The idea sounded straightforward: answer every call to a roofing business, collect the customer’s information, and remain available after office hours. The product became interesting when I stopped thinking of it as a talking form and started treating it like a front-desk teammate.
A real teammate remembers returning callers. The agent therefore begins with the phone number, looks for an existing customer, and brings forward useful context such as the name, address, and earlier conversations. It should not repeat every question when the answer is already known. At the same time, remembered information must remain visible and correctable; confident memory is dangerous when a customer has moved or a prior note was wrong.
Urgency created another important boundary. An active leak after a storm is not just another lead. The conversation must identify signals such as water entering the building, structural damage, or immediate safety risk, then mark the case clearly for human attention. The agent can organize and prioritize the handoff, but it should not pretend to make a professional safety assessment.
The final handoff is where the product earns trust. A transcript alone forces the roofing team to do the work again. A useful handoff contains the caller’s contact details, property address, problem, urgency, availability, and a short explanation of why the case was prioritized. The human should be able to continue the relationship without asking the customer to restart it.
The hackathon version was intentionally scrappy, but the design principle stayed clear: automation should preserve context and reduce repetition. If the customer feels heard and the team receives a clean next action, the agent is doing more than answering the phone.