PolicyLens AI — Smarter Insurance Comparison for Australia
Most insurance comparison platforms lead with price — the one metric that tells you almost nothing about real protection. Two policies at the same price can differ dramatically in sub-limits, exclusions, and benefit structures. The only way to spot those differences is to manually read 9 Product Disclosure Statements side by side — and nobody does that.
PolicyLens AI changes that. It lets Australian business owners compare coverage across 9 major insurers by actual protection strength — not price — through plain English questions and structured, side-by-side comparisons grounded in real policy wording.
Why Business Insurance Comparison Is Broken
The client was an Australian insurance broker who saw a fundamental problem: every comparison platform leads with price, but price tells you nothing about what happens when you need to make a claim. The information that actually matters lives buried inside Product Disclosure Statements — dense, 60+ page legal documents that each insurer formats differently. Manual comparison takes hours and doesn’t scale.
They’d tried simpler approaches — static comparison tables, PDF summaries, a basic FAQ chatbot. None worked. Static tables went stale immediately, and the FAQ chatbot couldn’t compare specific coverage details across insurers or drill into a single policy’s exclusions. They needed an AI system intelligent enough to understand insurance documents at a granular level and deliver answers through a simple conversational interface.
The Challenges
9 PDSs, Zero Standardisation
Nine insurers, nine different PDS formats, terminology, and structures — all needing to be normalised into one comparable framework.
Two Types of Questions, Two Different Architectures
Instant structured lookups and deep PDS retrieval needed to work seamlessly within the same system.
Regulatory Sensitivity
Every response needed strict guardrails — comparing policy wording only, never crossing into regulated financial advice.
Trust Through Source Attribution
Every comparison had to be traceable to a specific PDS clause — source attribution wasn't a feature, it was foundational.
Zero Tolerance for Hallucination
A fabricated detail could mean the wrong policy purchased — the system needed aggressive grounding, refusing to speculate when information wasn't in the PDS.
Technologies Used
Python
FastAPI
LangGraph
LangChain
AWS EC2
Docker
Pinecone
WebSocket
Compare Fast, Then Go Deep
We built a hub-and-spoke multi-agent system: one Orchestrator Agent handles all comparisons using pre-extracted structured data, and nine Specialist Agents — one per insurer — handle deep-dive queries using RAG retrieval from the full PDS text. The Orchestrator generates instant comparison tables; the Specialists answer detailed coverage questions with source citations. Users move seamlessly between comparison and deep-dive modes.
Orchestrator Agent — Instant Comparisons
Generates instant side-by-side comparison tables across all 9 insurers — with seamless handoff to Specialist Agents for deeper queries.
9 Specialist Agents — Deep Policy Expertise
Each Specialist Agent searches only one insurer's PDS — returning clause-referenced answers and explicitly stating when information isn't found.
Two-Layer Data Ingestion
Structured variable extraction powers instant comparison tables. Full-text RAG chunking with clause-level metadata in insurer-specific namespaces powers deep Specialist Agent retrieval.
Regulatory Guardrails & Anti-Hallucination
Non-bypassable safety layers enforce scope, block hallucinations, and prevent prompt injection — comparing policy wording only, never recommending a purchase.
Conversation State & Seamless Agent Hand-Offs
The Orchestrator tracks full conversation state — prior comparisons, active insurer, and routing decisions — so users never lose context switching between comparison and deep-dive modes.
Embeddable Widget with Real-Time Streaming
Single script tag deployment, no registration required. Real-time WebSocket streaming and structured comparison tables rendered directly in the chat interface.
Measurable Impact
Building for a vulnerable user population means impact is measured in trust and engagement, not just technical uptime. Here’s what Aura achieved:
9
<2s
100%
0
Quality Outcomes
Protection-First Buying
Users made coverage decisions based on sub-limits and exclusions rather than price alone — reducing under-insurance risk.
Broker Time Savings
Replaced 2–3 hours of manual PDS analysis per comparison with conversational queries that take minutes.
Zero Hallucination
Pre-extracted structured data for comparisons and grounded RAG for deep dives meant no fabricated coverage claims in production.
Scalable to New Insurers
Adding a new insurer requires ingesting their PDS and creating a templated Specialist Agent. No core system changes needed.
Regulatory Confidence
Zero compliance violations in production — no financial advice, no guardrail bypasses, full regulatory confidence from day one.
How We Built PolicyLens AI
What Makes PolicyLens AI Different
Protection-Led, Not Price-Led
Users compare actual coverage—limits, exclusions, and benefits—shifting decisions from cheapest to best protection.
Instant Comparison Tables
Comparison queries generate instant side-by-side tables from pre-extracted data—no slow document searches needed.
Deep Dive into Any Insurer
Seamlessly move from comparison to deep insurer analysis—covering exclusions, conditions, and clauses with source-backed references.
Embeddable Widget, Zero Friction
Single script deployment—no login required—with real-time responses and instant comparison tables in chat.
Per-Insurer Knowledge Isolation
Each insurer’s data lives in a completely separate vector database namespace. No cross-contamination between insurers’ policy details.
Source Attribution on Every Answer
Every claim is traceable to a specific PDS section and clause reference. Users can verify any answer against the source document.
Project Screenshots
Common Questions About This Project
A public-facing AI chatbot that compares business insurance coverage across 9 Australian insurers based on actual policy wording — sub-limits, exclusions, conditions, and benefit structures — rather than price. Users get instant comparison tables and can deep-dive into any single insurer’s PDS with source-attributed answers.
Existing tools lead with price. PolicyLens AI leads with protection. It delivers instant structured comparison tables from pre-extracted data and deep-dive RAG retrieval from full PDS text — with every answer citing the source section for verification.
Comparison queries use pre-extracted, verified structured data. Deep-dive queries are grounded exclusively in retrieved PDS text with source citations. When information isn’t in the PDS, the system says so explicitly rather than guessing.
No. PolicyLens AI compares policy wording only. It never recommends purchases, assesses insurer strength, or provides financial advice. These boundaries are enforced at multiple layers and cannot be bypassed.
Yes. Adding a new insurer requires ingesting their PDS and creating a templated Specialist Agent. The Orchestrator automatically incorporates new insurers into comparisons. No core changes needed.
Absolutely. The architecture — multi-agent RAG with document isolation, structured data extraction, source attribution, and regulatory guardrails — applies to any industry where users need to compare complex documents: legal contracts, healthcare policies, financial products, regulatory compliance.
Need an AI chatbot that understands your documents — with source attribution, regulatory guardrails, and answers your customers can trust?
We build document-grounded AI systems for insurance, legal, healthcare, and finance. If your domain demands accuracy and transparency, let’s talk.