Skip to content

System Design: Insurance Claims Processor

Domain: InsurTech · Pattern: Orchestrator + Specialist Sub-agents + HITL

← Back to Agentic AI Concepts

Interview Problem Statement

"Design an agentic system that helps an insurance company process claims against their homeowner's policy more quickly."

Clarifying Questions

The problem is intentionally vague. These questions must be answered before committing to a design:

Question Why It Matters
What evidence does the claimant submit? (photos, videos, police reports, written descriptions?) Determines whether we need vision, audio transcription, or document parsing capabilities
Which legacy systems must the agent interface with? (Claims Management System, CRM, policy DB?) Defines the integration surface — tools the agents need to call
Has an insurance assessor already visited the site? If yes, we have a structured assessment report; if no, the agent must infer damage from raw evidence alone
What is the claim value threshold for auto-approval vs. mandatory human review? Sets the HITL escalation policy
What data is available on past claims? (for fraud detection and cost benchmarking) Determines viability of RAG-based retrieval vs. fine-tuning

System Architecture Overview

┌──────────────────────────────────────────────────────────────────────┐
│                        CLAIMANT INTERFACE                             │
│         Web Portal / Mobile App / Phone (IVR + Transcription)        │
│                                                                       │
│   Submits: photos · videos · police report · written description     │
└──────────────────────────┬───────────────────────────────────────────┘
                           │
                           ▼
┌──────────────────────────────────────────────────────────────────────┐
│                    SECURITY & INGESTION LAYER                         │
│                                                                       │
│   • PII Redaction: strip names, addresses, policy numbers from       │
│     evidence before passing to LLMs                                  │
│   • Prompt Injection Guard: sanitize free-text fields                │
│   • Fraudulent Imagery Detection: EXIF metadata check,              │
│     reverse image search for stock/recycled photos                   │
│   • Evidence stored in secure object store (encrypted at rest)       │
└──────────────────────────┬───────────────────────────────────────────┘
                           │  Sanitised claim package
                           ▼
┌──────────────────────────────────────────────────────────────────────┐
│                       ORCHESTRATOR AGENT                              │
│                                                                       │
│   Model: Multi-modal Foundation Model (e.g., Gemini / Claude)        │
│   Pattern: Plan → Dispatch → Reflect → Synthesize                    │
│                                                                       │
│   • Creates claim context window (shared state for all sub-agents)   │
│   • Dispatches to specialist sub-agents in parallel                  │
│   • Detects conflicts between sub-agent findings                     │
│   • Applies escalation rules → routes to HITL or auto-decision       │
│   • Interfaces with legacy CMS and CRM via tool calls                │
└──────┬─────────────┬──────────────────────┬───────────────────┬──────┘
       │             │                      │                   │
       ▼             ▼                      ▼                   ▼
┌──────────┐  ┌──────────────┐   ┌─────────────────┐  ┌─────────────────┐
│  DAMAGE  │  │    POLICY    │   │  COST ESTIMATOR │  │  LEGACY SYSTEM  │
│  ANALYST │  │   AUDITOR    │   │                 │  │   CONNECTOR     │
│          │  │              │   │ • RAG over      │  │                 │
│ • Vision │  │ • Coverage   │   │   repair cost   │  │ • CMS + CRM     │
│   model  │  │   limits +   │   │   benchmarks    │  │   read/write    │
│ • EXIF   │  │   exclusions │   │ • ACV / RCV     │  │                 │
│   check  │  │ • Flags gaps │   │   per policy    │  │                 │
└─────┬────┘  └───────┬──────┘   └────────┬────────┘  └────────┬────────┘
      │               │                   │                    │
      └───────────────┴──────────────┬────┘                    │
                                     ▼                         │
┌────────────────────────────────────────────────────────────────────────┐
│                          SYNTHESIZER AGENT                              │
│                                                                         │
│   Combines: damage + policy coverage + cost estimate + claimant history│
│                                                                         │
│   RECOMMENDATION  │  CONDITIONS                                         │
│   ✅ Approve      │  Damage confirmed · Covered · Fraud risk: LOW       │
│   ℹ️ Request Info │  Evidence incomplete · Policy gap unclear           │
│   🔍 Dispatch     │  High severity · Fraud risk: MED/HIGH · Conflict    │
│      Human        │                                                     │
└──────────────────────────────┬──────────────────────────────────────────┘
                               │
           ┌───────────────────┼───────────────────┐
           ▼                   ▼                   ▼
    ✅ Auto-Approve       ℹ️ Request Info    🔍 Escalate to
    (HITL light-touch)   Claimant notified   Human Adjuster
    Adjuster 48hr        via portal          Full package
    override window                              │
           └───────────────────┴───────────────────┘
                               │
                               ▼
              ┌────────────────────────────────┐
              │        NOTIFICATION AGENT       │
              │  • Updates CMS claim status    │
              │  • Emails / SMS claimant       │
              │  • Triggers payment if approved│
              │  • Refers approved contractors │
              └────────────────────────────────┘

Agent Breakdown

1. Orchestrator Agent

Model: Multi-modal foundation model (Gemini 1.5 Pro or Claude) — single model maintains coherent reasoning across text and image evidence without context loss between handoffs.

