration layer so users feel more capable, not replaced.

Final Thoughts

AIDEX reinforced my belief that the best enterprise product design is not about adding AI features — it is about designing the human-AI collaboration layer so precisely that users feel more capable, not replaced.

Back to top

UX Case Study — Enterprise AI Platform

Ecomm — Document Intelligence, Redesigned.

Transforming a fragmented, manual document platform into a unified AI agent suite serving legal, compliance, and knowledge management teams.

Role

Lead UX Designer

Duration

14 Weeks

Platform

Web App · Desktop

Team

PM · 2 Engineers · SME

Year

2024

0%

Reduction in document processing time

0

AI agents shipped across 6 categories

0%

Task completion rate in Round 2 testing

0.0

Average satisfaction score out of 5

Contents

Overview
Problem
Research
Personas
Architecture
Design Process
Before / After
Interaction
Challenges
Testing
Impact
Handoff
Learnings

UX Case Study
Enterprise AI · 2024

01 — Problem Statement

The wrong tool for the job.

Legal and compliance teams were spending the majority of their working day doing tasks that were repetitive, error-prone, and cognitively exhausting.

"I always second-guess whether I caught everything. There's no way to know if I missed a PII item unless I re-read the whole document."

— Senior Legal Analyst, User Interview #4

01

Fragmented Tooling

Teams switched between 4–6 separate applications for summarization, classification, PII detection, and review.

02

No Trust in Output

Manually produced document summaries had inconsistent quality. No confidence scoring, no source attribution.

03

Zero Automation

Every document required a human to initiate, configure, run, and check each step independently.

Legacy Excel-based workflows — fragmented, manual, and error-prone

02 — Research & Insights

What the data actually said.

8 stakeholder interviews, a 2-week diary study, and competitive analysis across 6 enterprise document tools surfaced the real picture behind daily workflows.

Method

Stakeholder Interviews

8 interviews across legal analysts, compliance officers, and document operations managers. 60 minutes each with structured protocol.

Method

Diary Study

14-day diary study where participants logged every document-related task, tool used, time spent, and emotional state. 340+ task entries.

Method

Competitive Analysis

Evaluated 6 competing platforms: Kira Systems, Luminance, ContractPodAi, Harvey AI, Relativity, and Casetext.

"We've built elaborate Excel trackers, shared Google Docs, and Slack approval threads just to manage what should be one workflow."

— Document Operations Manager, Diary Study Week 2

Key Pain Points Identified

Pain Point

Context Switching

Users switched tools an average of 11 times per document review cycle. Each switch cost 3–7 minutes of reorientation time.

Pain Point

Output Anxiety

6 of 8 users expressed anxiety about whether they had caught everything. Manual review had no systematic completeness check.

Pain Point

Repetitive Cycles

70% of recurring workflows were identical week-to-week. No mechanism to save, reuse, or automate repeated task sequences.

Research

03 — Target Users & Personas

Who we designed for.

Three distinct user archetypes emerged from research, each with different workflows, anxieties, and success criteria. Click each to expand.

LA

Ananya — Legal Analyst

Mid-level · Legal & Compliance · 4 years

Processes 15–25 contracts per week. Needs speed and accuracy above all else. Gets anxious about missing PII or misclassifying documents.

Fears missing PIITool overloadManual repetitionFast auditable outputSingle workspace

Tap to expand →

CO

Rohan — Compliance Officer

Senior · Risk & Compliance · 8 years

Reviews analyst outputs before sign-off. Needs to trust the review process, not redo it. Manages a team of 5 analysts.

Can't verify AI outputNo audit trailConstant re-reviewConfidence scoresTeam visibility

Tap to expand →

DM

Priya — Document Ops Manager

Lead · Document Operations · 6 years

Owns the weekly document processing cycle. Needs to automate repetitive batch jobs without relying on engineering.

Manual batch cyclesNo workflow reuseEngineering dependencyReusable automationsBatch processing

Tap to expand →

04 — Information Architecture

Building the mental model.

The original platform had no clear mental model. I rebuilt the IA from scratch around three primary pillars.

AIDEX Platform

Home

Dashboard KPIs Activity Feed Projects

Agents

Hub Catalog Landing Pages Configure Flow Results

Workflows

Templates Builder Run History Results

Workspace

Projects Documents Members Settings

User Guide

Getting Started Agent Guides Workflow Help FAQs

05 — Design Process

How we got from research to pixels.

A 14-week process across 5 distinct phases, each building directly on the last. No phase was skipped, no shortcut was taken.

01

Week 1–2

Discovery & Research

8 stakeholder interviews, 2-week diary study, competitive analysis across 6 platforms. Synthesized 340+ diary entries into 12 core insight clusters.

02

Week 3–4

Information Architecture & User Flows

Rebuilt navigation model from scratch. Mapped task flows for all 12 agent types. Validated IA with 3 participants via card sorting.

03

Week 5–7

Wireframing & Ideation

Low-fidelity wireframes for all 13 primary screens. Explored 3 approaches for the workflow builder. Node canvas with template-first entry won.

04

Week 8–11

High-Fidelity Design & Design System

Built 40+ component library in Figma. Designed all 12 agent flows, workflow builder, workspace, and AI assistant.

05

Week 12–14

Usability Testing & Iteration

Two rounds of testing with 5 participants each. Task completion rate improved from 60% to 94% between rounds.

