To evaluate the responses of your AI chatbot, you can use developer platforms like Vellum to simulate conversations and evaluate the responses using ground data or LLMs. Once in production, you can further optimize and improve performance by testing the workflow with real user data.

To make an AI chatbot sound more human, make sure to invest some time to engineer a prompt that will give the best answer, incorporate tone and continuously test and improve with ground data or real users. Additionally, ensure there's an option for escalating to human customer support for queries the chatbot can't adequately address.

To build a chatbot from scratch follow these steps:

1. Define the purpose and scope of your chatbot;
2. Define the various intents users might;
3. Choose a framework or a AI development product like Vellum;
4. Design the conversation flow and user interactions;
5. Test the prototype with a bank of test cases to see how it performs;
6. Deploy to production, and monitor how users interact with it; and
7. Continuously monitor and update the chatbot for improvements.