Responsibility Details
Context management Creates and maintains the shared claim context window across all sub-agents
Parallel dispatch Runs Damage Analyst + Policy Auditor + Legacy Connector simultaneously
Conflict detection Compares sub-agent outputs for discrepancies before passing to Synthesizer
Escalation logic Routes to auto-approve, request-info, or human adjuster paths
Legacy integration Tool calls to CMS (update claim status) and CRM (read claimant history)

2. Damage Analyst

Model: Multi-modal foundation model · Tools: Vision analysis, weather event API, EXIF metadata extractor

  • Classifies damage type: water / fire / impact / theft / wind
  • Classifies damage severity: minor / moderate / major / total loss
  • Corroborates date and cause against weather records
  • Flags if imagery metadata (EXIF timestamps, GPS) conflicts with the claim narrative

3. Policy Auditor

Model: LLM with RAG over policy documents · Tools: Policy DB API, coverage rules engine

  • Pulls the claimant's active policy at the exact date of loss
  • Checks perils covered, deductibles, sub-limits, exclusions
  • Returns a structured policy summary — does not make a coverage decision

4. Cost Estimator

Tools: RAG over repair cost database (Xactimate / regional benchmarks), depreciation tables

  • Retrieves comparable repair costs from RAG — avoids hallucinated cost estimates
  • Applies ACV (Actual Cash Value) or RCV (Replacement Cost Value) per policy terms
  • Drafts a preliminary settlement offer amount for the human adjuster

5. Synthesizer Agent

Model: LLM reasoning over structured inputs from all other agents

Outputs one of three structured recommendations with a fraud risk score (Low / Medium / High):

Recommendation Trigger Conditions
Approve Claim Damage confirmed · Covered · Cost reasonable · Fraud risk: Low
Request More Information Evidence incomplete · Policy gap ambiguous
Dispatch Human Assessor High severity · Fraud risk Med/High · Evidence-claim conflict

Conflict Resolution Logic: - Photo shows minor damage but claimant text describes major → flag, downgrade, escalate - EXIF date on photos ≠ claimed date of loss → raise fraud flag - Cost estimate significantly exceeds regional benchmark → soft flag with justification required

6. Legacy System Connector

Tools: CMS API, CRM API — reads claimant history, writes claim status at each lifecycle transition.

7. Notification Agent

Tools: Email, SMS, customer portal, contractor referral DB — delivers decision letters, triggers payment, provides approved contractors.

Model Selection

Decision Choice Rationale
NLP + Vision Single multi-modal model Avoids context loss between separate silos
Build vs. fine-tune General-purpose model + RAG RAG over policy and cost DB keeps estimates grounded
RAG for cost data RAG over Xactimate / regional benchmarks Prevents hallucinated repair costs

Safety & Human-Centric Design

HITL: The agent is decision support, not a decision maker. The final approval rests with a human adjuster: - Auto-approve path: Adjuster receives summary + 48-hour override window - Escalation path: Adjuster receives full package — damage report, policy summary, cost estimate, fraud score, draft settlement offer

Security:

Threat Mitigation
PII exposure to LLMs PII redacted in Security Layer before any evidence reaches a model
Prompt injection via free-text Input sanitisation; system prompt isolation
Fraudulent imagery EXIF metadata check + reverse image search at ingestion
Adversarial image manipulation Perceptual hash comparison against known fraud imagery DB

Data Flow

Claimant submits evidence
        │
        ▼
Security Layer: PII redaction · prompt injection guard · image auth check
        │
        ▼
Orchestrator creates claim context window
        │
        ├──► [PARALLEL] Damage Analyst ──────► damage type + severity + flags
        ├──► [PARALLEL] Policy Auditor ──────► coverage summary + exclusions
        └──► [PARALLEL] Legacy Connector ───► claimant history from CRM
        │
        ▼ (all three complete)
Cost Estimator ──► itemised repair cost + draft settlement offer
        │
        ▼
Synthesizer ──► conflict detection ──► recommendation + fraud risk score
        │
   ┌────┴──────────┬──────────────────┐
   │               │                  │
Approve        Request Info      Dispatch Assessor
(HITL light)   (notify claimant) (full package to adjuster)
   │               │                  │
   └───────────────┴──────────────────┘
                   │
        Notification Agent ──► claimant + CMS update + payment trigger

Claim State Machine

SUBMITTED ──► EVIDENCE_REVIEW ──► PENDING_INFO ──► EVIDENCE_REVIEW
                    │
                    ▼
             SYNTHESIZING
                    │
        ┌───────────┼────────────────┐
        │           │                │
   AUTO_APPROVED  AWAITING_INFO  AWAITING_ASSESSOR
        │                            │
        └────────────┬───────────────┘
                     │
               ADJUSTER_REVIEW
                     │
              ┌──────┴──────┐
           APPROVED       DENIED
              └──────┬────────┘
                   CLOSED

