🏛 AXA☁️ Azure☁️ AWSRAG · Multi-agent · MCP · HIL
Case Study
Enterprise AI at AXA
Two complementary AI systems: one to centralise knowledge, one to automate claims, built on Azure and AWS.
2
AI projects
Delivered end-to-end
Azure
Knowledge Base
AI Search · Foundry · HIL
AWS
Claims automation
Bedrock · A2A · MCP
Multi
Agent coordination
Specialized agent pipelines
Project 1: Knowledge Base
Unified knowledge,
across every entity.
AXA needed a way for employees across multiple entities to ask natural language questions and get accurate, sourced answers, without leaving their workflow. We built a multi-tenant RAG system on Azure where each entity controls its own document library while employees query the entire group.
Knowledge siloed by entity
Each subsidiary managed documents in isolation. Employees couldn't find answers across the group.
Manual, time-consuming search
Finding the right policy or procedure required contacting several teams, a slow and error-prone process.
No quality control
When answers did surface, there was no way to validate their accuracy or flag outdated content.
How it works
01
Multi-entity document management
Each entity uploads and curates its own documents through a dedicated admin interface. Access is scoped per entity.
02
Azure AI Search: hybrid retrieval
Vector + keyword search across the combined corpus. Finds semantically relevant passages even with imprecise queries.
03
Azure AI Foundry: answer generation
Grounded response generation using retrieved context. The LLM cites its sources and stays within the knowledge base.
04
Automated evaluation pipeline
Each response is scored on relevance, groundedness and coherence. Low-quality answers are flagged before they reach users.
05
Human-in-the-Loop (HIL)
Domain experts review flagged answers and provide corrections. Feedback is fed back to continuously improve retrieval and generation.
Project 2: Claims & Service Automation
Automated claims,
from intake to resolution.
Claims processing and customer service interactions were largely manual: high volume, repetitive, and slow. We built a multi-agent system on AWS Bedrock where specialized agents collaborate to handle the full lifecycle, coordinated through A2A and MCP protocols.
📥
Intake Agent
Parses incoming claims, extracts structured data, and routes to the correct specialist agent.
🔍
Assessment Agent
Checks policy coverage, cross-references claim history, and flags potential fraud patterns.
💬
Communication Agent
Drafts customer updates, acknowledgements and resolution notices in the right tone and language.
🚨
Escalation Agent
Detects edge cases beyond automated scope and hands off to a human adjuster with full context.
🔗
A2A: Agent to Agent
Standardized protocol for agents to communicate directly: delegate tasks, share context, report results, without a central coordinator bottleneck.
🛠
MCP: Model Context Protocol
Internal tools (policy database, CRM, claims system) exposed as MCP endpoints. Agents call tools the same way regardless of the underlying system.
📊
Evaluation framework
Each agent decision is scored for accuracy and compliance. The evaluation layer mirrors what AXA's team used on Azure, adapted for Bedrock.
Tech Stack
Two clouds, one coherent architecture.
Knowledge Base stack
🔎
Azure AI Search
Hybrid vector + keyword retrieval across multi-entity document corpus
🧠
Azure AI Foundry
LLM orchestration, prompt management and grounded generation
✅
Azure Evaluation
Automated scoring on relevance, groundedness and coherence
👤
Human-in-the-Loop
Expert review and feedback loop to continuously improve quality
Claims automation stack
☁️
AWS Bedrock
Managed LLM backend powering all agent reasoning and generation
🔗
A2A Protocol
Direct agent-to-agent communication without central bottleneck
🛠
MCP Endpoints
Internal tools exposed as Model Context Protocol endpoints
📊
Evaluation layer
Per-decision scoring for accuracy, compliance and confidence
BotiqueAI
Custom AI systems for enterprise