Automated Carrier Claims Management Using AI Agents

fiege · Tech · 2024 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Carrier claims management

FIEGE needed to automate carrier claims for lost parcels, including reading customer emails, pursuing carriers, checking contractual obligations, and closing claims without letting generative AI improvise business-critical reimbursements.

Constraints:

• claims can arrive by email or other channels and include text, PDFs, images, and other documents
• communications may occur in European languages
• the system must not hallucinate compensation or bank-account responses
• humans need comprehensible reasoning and handoff when the agent cannot resolve a case
• enterprise adoption required using the existing IT ecosystem and ticketing processes

Architecture

FIEGE’s carrier-claims AI worker runs an end-to-end deterministic workflow around sandboxed generative AI. The system reads incoming claims, classifies/extracts/assesses document content, writes templated emails to carriers, waits for missing information, reminds carriers, checks carrier responses against contractual obligations, and closes claims or asks humans for help. The implementation uses Microsoft Azure Logic Apps and Function Apps, existing ERP/ticketing systems, confidence thresholds, templates, and fragmented permissions so the model cannot freely generate sensitive or unsupported answers.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  claim-intake-channel["claim-intake-channel"] --> claims-workflow-orchestrator["claims-workflow-orchestrator"]
  azure-logic-function-apps["azure-logic-function-apps"] --> claims-workflow-orchestrator["claims-workflow-orchestrator"]
  claims-workflow-orchestrator["claims-workflow-orchestrator"] --> document-assessment-agent["document-assessment-agent"]
  claims-workflow-orchestrator["claims-workflow-orchestrator"] --> template-email-generator["template-email-generator"]
  template-email-generator["template-email-generator"] --> carrier-communication-channel["carrier-communication-channel"]
  carrier-communication-channel["carrier-communication-channel"] --> contract-obligation-checker["contract-obligation-checker"]
  contract-obligation-checker["contract-obligation-checker"] --> claims-workflow-orchestrator["claims-workflow-orchestrator"]
  claims-workflow-orchestrator["claims-workflow-orchestrator"] --> confidence-handoff-gate["confidence-handoff-gate"]
  confidence-handoff-gate["confidence-handoff-gate"] --> human-claims-team["human-claims-team"]
  claims-workflow-orchestrator["claims-workflow-orchestrator"] --> customer-erp-ticketing["customer-erp-ticketing"]
  human-claims-team["human-claims-team"] --> customer-erp-ticketing["customer-erp-ticketing"]

Components

Data flow

A claim arrives by email or another channel with text, PDFs, images, or other documents. The AI worker reads and analyzes the claim, identifies missing information, sends a templated request within 60 seconds when needed, keeps claim information warm, writes carrier emails with relevant information, evaluates carrier responses and contractual obligations, either continues the claim process or closes it, and informs the client. When confidence is too low or the workflow blocks, a ticket is handed to the human team.

Why this architecture

The core design choice is to sandbox generative AI inside deterministic workflow boundaries. The model handles extraction, classification, assessment, multilingual reasoning, and missing-information detection, but templates and policies control outbound responses. Confidence thresholds and existing tickets let humans take over when automation is unsafe, while ERP/email checks prevent PII leakage and make adoption easier for existing teams.

Models used: sandboxed generative AI · Microsoft Azure Logic Apps · Azure Function Apps · templates · confidence threshold

Scale

• FIEGE is described as having about 22,000 employees.
  “it has about 22,000 employees”transcript, stated
• FIEGE is described as having about two billion turnover.
  “about two billion turnover”transcript, stated
• The system sends the first missing-information reply within 60 seconds.
  “a reply to the first email is sent within 60 seconds”transcript, stated
• The system can process 70-90% of claims fully automatically.
  “such a system can process 70 to 90% of all clim climes fully automatically”transcript, stated
• The presenter reports eight-digit revenue/refund improvement.
  “led to an improve from eight digit um cost in like revenue and refund”transcript, stated

Alternatives considered (roads not taken)

Let the LLM generate free-form claim responses

Rejected because: Carrier claims and reimbursements require policy-bounded answers, so the system uses templates instead of improvisation.

Expect the agent to resolve every case alone

Rejected because: The model must know when it cannot safely process the case and hand off to humans.

Replace the company’s existing IT workflow

Rejected because: Adoption improves when the agent uses existing systems and human ticket queues.

Patterns

tool-use-orchestration (agent-architecture)

The agent coordinates reading documents, sending emails, checking obligations, reminders, and closure through workflow tools.

guardrails-io-filtering (safety)

Templates, sandboxing, and fragmented permissions prevent the model from improvising sensitive or unsupported answers.

confidence-thresholding (handoff)

A confidence threshold determines when a case should be handed to a human worker.

human-in-the-loop-gate (operations)

Blocked or low-confidence cases become tickets for the human team.

structured-output-contract (workflow)

Claims communication is generated from templates and policies rather than free-form model output.

Tradeoffs

Constrain generation with templates

Chose templated deterministic responses over free-form LLM writing.

Rationale: Templates sacrifice open-ended flexibility but prevent improvisation in claims and reimbursement communications.

Use existing enterprise systems

Chose ERP and ticket queues already used by the team over a heavy new IT migration.

Rationale: Keeping the existing ecosystem reduced adoption friction for human teams.

Gotchas (what bit them)

Free-form generation is unsafe for reimbursement workflows.

Root cause: A hallucinated or over-broad response could mishandle claims, compensation, or private information.

Resolution: Keep generative AI sandboxed and use deterministic workflows, templates, and policy boundaries.

The AI worker cannot be trusted with unrestricted PII access.

Root cause: A malicious or mismatched email could ask for bank account or address details.

Resolution: Fragment access and tie sensitive data to existing ERP/customer identity checks such as matching email address.

Evaluation

The system is evaluated operationally through automation rate, handoff behavior, first-response latency, revenue/refund recovery, and user-interface acceptance by the human team.

Metrics: 70-90% claims fully automated · first reply within 60 seconds · eight-digit revenue/refund improvement · confidence threshold handoff rate · human team acceptance

Outcomes

• FIEGE deployed an AI worker for carrier claims management
• claims can be processed from first email through resolution
• 70-90% of claims can be fully automated
• first missing-information responses can be sent within 60 seconds
• the presenter reports eight-digit revenue/refund improvement

The system is described for a logistics provider with about 22,000 employees and about two billion turnover, with 70-90% full automation for claims and an eight-digit refund/revenue impact.