06 — Before vs After

The transformation.

The same user goal — review and process a contract — through two radically different experiences.

← drag →
Before
After
Before
After

Before

4–6 separate tools for summarization, classification, PII detection, and review

No shared workspace or project continuity between sessions

No confidence indicators or source attribution on any output

Cannot chain tasks or save a workflow for reuse

Generic grey interface with no visual hierarchy

Users described it as: "feels like a form, not a product"

Average contract review cycle: 4+ hours end-to-end

After

12 AI agents under one unified roof with consistent interaction patterns

Persistent workspace with project history, member collaboration, and run tracking

Confidence scores, source citations, and override controls on every result

Agentic workflow builder with 8 pre-built templates and unlimited custom chains

Dark sidebar + warm white content surface — designed with a clear point of view

Users described it as: "finally feels like it was built for us"

Average contract review cycle: under 25 minutes

07 — Interaction Design

One pattern. Twelve agents.

The most important design decision was establishing a single interaction grammar that all 12 agents follow. Learn one, know all twelve.

STEP 01

Configure

Upload documents via drag-and-drop zone. Select agent-specific settings. Every agent has a unique configuration panel but identical upload experience.

STEP 02

Processing

Animated progress with step-by-step status: Uploading → Parsing → AI Analysis → Generating → Finalizing.

STEP 03

Results

Agent-specific output view with confidence scores, source citations, export options, and direct path to Add to Workflow or Run Again.

Interaction

Home dashboard — KPIs, recent activity feed, and project grid

08 — Challenges & Constraints

What made this hard.

Every project has constraints. The goal is never to remove them — it's to design around them without compromising the outcome.

Constraint

12 Agents, 1 Pattern

+

Every agent has unique configuration needs and unique result formats. Making all 12 consistent required 3 rounds of pattern iteration.

Constraint

Workflow Builder Blank Canvas

+

A blank node canvas caused every user to hesitate in Round 1 testing. Solution: template-first entry.

Constraint

AI Trust Communication

+

Confidence scores were misunderstood. Relabeling to "AI Confidence: Very High" resolved the confusion entirely in Round 2.

09 — Usability Testing

Testing what we built.

Two rounds of moderated usability testing with 5 participants each — legal analysts and compliance officers from the target user group.

Round 1 — Week 12

60%

Task completion rate on core workflows

Workflow builder blank canvas caused hesitation in all 5 participants

Confidence scores misunderstood — "92%" read as "8% wrong"

Step indicator in agent flow not noticed by 3 of 5 users

Export options not discovered — hidden behind icon, not labeled

Round 2 — Week 13

94%

Task completion rate after iterations

Template-first entry resolved all workflow builder hesitation

Confidence label change resolved misunderstanding completely

Larger step indicator with labels visible to all 5 users

Export moved to labeled button — discovered by 5 of 5 users

"The AI suggestions were surprisingly useful. It suggested adding PII Detection before my export step — I hadn't thought to do that but it made perfect sense."

— Usability Testing Participant #3, Round 2

10 — Impact & Results

What changed.

Outcomes measured from usability testing, stakeholder review, and engineering handoff assessment.

0%

Reduction in steps required for a standard contract review cycle vs. legacy platform

0%

Task completion rate in Round 2 testing across all core document workflows

0.0

Average satisfaction score out of 5 across all tested agent flows

0%

Of participants said they would switch from their current tool to AIDEX

0%

Reduction in engineering design ambiguity at handoff due to full component coverage

0m

Average contract review cycle time, down from 4+ hours manually

11 — Developer Handoff

Built for engineers too.

The design system was built with engineering handoff as a first-class output — not an afterthought.

🎨

Design Tokens

All colors, typography, spacing, shadow, and motion values exported as named CSS variables.

🧩

40+ Components

Every component built with Auto Layout, all states (default, hover, focused, disabled, loading).

📐

Spacing System

Strict 4px base unit throughout. All padding, margin, and gap values are multiples of 4.

Motion Specs

All transitions documented: duration, easing curve, and trigger condition.

📋

13 Page Specs

Every screen annotated with component names, interaction notes, responsive behavior, and edge case states.

🤖

AI Architecture

Agent system architecture documented with input/output specs, confidence scoring logic, and processing state triggers.

12 — Key Learnings

What this project taught me.

The insights that will change how I design every AI product going forward.

01

Consistency is a feature.

Building one interaction pattern that all 12 agents followed was the single highest-leverage design decision.

02

Trust is a design problem.

Users don't distrust AI. They distrust outputs they can't verify. Transparency is not a nice-to-have — it is the product.

03

Template-first beats blank canvas.

Every time I gave users a blank canvas, they hesitated. Every time I gave them a starting point, they moved forward.

04

Empty states are the first impression.

Every empty state in AIDEX was designed as a guided entry point with a clear next action — never a blank wall.

05

Design systems pay off immediately.

Building the design system before screens felt slow at first. By screen 20, every new page took a fraction of the time.

06

Great design removes friction.

The best enterprise product design is not about adding AI features — it's about designing the human-AI collaboration layer precisely.

Final Thoughts

AIDEX reinforced my belief that the best enterprise product design is not about adding AI features — it is about designing the human-AI collaboration layer so precisely that users feel more capable, not replaced. The goal is never to impress. The goal is to make users feel like they are doing their best work.

Create a free website with Framer, the website builder loved by startups, designers and agencies.