Diagrams

End-to-End Agent Orchestration Flow

flowchart TD
    A[Claimant\nWeb · Mobile · Phone] -->|photos · videos · reports · description| B

    B[Security & Ingestion Layer\nPII Redaction · Prompt Injection Guard\nFraudulent Imagery Detection]

    B -->|Sanitised claim package| C[Orchestrator Agent\nMulti-modal Foundation Model\nGemini / Claude]

    C --> D[Damage Analyst\nVision model · Weather API · EXIF check\nClassifies type + severity]
    C --> E[Policy Auditor\nPolicy DB · Rules Engine\nCoverage · Deductibles · Exclusions]
    C --> F[Legacy System Connector\nCMS · CRM\nClaimant history + claim status]

    D -->|damage type · severity · flags| G[Cost Estimator\nRAG over repair cost DB\nXactimate · ACV / RCV]
    E -->|policy summary| G
    F -->|claimant history| G

    G -->|itemised estimate + draft offer| H[Synthesizer Agent\nConflict Detection · Fraud Risk Score\nStructured Recommendation]

    H -->|Approve · Low fraud · Clear coverage| I[✅ Auto-Approve\nAdjuster 48hr override window]
    H -->|Evidence incomplete| J[ℹ️ Request More Information\nClaimant notified via portal]
    H -->|Conflict detected · Fraud MED-HIGH · Complex| K[🔍 Dispatch Human Assessor\nFull package to adjuster]

    I --> L[Notification Agent\nEmail · SMS · CMS update · Payment trigger]
    J --> L
    K -->|Adjuster decision| L

Conflict Resolution Logic

flowchart TD
    A[Synthesizer receives\nDamage Report + Policy Summary\n+ Cost Estimate + Claimant History]

    A --> B{Evidence vs.\nClaim consistent?}

    B -->|Photo shows MINOR\nText claims MAJOR| C[🚩 Flag discrepancy\nDowngrade recommendation\nEscalate to adjuster]
    B -->|EXIF date ≠\nclaimed date of loss| D[🚩 Raise fraud flag\nBlock auto-approve\nRequire adjuster sign-off]
    B -->|Cost estimate >>\nregional benchmark| E[⚠️ Soft flag\nSynthesizer must justify\nin recommendation]
    B -->|Policy exclusion\npartially applies| F[⚠️ Ambiguity flag\nEscalate — do not decide]
    B -->|All consistent| G[Proceed to recommendation\nwith fraud risk score]

    C --> H[DISPATCH HUMAN ASSESSOR]
    D --> H
    E --> I[APPROVE with justification\nor ESCALATE]
    F --> H
    G --> J{Fraud Risk Score?}
    J -->|Low| K[APPROVE CLAIM]
    J -->|Medium| L[REQUEST MORE INFO\nor ESCALATE]
    J -->|High| H

Claim State Machine

stateDiagram-v2
    [*] --> SUBMITTED : Claimant files claim
    SUBMITTED --> EVIDENCE_REVIEW : Security layer clears · Orchestrator dispatches agents
    EVIDENCE_REVIEW --> PENDING_INFO : Evidence incomplete · info requested
    PENDING_INFO --> EVIDENCE_REVIEW : Claimant provides missing evidence
    EVIDENCE_REVIEW --> SYNTHESIZING : All agents complete
    SYNTHESIZING --> AUTO_APPROVED : Low fraud · Clear coverage · Consistent evidence
    SYNTHESIZING --> AWAITING_INFO : Policy gap or ambiguity
    SYNTHESIZING --> AWAITING_ASSESSOR : Conflict detected · Fraud MED-HIGH · High value
    AWAITING_INFO --> EVIDENCE_REVIEW : Claimant responds
    AWAITING_ASSESSOR --> ADJUSTER_REVIEW : Human assessor assigned
    AUTO_APPROVED --> ADJUSTER_REVIEW : Adjuster exercises 48hr override
    ADJUSTER_REVIEW --> APPROVED : Adjuster approves
    ADJUSTER_REVIEW --> DENIED : Adjuster denies
    AUTO_APPROVED --> CLOSED : Override window passes · Payment processed
    APPROVED --> CLOSED : Payment processed
    DENIED --> CLOSED : Denial letter sent

Parallel Agent Execution Timeline

gantt
    title Claim Processing Timeline — Target under 10 min for auto-approve
    dateFormat  mm:ss
    axisFormat  %M:%S

    section Ingestion
    Security Layer — PII redaction · image auth   :a0, 00:00, 00:30

    section Parallel Analysis
    Damage Analyst — vision + weather + EXIF       :a1, after a0, 02:00
    Policy Auditor — coverage + exclusions         :a2, after a0, 00:45
    Legacy Connector — CMS + CRM lookup            :a3, after a0, 00:30

    section Cost & Synthesis
    Cost Estimator — RAG retrieval + ACV/RCV       :a4, after a1, 00:45
    Synthesizer — conflict check + recommendation  :a5, after a4, 00:30

    section Decision & Notify
    Orchestrator routing                           :a6, after a5, 00:10
    Notification Agent                             :a7, after a6, 00:20

← Back to Agentic AI Concepts · All System Designs