AI-Powered Policy Agent

Agentic AI transforms policy generation-delivering Nova sonic voice-driven, explainable & complaint decisions for insurance docs, powered by AWS Bedrock AgentCore, Strands, Kiro IDE & Nova models

Comment

AI-Powered Policy Agent

🩺 Inspiration

The healthcare insurance industry faces a critical challenge: complex policy documents that are difficult for customers to understand, leading to poor decision-making and increased claim disputes. Traditional insurance advisors are expensive and not available 24/7.

We envisioned an intelligent, conversational AI agent that could democratize access to insurance expertise, making policy selection as simple as talking to a trusted friend.

Our inspiration came from observing how people struggle with insurance jargon, premium calculations, and policy comparisons. We asked ourselves:

"What if we could create an AI agent that not only reads and understands complex insurance documents but also explains them in natural language, provides personalized recommendations, and ensures regulatory compliance—all through voice interaction?"

This vision led us to build Alan, a sophisticated AI-powered health insurance advisor that combines AWS Bedrock’s Amazon Nova models with advanced agentic workflows to deliver intelligent, explainable, and compliant insurance guidance.

💡 What It Does

This project is an enterprise-grade, multi-agent AI system that revolutionizes health insurance policy management and customer advisory services.

Conversational Policy Advisory

Users interact with Alan through natural voice conversations powered by Amazon Nova Sonic, discussing their insurance needs, asking questions about policy details, and receiving personalized recommendations in real-time.

Intelligent Document Processing

The system automatically ingests, validates, and analyses complex health insurance policy documents using a team of eight specialized AI agents, each focusing on specific aspects like medical risk assessment, financial analysis, compliance verification, and lifestyle factors.

Explainable AI Decision-Making

Every recommendation comes with transparent reasoning, showing users exactly how premiums were calculated, why specific coverage was recommended, and what factors influenced the decision—building trust through transparency.

Automated Policy Generation

Upon successful underwriting analysis, the system automatically generates professionally formatted PDF policy documents with all necessary legal clauses, coverage details, and personalized terms.

Regulatory Compliance

Built-in compliance agents ensure all processing adheres to HIPAA regulations, state insurance laws, and company policies, with automatic PHI masking and audit trail generation.

Real-Time Status Tracking

A sophisticated S3-based status tracking system provides live updates on document processing, allowing frontend applications to display progress as each agent completes its analysis.

⚙️ How We Built It

Architecture Foundation

We architected SDIS as a hybrid cloud-native application leveraging AWS services extensively.

  • Frontend: Flask with SocketIO for real-time communication
  • Backend: AWS Bedrock AgentCore for multi-agent processing

Multi-Agent Orchestration

We designed eight specialized Nova Pro agents, each with carefully crafted system prompts encoding domain expertise:

  1. Data Intake Agent: Handles document ingestion and initial categorization
  2. Document Verification Agent: Validates authenticity and completeness
  3. Medical Risk Assessment Agent: Evaluates health-related mortality risks
  4. Financial Analysis Agent: Assesses coverage appropriateness and affordability
  5. Driving Record Agent: Analyzes transportation-related risk factors
  6. Compliance Verification Agent: Ensures regulatory adherence
  7. Lifestyle Behavioural Analysis Agent: Evaluates behavioural risk patterns
  8. Summary Generation Agent: Produces comprehensive HTML reports

The orchestrator agent coordinates this workflow, ensuring all agents execute sequentially and their insights are properly synthesized.

Additional Components

  • Voice Integration: Integrated Amazon Nova Sonic for natural voice interactions with real-time audio streaming, speech-to-text conversion, and contextual conversation tracking.
  • Document Processing Pipeline: Built using PyPDF2 with robust error handling and recovery mechanisms.
  • Session Management: Session data and context stored in S3 using unique timestamped identifiers for persistent tracking.
  • Policy Generation Engine: Generates professional PDFs with ReportLab from structured data extracted by Nova Pro.
  • Real-Time Updates: S3-based JSON status tracking, continuously polled by the Flask frontend.
  • Deployment Architecture: Entire backend deployed on AWS Bedrock AgentCore with serverless scalability and built-in security.

🧩 Challenges We Ran Into

  • Credential Management Complexity: Handling AWS STS tokens and automatic refresh mechanisms.
  • Agent Coordination and Error Recovery: Sequential reliability and error tolerance across agents.
  • Nova Sonic Audio Streaming: Managing real-time bidirectional audio streams with WebSockets.
  • Encoding Issues in PDF Extraction: Handling non-standard fonts and corrupted files.
  • Context Preservation Across Agents: Balancing token limits with contextual accuracy.
  • S3 Status Synchronization: Managing consistency, race conditions, and polling frequency.
  • Policy Generation Accuracy: Reliable extraction of structured data from unstructured text.
  • Voice Conversation Flow Control: Preventing premature AI responses through state management.

🏆 Accomplishments That We're Proud Of

  • Production-Ready Multi-Agent System: Fully functional end-to-end underwriting automation.
  • Seamless Voice Experience: Natural, dynamic insurance conversations powered by Nova Sonic.
  • True Explainable AI: Transparent reasoning for every decision.
  • Automated Compliance: HIPAA adherence with PHI masking and audit trails.
  • Professional Policy Documents: Insurance-grade PDF generation with legal formatting.
  • Hybrid Architecture Excellence: Distributed system integration via S3.
  • Robust Error Handling: Graceful failure recovery with informative error messages.
  • Real-Time Status Visibility: Transparent processing pipeline for users.

📚 What We Learned

  • Agent Design Patterns: Clear, narrowly scoped responsibilities improve consistency.
  • Prompt Engineering Mastery: Precision in instructions yields structured and reliable outputs.
  • AWS Credential Management: Importance of token lifecycle validation and fallback strategies.
  • Conversational AI Nuances: Explicit user intent detection outperforms assumption-based triggers.
  • Document Processing Resilience: Edge case handling is essential for production reliability.
  • S3 as Coordination Layer: Effective state management using structured JSON updates.
  • Importance of Observability: Comprehensive logs for debugging and monitoring.
  • Model Selection Matters: Optimal pairing of Nova Pro (reasoning) and Nova Sonic (voice).

🚀 What's Next for AI-Powered Policy Agent

  • Enhanced Personalization: Memory of past interactions for tailored recommendations.
  • Multi-Language Support: Expansion into Spanish, Mandarin, and more.
  • Integration with Insurance Carriers: Real-time quotes, comparisons, and instant enrollment.
  • Advanced Risk Prediction: Predictive models for future coverage needs.
  • Mobile Application: Native iOS and Android apps with offline and voice features.
  • Claims Assistance: Guided claims filing and tracking.
  • Telemedicine Integration: Connect with telehealth platforms for underwriting data.
  • Continuous Learning Pipeline: Feedback loops for model improvement via RAG updates.
  • Blockchain-Based Policy Records: Immutable, transparent policy recordkeeping.

Alan represents a new era of agentic AI in insurance, combining intelligence, compliance, and empathy to empower users to make confident healthcare decisions.

Built With

Share this project:

Updates