Unifying Business Logic Across Voice and Text Channels

A customer asks about a refund policy on WhatsApp and receives a specific set of requirements. Ten minutes later, they call the support line to start the process but find that the phone system provides conflicting information. This discrepancy happens because the web chat and the phone script live in different software silos managed by different teams. The customer loses trust the moment they realize the business is not speaking with one voice.

Fragmented support logic is a common failure mode for growing companies. Maintaining a consistent set of rules across text and voice is difficult when each channel requires its own unique integration and separate database configuration. Most businesses experience knowledge drift where the website is updated frequently while the phone tree remains static for months.

Duvi addresses this by using a single agent for every channel. When you build an agent, you deploy it to a website via a script tag or connect it to communication providers like Meta and Twilio. The brain of the agent does not change between these surfaces. This architecture prevents the conversation persistence tax where losing customer context resets resolution time. If a customer sends an asynchronous message on WhatsApp, they get a reply on WhatsApp. If they call the same number, they speak to the same agent using the same knowledge.

The knowledge base is the source of truth for these interactions. Instead of writing manual scripts for every possible question, you point Duvi toward your existing documentation. The system crawls your website and parses uploaded files into a vector store. This ensures that the agent provides answers based on real content rather than a generic model intuition. You can monitor the indexing process through four distinct states to ensure the data is ready for retrieval. This approach helps solve the friction where a knowledge base is hidden from the people who need it most.

Writing the instructions for these agents requires a specific structure to maintain control over how they behave. Duvi uses a system prompt divided into six blocks. Every rule you give the agent must include a reason. This prevents the agent from making arbitrary decisions and helps it understand the context of the business logic. While the core facts remain the same, you can tune the prompt specifically for voice to ensure the agent handles natural interruptions or the pacing of a phone call. This method avoids the trap of generic chatbots that fail to handle specific technical inquiries.

Answering questions is often just the first step in a support interaction. Real resolution requires taking an action. An agent that can only quote a PDF is just a search bar. To solve actual tickets, agents need to interact with your existing infrastructure through connectors for HTTP APIs, MCP servers, Salesforce, and external databases. These tools are available across all channels simultaneously. A customer on the website can check an order status through an API call, and a caller on a phone line can perform the same check using the same tool. There is no additional integration to write for the second channel. The agent understands when to trigger a tool based on the intent of the conversation. This moves the interaction beyond search and into execution.

Managing these interactions does not require a large engineering lift. Launching an agent is done through the interface, and usage is managed through a credit system. Businesses connect their own numbers through providers like Meta or Twilio, keeping the communication infrastructure under their own control. For cases where the AI cannot resolve the issue, the system supports a handover to a human. This ensures that complex problems or high value customers are not left in a loop. After a conversation ends, the system analyzes the interaction to provide insights into what customers are asking and where the knowledge base might need more detail. This feedback loop allows for continuous refinement of the agent without needing to rebuild the underlying logic for each channel.