EMA: The Universal AI Employee for Insurance Claims

EMA (Enterprise Multi-Agent Assistant) is an autonomous, agentic AI system designed to revolutionize the motor claims lifecycle. It moves beyond simple chatbots to act as a proactive "Universal AI Employee" that can execute complex workflows, analyze unstructured data, and make intelligent decisions.

---

📉 Industry Context: The "Claims Crisis"

Motor insurance is a high-frequency, moderate-severity business. The claims process is the core "moment of truth" for customers, yet it is bogged down by manual work.

The End-to-End Claims Value Chain

Currently, this process involves multiple handoffs and dependencies on external partners.

graph LR
    FNOL[First Notice of Loss] --> Triage[Triage & Assignment]
    Triage --> Investigate[Investigation & Evidence]
    Investigate --> Decision[Liability & Coverage]
    Decision --> Settlement[Settlement & Payout]
    
    style FNOL fill:#f9f,stroke:#333,stroke-width:2px
    style Settlement fill:#bbf,stroke:#333,stroke-width:2px

Loading

The Pain Points (MECE Problem Tree)

The core problem is "High Cost + Low NPS".

mindmap
  root((High Cost + Low NPS))
    Customer Experience
      Long cycle times
      Lack of transparency
      Repeated document requests
    Operational Efficiency
      Adjusters chasing info
      Data re-entry (Swivel Chair)
      Manual coordination
    Risk & Fraud
      Inconsistent adjudication
      Late fraud detection
      Claims leakage

Loading

---

🧠 Core Concept: Agentic AI as the Step-Change

Legacy automation (RPA) waits for humans to decide. Agentic AI reverses this. EMA is an AI employee that can:

1. Autonomously Reason: "This photo shows damage inconsistent with the description."
2. Plan: "I need to ask the claimant for a police report."
3. Act: Sends the email, updates the DB, and triggers the payment.

The Solution: EMA Claims Orchestrator

EMA acts as a layer of intelligence on top of your existing systems.

graph TD
    subgraph "Experience Layer"
        AdjusterConsole[Adjuster Console]
        ClaimantUI[Claimant Interface]
    end

    subgraph "AI Employee Layer (EMA)"
        Orchestrator{Claims Orchestrator}
        EvidenceAgent[Evidence Agent]
        VisionAgent[Vision Estimator]
        PolicyAgent[Policy Agent]
        FraudAgent[Fraud Signal Agent]
    end

    subgraph "Enterprise Integration Layer"
        CoreSystem[(Core Claims DB)]
        DMS[(Document Mgmt)]
        Payments[Payment Rails]
    end

    AdjusterConsole <--> Orchestrator
    Orchestrator <--> EvidenceAgent
    Orchestrator <--> VisionAgent
    Orchestrator <--> PolicyAgent
    Orchestrator <--> FraudAgent
    
    EvidenceAgent <--> DMS
    Orchestrator <--> CoreSystem
    Orchestrator --> Payments

Loading

---

🤖 Agent Pod Design

EMA isn't a single model; it's a coordinated pod of specialized agents.

Agent	Role	Capabilities
Intake Sync Agent	The Scribe	Retrieves claim/FNOL data, creates case files.
Evidence Agent	The Investigator	Interacts with claimants, chases documents, reads emails.
Vision Damage Agent	The Appraiser	Uses Computer Vision to estimate repair costs from photos.
Policy Agent	The Lawyer	Translates PDF policies into executable code/rules.
Fraud Signal Agent	The Detective	Performs anomaly detection and SIU flagging.

---

📱 Prototype Journey & Wireframes

The current prototype demonstrates the "Happy Path" for a Type A/B (Low/Medium Complexity) claim.

1. The Command Center (Dashboard)

• Visual: A unified workspace showing the Claims Inbox.
• Feature: "Risk Score" column pre-calculated by the Orchestrator.
• Action: Adjusters can sort by "Severity" to prioritize work.

2. Claim 360 & Evidence Graph

• Visual: A dynamic graph visualization.
• Concept: All entities (Claimant, Vehicle, Photos, Policy) are nodes. As agents work, they add edges (relationships) to this graph.
• UI: The Agent Activity Log on the right shows real-time updates ("Verified Coverage", "Flagged Risk").

3. Smart Document Analysis

• Visual: Split screen. PDF on the left, Extracted Data on the right.
• Tech: GPT-4o reads raw repair estimates.
• Outcome: Converts unstructured PDF text into structured Line Items (Parts, Labor, Tax) and validates against policy limits.

4. Settlement Recommendation

• Visual: A drawer/modal popping up from the bottom.
• Content: Recommended Payout Amount, Confidence Score, and "Approve" button.
• Governance: High-value payouts require human click; low-value can be auto-approved (future scope).

---

🛠️ Technology Stack

• Frontend: Next.js 14 (App Router), Tailwind CSS, React Flow (for graphs).
• Backend: FastAPI (Python), Pydantic.
• AI Orchestration: LangChain & LangGraph.
• LLM: Azure OpenAI Service (GPT-4o).
• Database: Supabase (PostgreSQL).

---

📦 Installation & Setup

1. Clone: git clone https://github.com/Kulraj69/ema.git
2. Backend:

   cd backend
   python -m venv venv
   source venv/bin/activate
   pip install -r requirements.txt
   uvicorn main:app --reload

3. Frontend:

   cd frontend
   npm install
   npm run dev

---

📄 License

MIT