Build a Voice AI Agent to Answer Insurance Policy Questions

Independent insurance agencies often struggle to efficiently handle the high volume of routine policy questions, diverting valuable agent time from more complex tasks like claims triage or in-depth policy comparisons. Many agencies first attempt to address this bottleneck with their existing phone system's Interactive Voice Response (IVR).

Standard IVRs, however, are inherently rigid. They rely on fixed, menu-based trees ('press 1 for sales, press 2 for billing') and are incapable of understanding natural language. They cannot answer dynamic questions such as, 'Am I covered for a cracked windshield if I live in Florida?' Instead of resolving inquiries, they typically only route calls, often after frustrating the customer, which ultimately increases an agent's workload rather than decreasing it.

Following the limitations of IVRs, some agencies explore off-the-shelf voicebots offered by CRM or marketing platforms. While these tools can effectively deliver pre-scripted answers to general FAQs, they consistently fail when a customer asks a question about their *specific* policy details. For instance, a generic bot might state, 'Most auto policies include glass coverage,' but it lacks the direct integration with your AMS (like Applied Epic or HawkSoft) to confirm if *that specific customer's policy* actually includes it. This inability to access and interpret individual policy data from the core system means that any important, personalized inquiry must be escalated to a human agent.

This gap creates significant inefficiency. An agent might spend several minutes logging into Vertafore, searching for a policy number, and manually reading relevant clauses to answer a question that a client expected an instant response for. This not only delays service but also ties up agents who could be focused on tasks requiring human judgment, such as reviewing complex FNOL reports or addressing detailed benefits enrollment queries. The current workflow often adds an unnecessary step for both the customer and the agent, leading to dissatisfaction and increased operational costs.

Syntora's approach to engineering a voice AI system for policy questions begins with a thorough discovery phase. This phase would involve connecting to your chosen telephony provider (e.g., Twilio) and establishing read-only API access to your Agency Management System, such as Applied Epic, Vertafore, or HawkSoft. We would analyze existing call logs and support tickets to identify the most common and time-consuming policy questions agents currently handle, focusing on those with a high potential for accurate automation.

The core voice agent would be engineered as a Python service using FastAPI, designed for deployment on AWS Lambda. This architecture ensures high efficiency, scalability, and cost-effectiveness. When a customer call is received, a service like Deepgram would perform real-time speech-to-text transcription. The architecture would aim for transcription latency under 2 seconds. The transcribed text would then be sent to the Claude API, which performs intent recognition to classify the caller's question (e.g., 'coverage_check', 'billing_inquiry', 'update_address', or 'claim_status'). We've used Claude API for document parsing in financial contexts, and its capabilities are directly applicable to understanding nuanced insurance queries.

Once the intent is recognized, the FastAPI service would securely query your AMS API for the caller's specific policy details. Relevant information extracted from the policy—such as coverage limits, deductibles, or specific endorsements—would be passed as context to a second Claude API call. This allows the AI to generate a precise, natural language answer tailored to the individual policy. This answer is then converted back to audio using a text-to-speech API and delivered to the customer. The design target for the entire interaction, from the customer's question to the AI's answer, would be under 5 seconds.

If the AI's confidence in an answer falls below a defined threshold, or if the caller explicitly requests to speak with a human, the system would automatically transfer the call to a live agent queue. Every transcript, AI response, and system action would be logged in a Supabase database. This logging provides a comprehensive audit trail and a valuable data source to identify new question types or edge cases for potential future automation, allowing for continuous improvement of the system over time. Client deliverables would include the deployed system, comprehensive documentation, and knowledge transfer to internal teams.