Back to POC Lab
LiveGenAI Agent

CPC Comparison AI Agent

Multi-Agent Benefits Comparison System

Google ADK · Vertex AI Agent Engine · Health Insurance Operations

Google ADKVertex AI Agent EnginePythonHealth Insurance APIsExcel

THE CHALLENGE

Manual Benefits Administration is Broken

Benefits Administrators spend significant time manually cross-referencing health plan configurations — comparing deductibles, copays, and benefit categories across products to ensure accurate benefits coding.

01

Time-Intensive

Hours spent cross-referencing spreadsheets across multiple plan documents and benefit categories

02

Error-Prone

Manual data entry creates risks of miscoding benefits, leading to claims issues and compliance concerns

03

Spreadsheet-Driven

No structured tooling — comparisons exist only in ad hoc Excel files without version control or audit trails

THE SOLUTION

A Pipeline of Specialized Sub-Agents

1Orchestrator Agent — Routes intent, manages multi-turn session state
2CPC Intake Agent — Extracts plan codes & effective dates from natural language
3Documents Agent — Retrieves plan docs from health insurance APIs
4Comparison Agent — Structured diff + color-coded Excel export

Outcome: Natural language input: "Compare Plan A effective 2024-01-15 with Plan B" → Fully formatted Excel comparison report generated automatically — no manual steps.

ARCHITECTURE

System Architecture Diagram

AI Lab Prototype

Multi-Agent Benefits Comparison System

Google ADK · Vertex AI Agent Engine · Health Insurance Operations

Natural Language Input

👤
Benefits Administrator
“Compare Plan A effective 2024-01-15 with Plan B”

Orchestration Layer

🧠
Orchestrator AgentPrimary Agent

Interprets multi-turn conversations · Routes intent to sub-agents · Manages session state across turns · Coordinates pipeline execution

Intent ClassificationMulti-turn StateTool RoutingGoogle ADK

Specialized Sub-Agents

📋CPC Intake Agent

Extracts plan codes and effective dates from free-form natural language queries

Plan Code ExtractionDate ParsingNLP
📄Documents Agent

Retrieves Plan Highlights, Benefit Details, and Plan Summaries from health insurance APIs

API IntegrationPlan HighlightsBenefit Details
⚖️Comparison Agent

Structured diff across all benefit categories · Color-coded Excel report with 5 analytical sheets

Structured DiffExcel Export5 Sheets

Automated Output

📊
Color-Coded Excel Comparison Report
Generated automatically · No manual steps · 5 analytical sheets
✓ Deductibles✓ Copays✓ Benefit Categories
Benefit Category
Plan A
Plan B
Status
Annual Deductible (Ind)
$1,500
$2,000
DIFF
Primary Care Copay
$25
$25
MATCH
Specialist Copay
$45
$60
DIFF
Out-of-Pocket Max
$6,000
$7,500
DIFF
Emergency Room
$350
$350
MATCH
Google ADK
Vertex AI Agent Engine
Health Insurance APIs
Multi-turn Session Mgmt
⚡ Mock Mode🔴 Live API Mode

AGENT CAPABILITIES

What Each Agent Does

🧠

Orchestrator Agent

Primary Agent
  • Interprets multi-turn conversation context
  • Routes user intent to correct sub-agent
  • Manages state across conversation turns
  • Coordinates end-to-end pipeline execution
Google ADK · Session Management · Intent Routing
📋

CPC Intake Agent

Sub-Agent
  • Extracts plan codes from free-form natural language
  • Parses effective dates (e.g. 2024-01-15)
  • Validates extracted parameters before downstream use
  • Handles ambiguous or incomplete query inputs
NLP Extraction · Date Parsing · Parameter Validation
📄

Documents Agent

Sub-Agent
  • Retrieves Plan Highlights from insurance APIs
  • Fetches Benefit Details and Plan Summaries
  • Handles API authentication and error states
  • Stores structured JSON for downstream comparison
REST API Integration · JSON Storage · Error Handling
⚖️

Comparison Agent

Sub-Agent
  • Performs structured diff across all benefit categories
  • Identifies deductible, copay, and OOP differences
  • Exports color-coded Excel with 5 analytical sheets
  • Flags MATCH vs DIFF per benefit line item
Structured Diff · Excel Export · 5 Analytical Sheets

EVALUATION FRAMEWORK

Evaluation Was Harder Than Building

A single query can trigger: intent classification → tool selection → API calls → data transformation → file generation. Any break in that chain produces a wrong answer — traditional unit tests don't catch routing failures or hallucinated tool arguments.

01
RoutingEvaluator
  • Intent → agent sequence mapping
  • Correct agent routing per query
  • Routing failure detection
02
ApiDataEvaluator
  • JSON schema integrity check
  • Data structure validation
  • Saved response verification
03
ComparisonEvaluator
  • Real difference detection
  • False positive prevention
  • Benefit category coverage
04
ExcelEvaluator
  • Sheet structure validation
  • Plan code header alignment
  • Difference highlighting check
⚡ Mock Mode
Deterministic, no credentials · Fast inner-loop development
🔴 Live API Mode
Real data validation · Production signal

KEY LEARNINGS

What the AI Lab Taught Us

1

Agent Specialization Matters

Smaller, focused agents are easier to route, test, and iterate on independently. Avoid monolithic agents.

2

Build Evaluators Early

Multi-turn systems need custom evaluation logic before you can trust outputs. Retrofitting evals is painful.

3

Mock + Live Eval Modes

Essential separation: Mock mode for fast inner-loop dev; Live API mode for real-world validation signal.

4

Structured Output Contracts

Defining expected tool args, agent sequences, and output schemas upfront made debugging dramatically faster.

5

Google ADK + Vertex AI

A promising foundation for enterprise multi-agent deployment with built-in session management and tool routing.

NEXT STEPS

From Prototype to Production

Production Hardening

Error recovery, retry logic, input sanitization, and security review

Data Quality & Governance

Audit trails, output validation, and human-in-the-loop checkpoints

Workflow Integration

Connect to existing Benefits Admin tools and claims systems

Scale Evaluation

Expand eval coverage with more plan types and edge case queries

Note: This is an AI Lab prototype — meaningful work ahead before production deployment. Exploring multi-agent patterns in regulated industries requires careful evaluation design.