InsureVoice — AI Voice Insurance Advisor

A natural AI voice agent automating insurance: from plan recommendation to form filling to document verification and underwriting , completing the sales journey in minutes!

Comment

InsureVoice — AI Voice Insurance Advisor

Inspiration

Insurance advisory remains one of the most manual and compliance-sensitive processes in financial services. Agents often spend 60–90 minutes collecting customer information, evaluating eligibility criteria, comparing products, and ensuring recommendations satisfy regulatory requirements. Every recommendation must be accurate, explainable, and compliant.

We saw an opportunity to rethink this experience using Agentic AI. Instead of navigating forms and comparing policies manually, what if a customer could simply speak naturally and receive compliant, personalized insurance recommendations within seconds?

Our broader vision extends beyond recommendations. We believe voice-first AI agents can eventually support the entire insurance lifecycle—from product discovery and premium estimation to underwriting, verification, policy issuance, and customer servicing. InsureVoice is our first step toward that future.

What it does

InsureVoice is a voice-driven AI insurance advisor that helps customers discover suitable insurance products through natural conversation.

A customer provides eight profile attributes using voice:

  • Name
  • Age
  • Smoker status
  • Income
  • Health condition
  • Family size
  • Coverage goal
  • Desired sum assured

The system performs semantic product discovery using Elastic ELSER v2, validates eligibility and compliance requirements, ranks suitable products, and generates a personalized Top-3 recommendation with clear explanations for each option.

The platform also intelligently handles unsupported insurance categories such as car, bike, and travel insurance by redirecting users appropriately rather than producing unreliable responses.

How we built it

We built InsureVoice as a hybrid Agentic AI architecture on Google Cloud that combines deterministic workflows with LLM-powered reasoning.

Phase 1: Deterministic Orchestration (<50ms)

A lightweight Python orchestration layer handles:

  • Out-of-scope request detection
  • Customer profile collection
  • State management
  • Follow-up question routing

This layer resolves the majority of conversational tasks without requiring LLM intervention, significantly reducing latency and increasing reliability.

Phase 2: Agentic Intelligence Layer (~8s)

The reasoning layer uses Google ADK and Gemini models.

Components include:

  • Google ADK LlmAgent
  • Gemini 2.5 Flash Lite
  • ADK before_model_callback
  • Elastic MCP Server
  • Compliance validation Cloud Functions
  • Product ranking Cloud Functions
  • Gemini-powered recommendation sub-agent

The workflow orchestrates multiple specialized capabilities:

  1. Semantic product retrieval
  2. Compliance validation
  3. Suitability ranking
  4. Recommendation generation
  5. Personalized explanation generation

Rather than relying on a single prompt, each capability operates as an independent tool within the agent workflow.

Phase 3: Voice Response Layer (<200ms)

The final response pipeline includes:

  • Response assembly
  • Audit logging
  • Google Chirp 3 HD Text-to-Speech

This provides a consistent and natural end-to-end voice experience.

Challenges we ran into

Building a voice-first AI system for a regulated domain required balancing conversational intelligence with deterministic business controls.

One challenge was ensuring reliable tool orchestration. While LLMs are excellent reasoning engines, insurance workflows require predictable execution paths and guaranteed compliance checks. We addressed this using Google ADK callback controls to enforce tool execution behavior and maintain workflow consistency.

Another challenge involved preserving context across multiple tools and agents. Product search, compliance validation, ranking, and recommendation generation all require access to a shared customer profile. We solved this through centralized session-state management that enables reliable data flow across the workflow.

We also encountered challenges with real-time voice streaming during periods of prolonged silence. By introducing keepalive mechanisms within our WebSocket bridge architecture, we improved session stability and maintained a smooth conversational experience.

Finally, we intentionally designed the platform for future extensibility. Our modular architecture allows additional capabilities such as underwriting, KYC verification, premium simulation, and policy issuance to be integrated without redesigning the core system.

Accomplishments that we're proud of

  • Built a production-inspired multi-agent workflow using Google ADK and Gemini models.
  • Integrated Elastic ELSER v2 through MCP to enable semantic insurance product discovery beyond keyword matching.
  • Achieved a 0% compliance false-pass rate through deterministic compliance validation.
  • Designed a hybrid architecture where more than 80% of conversational turns are handled through low-latency deterministic processing.
  • Delivered personalized recommendations with explainable reasoning for every suggested product.
  • Implemented a complete voice experience using Google Speech technologies and Chirp 3 HD.
  • Established a scalable Google Cloud foundation for future insurance automation workflows.

Most importantly, we demonstrated how Agentic AI can orchestrate real business processes while maintaining compliance, explainability, and operational reliability.

What we learned

This project reinforced a key lesson about enterprise AI: successful AI systems are not built around language models alone. They are built around the orchestration of agents, tools, business rules, and trusted data sources.

We learned that deterministic systems and AI reasoning complement each other exceptionally well. The most reliable architecture is one where LLMs focus on reasoning and decision support while deterministic services handle validation, compliance, and workflow execution.

Google ADK proved valuable for introducing structured agent orchestration and operational control. MCP-based integrations provided a clean mechanism for connecting enterprise tools and retrieval systems, while Elastic ELSER v2 demonstrated the power of domain-specific semantic search without requiring custom model training.

Most importantly, we learned that the future of enterprise automation lies in Agentic AI ecosystems where intelligent agents can retrieve information, execute actions, validate outcomes, maintain auditability, and collaborate across complex business workflows.

What's next for InsureVoice — AI Voice Insurance Advisor

Our current implementation focuses on insurance discovery and recommendation. The next phase is expanding InsureVoice into a complete AI-powered insurance onboarding platform.

Phase 1: Interactive Form Filling & Real-Time Aggregator Integration

  • Dynamic voice-based form filling that extracts application details naturally during conversation.
  • Secure integrations with financial and identity data providers to pre-fill and validate customer information.
  • Real-time actuarial premium simulations using Cloud Functions.
  • Expansion of the product catalog from 28 to 48+ insurance products.

Phase 2: AI-Powered Automated Underwriting & Instant Verification

  • KYC workflows powered by Google Cloud Document AI.
  • Automated identity verification using document and biometric validation.
  • AI-driven underwriting agents capable of evaluating risk and compliance requirements.
  • Firestore-backed session architecture for enterprise-scale multi-modal workflows.
  • Automated policy generation, audit logging, digital signatures, and payment integration.

Our long-term vision is to transform InsureVoice into a fully autonomous, voice-first insurance platform capable of supporting discovery, compliance, underwriting, verification, issuance, and servicing through a unified Agentic AI ecosystem.

Built With

Share this project:

Updates

Submission history