GDP-Agent: AI-Enhanced PII Detection Architecture

Executive Summary

GDP-Agent is an intelligent email monitoring system that combines Microsoft Presidio for pattern-based PII detection with Azure AI Foundry for contextual analysis and false positive filtering. The system transforms raw technical findings into actionable business intelligence and professional communication.

Core AI-Enhanced Business Flow

graph TD
    A[📧 Email Sent] --> B[Email Monitoring]
    B --> C[Basic Email Processing]
    C --> D[🔍 Presidio PII Scan]
    D --> E{Raw PII Detected?}
    E -->|No| F[✅ No Action Needed]
    E -->|Yes| G[🧠 AI Analysis & Filtering]
    G --> H[🎯 False Positive Review]
    H --> I{Real PII Confirmed?}
    I -->|No| J[✅ Filtered Out - No Alert]
    I -->|Yes| K[📊 Risk Assessment]
    K --> L[🤖 AI Communication Generation]
    L --> M[📧 Professional Follow-up Email]
    M --> N[📋 Business Action Items]
    
    style A fill:#e1f5fe
    style G fill:#f3e5f5
    style H fill:#fff3e0
    style L fill:#e8f5e8
    style M fill:#fce4ec

AI-Powered PII Detection Pipeline

1. Dual-Layer AI Detection → Enhancement

sequenceDiagram
    participant Email as 📧 Email Content
    participant PresidioCore as 🔍 Presidio Core Patterns
    participant PresidioAI as 🤖 Presidio AI Extension
    participant AIAnalysis as 🧠 Azure AI Analysis
    participant Filter as 🎯 False Positive Filter
    participant Business as 📊 Business Logic
    
    Email->>PresidioCore: Standard pattern detection
    Email->>PresidioAI: AI-enhanced entity detection
    PresidioCore->>AIAnalysis: Pattern results (8 entities)
    PresidioAI->>AIAnalysis: AI context results (5 entities)
    AIAnalysis->>Filter: Combined results (10 unique entities)
    Filter->>AIAnalysis: "7 are false positives"
    AIAnalysis->>Business: Confirmed: 3 real PII entities
    Business->>AIAnalysis: Generate business-friendly report
    AIAnalysis-->>Email: Professional notification sent

2. Extended Presidio with AI Enhancement

graph TB
    subgraph "Extended Presidio Engine"
        A[📧 Email Content] --> B[Core Presidio Analyzers]
        A --> C[🤖 AI Enhanced Recognizer]
        
        B --> D[Pattern-Based Detection]
        D --> E[SSN: 123-45-6789<br/>Phone: 555-0123<br/>Email: contact@company.com]
        
        C --> F[AI Context Detection]
        F --> G[Unusual Name Formats<br/>Business Context Entities<br/>Contextual PII]
        
        E --> H[Combined Results]
        G --> H
    end
    
    subgraph "AI Analysis Layer"
        H --> I[🧠 Azure AI Foundry Analysis]
        I --> J[Business Context Filter]
        I --> K[False Positive Detection]
        I --> L[Risk Assessment]
    end
    
    subgraph "Business Intelligence"
        L --> M[📊 Final PII Assessment]
        M --> N[Professional Communication]
        M --> O[Compliance Guidance]
    end
    
    style B fill:#ffebee
    style C fill:#e8f5e8
    style I fill:#f3e5f5
    style M fill:#fff3e0

Extended Presidio Architecture

AI-Enhanced Recognizer Integration

GDP-Agent extends the standard Presidio framework by injecting a custom AI Enhanced Recognizer directly into Presidio's analyzer engine. This creates a hybrid detection system that combines pattern-based precision with AI contextual understanding.

graph TB
    subgraph "Standard Presidio Framework"
        A[AnalyzerEngine] --> B[Built-in Recognizers]
        B --> C[Regex Patterns]
        B --> D[ML Models]
        B --> E[Spacy NLP]
    end
    
    subgraph "GDP-Agent Extension"
        F[AIEnhancedRecognizer] --> G[Azure AI Foundry]
        G --> H[Prompty Template]
        H --> I[Context Analysis]
        I --> J[Unusual Format Detection]
    end
    
    A --> F
    F --> A
    
    subgraph "Detection Results"
        K[Core Presidio Results] --> L[Combined Entity List]
        M[AI Enhanced Results] --> L
    end
    
    C --> K
    D --> K
    E --> K
    J --> M
    
    style A fill:#ffebee
    style F fill:#e8f5e8
    style G fill:#f3e5f5
    style L fill:#fff3e0

Key Features of AI Extension

1. Seamless Integration: Plugs directly into Presidio's AnalyzerEngine as a custom recognizer
2. Contextual Detection: Identifies PII that pattern-based methods miss
3. Confidence Calibration: AI results use lower confidence scores to let built-in recognizers take precedence
4. Format Flexibility: Detects names, addresses, and identifiers in unusual formats
5. Business Context: Understands when entities are business-relevant vs. personal

