Autonomous Insurance Claims Processing Agent

A lightweight FNOL (First Notice of Loss) claims agent for the Synapx assessment. It reads TXT or text-based PDF FNOL documents, extracts key fields, identifies missing mandatory data, classifies the claim, and returns a routing decision as JSON.

What It Does

• Extracts policy, incident, involved-party, asset, estimate, attachment, and claim-type fields.
• Detects missing mandatory fields.
• Routes claims using the assessment rules:
  • Missing mandatory fields -> Manual Review
  • Fraud indicators (fraud, inconsistent, staged) -> Investigation Flag
  • Injury claim type -> Specialist Queue
  • Estimated damage below 25,000 -> Fast-track
  • Otherwise -> Standard Review
• Explains the routing decision in plain English.

When more than one rule could match, the agent prioritizes missing mandatory data first, then investigation keywords, then injury, then fast-track.

Project Structure

.
├── README.md
├── pyproject.toml
├── samples/
│   ├── fnol_fast_track.txt
│   ├── fnol_injury.txt
│   ├── fnol_investigation.txt
│   └── fnol_manual_review.txt
├── src/fnol_agent/
│   ├── cli.py
│   ├── extractors.py
│   ├── models.py
│   ├── parser.py
│   ├── pipeline.py
│   └── router.py
└── tests/
    └── test_pipeline.py

Run Locally

Requires Python 3.9 or newer.

python3 -m venv .venv
source .venv/bin/activate
pip install -e .

Process one sample:

fnol-agent samples/fnol_fast_track.txt --pretty

Process multiple documents:

fnol-agent samples/*.txt --pretty

Run without installing:

PYTHONPATH=src python3 -m fnol_agent.cli samples/fnol_fast_track.txt --pretty

Example Output

{
  "extractedFields": {
    "policyNumber": "POL-100245",
    "policyholderName": "Aisha Khan",
    "effectiveDates": "2026-01-01 to 2026-12-31",
    "incidentDate": "2026-04-18",
    "incidentTime": "09:35",
    "incidentLocation": "Dubai Marina, Dubai",
    "description": "Rear bumper and parking sensor damaged in a low-speed collision.",
    "claimant": "Aisha Khan",
    "thirdParties": "Omar Ali, driver of vehicle DXB-77821",
    "contactDetails": "aisha.khan@example.com, +971501234567",
    "assetType": "Private vehicle",
    "assetId": "VIN-1HGBH41JXMN109186",
    "estimatedDamage": 8750.0,
    "claimType": "Motor",
    "attachments": ["police report", "damage photos"],
    "initialEstimate": 8750.0
  },
  "missingFields": [],
  "recommendedRoute": "Fast-track",
  "reasoning": "Estimated damage is below 25,000 and all mandatory fields are present.",
  "sourceDocument": "samples/fnol_fast_track.txt"
}

Tests

PYTHONPATH=src python3 -m unittest discover -s tests

Notes

• PDF support is intended for text-based PDFs. Scanned PDFs need OCR before processing.
• The extractor is intentionally deterministic and dependency-free so the project is easy to run and review.
• The parser is regex-based and works best with label/value FNOL documents, which matches the dummy assessment format.