ClaimFlow AI: A Hybrid Multi-Agent Adjudication Platform in the Insurance Claims Field

ClaimFlow AI is a hybrid multi-agent claims adjudication platform for auto and health insurance. It combines LLM workflows with a deterministic rule engine, with the core concept of 'LLMs handle understanding, rule engines handle decision-making'. It aims to address the issues of low efficiency in traditional manual claims processing and security risks in pure LLM solutions, enabling explainable and auditable automated claims decisions.

Insurance claims review faces four core challenges: complex and diverse documents (medical records, accident reports, etc., in various formats), high decision-making risks (incorrect approvals leading to losses or complaints), strict compliance requirements (decisions need to be traceable and explainable), and high labor costs (U.S. doctors spend 40 hours per week on administrative tasks). ClaimFlow AI balances efficiency and security through a hybrid architecture, leveraging AI to understand unstructured data while retaining the determinism and auditability of the rule system.

Layered Architecture

• Frontend Layer: Next.js review interface (queue management, case details, workflow timeline, etc.)
• API Layer: FastAPI backend (CRUD, document management, workflow query, etc.)
• Core Engine Layer: State machine pattern manages the claims lifecycle (intake→extraction→validation→coverage_review→fraud_review→decision→communication)
• Agent Layer: Collaboration of multiple dedicated agents (Intake, Document Extraction, Contradiction, etc.)
• Decision Layer: Rule engine makes decisions by priority (hard rejection → suspension due to missing information → high risk referred to manual → normal approval)

Dual-Domain Support

• Auto Insurance: 7 typical scenarios (standard accident approval, invalid policy rejection, etc.)
• Health Insurance: 7 common cases (regular coverage approval, invalid member rejection, etc.) The dual-domain design enhances applicability and the possibility of cross-domain learning.

• Defensive Prompt Engineering: Agent prompts prevent injection; versioned management allows traceability and rollback
• Human Intervention Mechanism: Automated decisions can be overwritten; overwrites generate audit events; high-risk cases are forced to manual processing
• Complete Audit Trail: Records the full history of case processing (workflow steps, agent data, rule basis, manual operations)

The project has a built-in complete evaluation framework:

• Synthetic dataset: 14 seed scenarios (7 auto insurance + 7 health insurance)
• Decision quality evaluation: Automatic accuracy check
• Unit tests: Cover core components such as rule engines and decision strategies
• End-to-end integration tests: Verify the complete workflow The evaluation report outputs structured JSON, facilitating monitoring of model performance degradation.

Technical Details

• Project structure: apps/api (FastAPI backend), apps/web (Next.js frontend), data/sample_claims, etc.
• LLM integration: Supports environment variable configuration for providers (default Ollama local deployment); LLMs are only used in stages like information extraction and contradiction detection; final decisions do not rely on LLMs

Open-Source Significance

1. Architectural reference: Demonstrates safe use of LLMs in high-risk scenarios
2. Compliance template: Example of an auditable AI system implementation
3. Domain knowledge: 28 seed scenarios as industry benchmarks
4. Evaluation method: Reusable decision quality framework For insurance companies, it can serve as a starting point for their own systems; for researchers, it is a case of human-machine collaborative decision-making.

ClaimFlow AI balances efficiency and security in insurance claims through the 'LLM understanding + rule decision-making' hybrid architecture. Its designs such as multi-agent collaboration, deterministic decision gates, and complete audit trails provide important references for AI applications in sensitive industries like finance and healthcare. With iterations, it is expected to become an important part of the insurance technology open-source ecosystem.