Discovery & Taxonomy
We studied CADEX’s operational workflows and mapped the real-world email landscape.
- Email analysis
- Operations workshops
- Category mapping
- Edge-case identification
- Workflow analysis
CLIENT
CADEX — AI Email Workflow Automation
TEAM
UX/UI Designers, Product Team, Developers & QA
Services Used
The Challenge
CADEX handles hundreds of customers and partner emails every day. Messages can contain multiple issues, varying urgency, and important information hidden inside PDFs, scans, or screenshots.
An earlier machine-learning model achieved over 90% accuracy in testing but dropped to around 40% when exposed to real-world email traffic.
The challenge was to turn a promising AI experiment into a reliable, production-ready system that could handle messy email data at scale.
Key Pain Points
- Unstructured email content
- Multi-intent emails
- Inconsistent real-world data
- Manual classification and routing
- Low production accuracy
- Risk of SLA delays
Research & Discovery
User & Operational Insights
We worked closely with CADEX’s operations team to understand how emails were handled in real-world conditions.
Team Members Interviewed24
Role Coverage6-roles
Email Types Reviewed50+
Analytics Review
We analyzed existing email workflows to identify patterns and establish a clear classification structure for automation.
Emails Analyzed50K+
Time Period6 months
SLA Sensitive RatePersonalized resume experience
Key Research Findings
73%
Emails contained multiple intents
81%
Production data was unpredictable
65%
Attachments contained critical information
Intelligent Email Preprocessing
The preprocessing layer prepares every message for downstream AI models.
- Email cleaning
- Content normalization
- PDF extraction
- Image OCR
- Attachment processing
- Structured data preparation
Hierarchical Intent Classification
Instead of forcing every email into a single category, the system progressively determines the message’s intent.
- Coarse intent detection
- Fine-grained classification
- Multi-intent handling
- Edge-case detection
- Manual review routing
Smart Workflow Routing
Once an email is classified, the system determines where it belongs operationally.
- Intent-based routing
- Queue assignment
- Context preservation
- Workflow-specific tagging
- Automated operational handoff
Measurable Impact
40%
Classification Accuracy
AI classification improved dramatically when moving from the earlier production model to the new workflow.
Before:40%
After:98%
+65%
SLA Performance
Improved routing and faster handling helped increase service-level adherence.
Before:Baseline
After:+65%
−30%
Average Handling Time
Automated classification and routing reduced the time agents spent on email triage.
Before:Manual-heavy handling
After:30% reduction
−50%
Manual Workforce Requirement
Automation reduced the workforce requirement associated with email triage.
Before:Fully manual-heavy process
After:~50% reduction
Model Adaptability
Human feedback transformed the AI from a static model into a continuously improving system.
Before:One-time model
After:Feedback-driven learning
Methodology & Testing
Email Taxonomy Design
Mapped real-world email types, edge cases, and operational workflows into a structured classification system.
Multi-Agent AI Design
Separated preprocessing, classification, routing, and feedback into specialized AI responsibilities.
Hierarchical Classification
Used progressive classification to handle complex and multi-intent messages more reliably.
“We’d tried machine learning before, but it never survived contact with our real email traffic. Galaxy helped us turn that idea into a production system that actually works. Accuracy jumped, SLAs impr
Operations Leadership, CADEX
Operations Leadership, CADEX
Key Learnings & Insights
Real-world data changes everything
A model that performs well on curated data can behave very differently in production.
Complex problems need modular AI
Separating responsibilities makes AI systems easier to monitor, tune, and scale.
Automation needs human oversight
Manual review provides a safe path for ambiguous or high-risk cases.
Feedback creates smarter systems
Human corrections can become structured learning signals for continuous improvement.
Design Process Deep Dive
AI Architecture & Prototyping
We translated the taxonomy into a modular AI workflow.
- Multi-agent architecture
- Classification models
- Routing logic
- OCR integration
- Feedback mechanisms
Testing & Optimization
The system was refined against real-world conditions and operational feedback.
- Classification testing
- Edge-case validation
- Human review
- Model tuning
- Performance monitoring
Implementation & Handoff
AI Workflow Integration
Integrated the AI pipeline into CADEX’s existing email and support environment.
Queue Integration
Connected classifications and tags to the appropriate operational workflows.
Monitoring
Enabled visibility into accuracy, SLA performance, handling time, and email volumes.
Developer Handoff Process
Architecture
Defined agent responsibilities, workflow logic, classification hierarchy, and routing rules.
Integration
Connected the AI workflow with existing email systems, queues, and operational processes.
Testing
Validated classification, routing, attachments, edge cases, and human-review scenarios.
Monitoring
Tracked model performance and operational metrics to support continuous optimization.
Ready to Transform Your Conversion Rates?
See how we can apply similar strategies to your product. Let's discuss your conversion optimization challenges and explore proven solutions.
Want Similar Results for Your Product?
Our team of UX experts is ready to help you achieve measurable improvements in your conversion rates and user satisfaction. Let's discuss your specific challenges and goals.
Free Consultation
30-minute strategy session to discuss your goals
UX Audit
Comprehensive analysis of your current experience
Custom Strategy
Tailored approach based on your unique needs
No commitment required. Let's explore how we can help you achieve similar results.