Detection Enhancement Examples

Scenario	Core Presidio	AI Enhancement	Combined Result
"John Smith from ABC Corp"	Detects "John Smith" as PERSON	Detects "ABC Corp" as ORGANIZATION	Both entities with appropriate context
"Call me at five-five-five-zero-one-two-three"	Misses (unusual format)	Detects as PHONE_NUMBER	Phone number identified
"CEO Jane Doe"	Detects "Jane Doe" as PERSON	Adds business context (title)	Enhanced person entity with role
"john.smith.ceo@company.co.uk"	Detects email pattern	Identifies professional context	Email with business relevance

AI Template System Architecture

The system uses three specialized Prompty templates:

1. Presidio Analysis Template (presidio_analysis.prompty)

Purpose: Transform raw Presidio detections into business-intelligent insights

Key Functions:

• Filter false positives (company names detected as persons)
• Apply business context to PII findings
• Determine actual risk levels based on data sensitivity
• Generate actionable recommendations

graph LR
    A[Raw Presidio Data] --> B[presidio_analysis.prompty]
    B --> C[Filtered Results]
    B --> D[Risk Assessment]
    B --> E[Business Context]
    
    style B fill:#f3e5f5

2. Follow-up Report Template (follow_up_report.prompty)

Purpose: Generate professional business communications

Key Functions:

• Translate technical findings to business language
• Create anonymized examples for user education
• Provide practical impact assessments
• Generate actionable next steps

graph LR
    A[Technical Findings] --> B[follow_up_report.prompty]
    B --> C[Business Summary]
    B --> D[Impact Assessment]
    B --> E[Action Items]
    
    style B fill:#e8f5e8

3. Entity Detection Template (pii_entity_detection.prompty)

Purpose: Enhance pattern-based detection with AI context understanding

Key Functions:

• Identify PII in unusual formats
• Understand contextual relevance
• Detect entities missed by pattern matching
• Provide confidence scoring

Technology Integration Points

graph TD
    subgraph "External Services"
        A[Microsoft Graph API]
        B[Azure AI Foundry]
        C[Microsoft 365]
    end
    
    subgraph "AI Components"
        D[Presidio Engine]
        E[Prompty Templates]
        F[GDPAgentManager]
    end
    
    subgraph "Business Logic"
        G[Enhanced Email Processor]
        H[Notification Service]
        I[Report Generator]
    end
    
    C --> A
    A --> G
    G --> D
    D --> F
    F --> B
    E --> F
    F --> I
    I --> H
    H --> A
    
    style B fill:#e3f2fd
    style D fill:#f3e5f5
    style F fill:#fff3e0

Business Value Proposition

Before AI Enhancement (Standard Presidio)

• Limited pattern detection → Misses contextual PII
• High false positive rate → Company names flagged as people
• Format restrictions → Misses "five-five-five" phone numbers
• Technical jargon → Users confused about actions
• Generic recommendations → Poor adoption

After Dual-Layer AI Enhancement

• Extended detection coverage → AI finds contextual and unusual format PII
• Intelligent false positive filtering → 70% reduction in noise
• Context-aware analysis → Business vs. personal data distinction
• Professional communication → User-friendly explanations
• Actionable guidance → Specific privacy improvement steps

Detection Improvement Examples

• Coverage: Standard Presidio finds 8 entities → AI extension adds 5 more → Combined system identifies 10 unique entities
• Accuracy: Raw 10 detections → AI filters 6 false positives → 4 confirmed real PII
• Context: "John Smith, ABC Corp CEO" → System understands John Smith is a person, ABC Corp is a company, CEO is a business title

Key AI Capabilities

Enhanced Detection Layer

1. Extended Pattern Recognition: AI recognizer finds PII in unusual formats that regex patterns miss
2. Contextual Entity Detection: Identifies business entities, professional titles, and industry-specific terms
3. Format Flexibility: Detects phone numbers written as "five-five-five-one-two-three-four"
4. Confidence Calibration: AI results complement rather than compete with core Presidio patterns

Analysis & Filtering Layer

5. Contextual Understanding: Distinguishes between "John Smith" as a customer vs. email signature
6. Business Relevance: Identifies when phone numbers are company contacts vs. personal data
7. False Positive Intelligence: Filters out company names incorrectly detected as person names
8. Risk Calibration: Adjusts severity based on data type and business context

Communication Layer

9. Professional Communication: Converts technical findings into user-friendly explanations
10. Educational Approach: Provides learning-focused guidance rather than punitive warnings
11. Compliance Intelligence: Delivers relevant regulatory guidance based on detected PII types
12. Actionable Recommendations: Offers specific steps users can take to improve privacy practices

Compliance & Privacy Features

• Data Anonymization: All PII examples in communications are masked
• Audit Trail: Complete logging of AI decisions and filtering logic
• Risk Categorization: GDPR/CCPA-aligned risk level assessment
• Professional Notifications: Business-appropriate communication tone
• False Positive Learning: AI continuously improves context understanding

This architecture enables proactive privacy protection while minimizing user disruption through intelligent filtering and professional communication.