One of the project’s main challenges was creating a conversational intake experience that could perform effectively across voice and web chat channels. While customers are generally comfortable sharing information through chat, collecting accurate details over the phone introduces greater friction. In particular, contact information such as names, phone numbers, email addresses, and policy-related details can easily be misheard, fragmented, or provided in inconsistent formats during spoken interactions. Because the agent needed to serve as a reliable first point of contact and support smooth handoffs to live staff, the intake flow had to be carefully structured to maintain a natural customer experience while capturing complete and usable data.

Another significant obstacle was preparing the knowledge base that would power the agent’s property and casualty insurance question-answering capabilities. The relevant information existed across the company’s broader ecosystem rather than in a single clean, deployment-ready source. This meant the project required extraction, organization, and refinement of business information so the agent could deliver accurate and consistent responses across channels.

Ensuring dependable performance over time presented another significant challenge. Because the assistant was expected to operate continuously and represent the agency in customer-facing interactions, testing could not be limited to occasional manual reviews. Manual validation alone would have been too slow and too narrow to cover the range of real-world conversations, edge cases, and failure scenarios that a production-grade receptionist and insurance support agent would encounter.

To address the complexity of intake across voice and chat, the solution centered on careful prompt engineering. The conversational flows were designed to guide users through information capture while maintaining a natural experience and increasing accuracy and completeness. Special attention was given to how the agent requested, confirmed, and reformatted contact details in phone-based interactions, where ambiguity is more common. By structuring prompts to encourage step-by-step clarification and confirmation, the agent was better able to collect critical customer information reliably while maintaining a professional and efficient experience across channels.

To support the agent’s insurance knowledge capabilities, relevant content was extracted and cleaned from the client’s existing ecosystem using a human-assisted scraping process supported by LLMs. This helped transform distributed business information into a more usable knowledge foundation for the assistant and improved consistency in how it handled common property and casualty insurance questions.

Given the importance of reliability, the agent was validated through a heavily automated testing workflow in a sandbox environment. It was tested against synthetic scenarios generated from real inbound call patterns, which allowed the team to evaluate performance at scale using realistic conversation types rather than relying solely on manual checks. In addition to general functionality testing, adversarial testing was conducted to expose weaknesses, stress conversational boundaries, and identify failure points before deployment. This approach significantly improved testing speed and coverage while helping ensure the system could perform more consistently in live customer interactions.