Moving Beyond Static Answers to Actions Driven by APIs

A customer visits your website to ask about product compatibility. The bot answers correctly by reading technical documentation. When the same customer calls to check if the part is in stock at a specific warehouse, the system often fails. This failure happens when the transition from static knowledge to dynamic data is not handled. The gap between answering a question and solving a ticket occurs because most systems treat web chat and phone lines as separate entities with different logic. Duvi allows a business to deploy the same agent across the website and phone lines. This agent uses the same vector store for knowledge and the same tools for actions.

An agent that only reads from a crawled website is a search interface. To solve actual tickets, the agent must interact with internal systems. Duvi agents use tools to call HTTP APIs or MCP servers. They can also connect to Salesforce. When a customer provides an order ID on a phone call, the agent triggers a GET request to your backend. It retrieves the status and speaks the answer to the customer. Connecting support agents to live business data allows for real time resolution using current information instead of static FAQs. Documentation only agents increase the manual workload for support teams by forcing humans to perform the actual data entry after the AI finishes the call.

This process requires control over the system prompt. Duvi uses a structure of six blocks to define agent behavior. You keep facts in the knowledge base and use the prompt to define the logic for calling an API. Every rule in the prompt requires a clear reason. This prevents the model from hallucinating a tool call when it should provide a text answer. This precision helps avoid the cost of the internal ticket transfer, which typically costs twenty dollars per incident in labor.

In a text based WhatsApp conversation, a three second delay while an API returns data is acceptable. On a live phone call, three seconds of silence causes callers to think the connection dropped. The latency threshold is a major factor in the success of AI voice. Duvi provides a combined latency estimate within the agent builder. You can choose specific transcribers or models to hit the targets required for a natural conversation. You also select the voice. If a tool call to your database takes too long, the experience degrades. You can tune the model selection to prioritize speed for voice interactions. This helps the system parse complex JSON responses without causing long pauses.

Businesses often manage a WhatsApp account through one provider and a phone system through another. This leads to fragmented customer history. Duvi connects to existing Twilio or Meta accounts. A reply always goes back through the channel the customer used. The surface parity friction explains how disparate logic for web and phone support creates knowledge silos. With Duvi, the knowledge base stays synchronized. When you upload a new PDF of technical specifications, that information is available to the web widget and the WhatsApp bot. If the agent cannot resolve the issue through its connected tools, it can hand the conversation to a human.

Once an agent is active, the focus shifts to how well it uses the provided tools. Completed conversations are analyzed into analytics. You can see where the agent successfully called an API. You can also see where it struggled to find information in the crawled site content. This data allows for adjustments to the system prompt or the knowledge base files. Launching an agent does not require an engineering team. You describe the job in a conversation within the agent builder, index your files, and embed the script tag. The technical work happens in the background so the business can focus on the logic of the customer interaction.