A central challenge in this project was designing a conversational intake agent capable of performing reliably across voice and chat environments. While text-based interactions allowed for relatively structured information capture, voice introduced greater complexity, particularly when collecting critical customer details such as phone numbers and email addresses. In spoken conversations, users often provide this information in inconsistent formats, repeat themselves, or speak quickly, creating a greater risk of incomplete or inaccurate intake. Because the agent’s role was to support downstream help-desk workflows, maintaining precision in data capture was essential to creating support tickets efficiently and without manual correction.

An additional challenge was the requirement for multilingual performance in both English and Spanish. The agent needed to preserve the same level of clarity, responsiveness, and intake accuracy regardless of language while also adapting naturally to different phrasing patterns and conversational expectations.

A scalable and effective testing strategy was also necessary for the project. Evaluating conversational agents, particularly those operating across channels and languages, can be time-intensive when performed manually. The team needed to validate whether the agent could respond appropriately, consistently gather the right information, remain aligned with the support intake objective, and handle realistic customer scenarios with a high degree of reliability.

To address the complexity of cross-channel conversational intake, the solution centered on careful prompt engineering supported by rigorous testing. The prompting framework was designed to help the agent guide users through support-focused conversations while capturing the essential information required for ticket creation. Special attention was given to how the agent requested and confirmed sensitive contact details, particularly in voice interactions where misheard or incomplete information could disrupt the handoff to the help desk. By refining conversational instructions and response behavior through repeated evaluation, the team improved the agent’s ability to create a smoother and more dependable intake experience across voice and chat.

For multilingual delivery, the approach combined prompt engineering with appropriate platform configuration and structured testing in both English and Spanish. The agent was tuned to maintain consistent support workflows across languages while accommodating linguistic variation in how users describe issues and provide their details. Proper configuration within the Wildix environment ensured the agent could operate effectively in each language context while multilingual testing helped validate that the experience remained coherent, accurate, and aligned with operational requirements regardless of language.

To overcome the limitations of manual conversational agent testing, the team implemented an evaluation workflow in a sandbox environment that enabled faster and safer iteration. Testing was strengthened through the use of an LLM User Simulator and an LLM-as-a-Judge methodology powered by the DeepEval framework, allowing the team to model realistic interactions and assess agent performance at scale. In addition, synthetic test cases were generated from real support scenarios identified in historical inbound calls to the IT support line, making the evaluation process more representative of actual user behavior. This testing strategy made it possible to systematically measure performance, identify weaknesses quickly, and accelerate improvement before deployment.