Why Diagnostic Support Queries Break Standard Chatbot Logic

A customer calls with a specific error code like E4 on a high end industrial boiler. A standard AI model might suggest checking the manual or offer a generic list of plumbing tips. This happens because the model prioritizes conversation flow over technical precision. For businesses with complex products, a polite guess is worse than a redirect to a human. It wastes time and risks equipment damage.

Generic language models are trained to be helpful. This often results in polite guesses when they face technical ambiguity. In a diagnostic scenario, the agent needs to access the exact documentation for that specific model. Duvi handles this by crawling the business website and parsing uploaded files into a vector store. When the customer mentions an error code, the agent pulls the response from the actual technical manual instead of relying on its training data. This grounding ensures the answer is based on real content. Even with access to manuals, developers often find that Retrieval-Augmented Generation is insufficient for transactional support when a specific action must follow the answer.

Solving a technical fault requires more than reading a manual: it requires knowing the state of the machine. An agent that only talks is a bottleneck. If the boiler is connected to a cloud system, the agent should check the pressure readings before giving advice. Duvi agents use tool calls to interact with HTTP APIs or MCP servers. The agent can pull live telemetry from a database or log a service ticket in Salesforce without a human intervening. This moves the interaction from a search task to a resolution task. This shift addresses the gap between answering a question and solving a ticket.

Relying solely on static data often leads to failure in scheduling or active troubleshooting. This is a common friction point for agents that only read documentation. Consistency across channels is another failure point for support teams. A technician might start a conversation on the website chat while at their desk, then switch to a phone call once they are in the boiler room. If the website bot and the phone system use different logic, the technician receives conflicting instructions. Duvi uses the same agent and the same knowledge base for every surface. This prevents the tax of fragmented front doors where customers repeat themselves across channels.

The business connects its own Twilio or Meta accounts to handle phone and WhatsApp traffic. Whether the customer sends a WhatsApp message or calls a dedicated number, they interact with the same system prompt and tools. Voice interactions introduce a specific constraint that text does not: latency. A two second pause in a chat window is acceptable, but a two second pause on a phone call feels like a dropped connection. Duvi allows users to choose the transcriber, model, and voice for their agent to hit specific latency targets. The platform provides a combined latency estimate so the business can see the impact of their choices before they go live.

Tuning the system prompt for voice also helps. Shorter sentences and clearer instructions prevent the agent from rambling while the customer is trying to work. Large scale support operations often struggle with the transition from automation to human assistance. When a diagnostic check reveals a hardware failure that requires a physical part, the agent must hand over the conversation. Duvi supports live chat handovers to humans. If the interaction happens over the phone, the call can be recorded and analyzed. These completed conversations are turned into analytics. This data helps the business identify which error codes are most common and which parts of the documentation are confusing for the agent.

Launching an agent with these capabilities does not require an engineering team. The process starts by describing the job of the agent in a conversation. The builder then provides sections for the system prompt, which is written in six distinct blocks. These blocks define the rules and the logic the agent follows. Facts stay in the knowledge base, while the prompt handles the behavior. This structure allows a support manager to update the agent logic without writing code or waiting for a sprint cycle.