How to build a Claims Processing Agent

Workflow Overview

1. User Input Collection

The workflow begins by collecting all necessary information from the user through a series of input nodes:

• Policy Number

• Full Name of Insured Person

• Email

• Phone Number

• Date and Time of Incident

• Location

• Description of Damage or Loss

• Device Type

• Make and Model (e.g., iPhone 15 Pro)

• Date of Purchase

Additionally, users can upload:

• Proof of Purchase (e.g., receipt, processed with OCR for text extraction)

• Supporting Documentation (e.g., photos, police reports, also processed with OCR)

2. Policy Verification (Snowflake Search)

• The Snowflake Search node (an LLM node) uses the provided policy number, customer name, and/or device serial to search your Snowflake database for the relevant policy.

• It prefers an exact match on policy number, but can also use name and device serial if needed.

• The LLM is instructed to return only the best-match policy record, or indicate if no match is found.

3. Ownership & Eligibility Verification

• The Verify Ownership and Purchase Date node (another LLM) takes the policy data and all user-provided evidence.

• It checks:

  • Is the policy active?

  • Did the incident occur within the coverage period?

  • Does the receipt show a purchase date before the incident?

  • Does the device info match the policy?

• If information is missing or contradictory, it flags that ownership cannot be verified.

4. AI Routing: Decision Logic

• The AI Routing node uses AI to categorize the claim outcome into one of three categories:

  1. Policy and ownership verified (claim can be approved)

  2. Policy verified, but ownership cannot be verified (likely needs more info)

  3. Policy not verified (claim cannot be approved)

5. Outcome Handling: Approval or Rejection

Depending on the AI Routing result, the workflow branches:

If Approved:

• The Accepted LLM node generates a personalized approval email for the policyholder, confirming the claim is approved and outlining next steps.

• The Send Email action node sends this approval email to the user’s email address.

If Rejected:

• The Rejected LLM node generates a personalized rejection email, including the reason for denial and instructions for next steps or appeal.

• The Send Email 1 action node sends this rejection email to the user’s email address.

6. Output

• The final output node displays the result to the user (typically the approval message, but can be customized).

Node Map (Simplified)

Special Features

• Document Uploads: Proof of purchase and supporting docs are processed with OCR for text extraction.

• Database Integration: Policy verification is done live against your Snowflake database.

• Automated Emailing: Approval or rejection emails are sent automatically using Gmail integration.

• AI Decisioning: All routing and eligibility checks are handled by AI, reducing manual review.

In summary:
This workflow automates the entire insurance claim intake, verification, and communication process—collecting all necessary info, verifying policy and ownership, making a decision, and notifying the user, all with minimal manual intervention.