Why Chat Prompts Fail on a Phone Call

The Verbosity Problem

Most AI agents built for chat fail when they answer a phone call. A customer expects a short response. They do not want a 300 word paragraph read back by a robotic voice. If the agent takes four seconds to think before speaking, the caller will assume the line is dead and hang up. The transition from a web widget to a live phone line requires a fundamental shift in how the agent processes knowledge. Text based models are trained to be helpful and thorough. On a website or over WhatsApp, a detailed explanation is useful because the customer can skim the text to find relevant details. On a phone call, this behavior is a liability. The caller cannot skim audio and must listen to every word in sequence. This often leads to Knowledge Drift and Why Phone Lines Struggle to Match Website Content when voice and web channels use separate logic.

Duvi addresses this by allowing specific tuning for voice within the agent builder. When you describe the job in a conversation, the system helps you structure the prompt for brevity. Long lists must be converted into options. Instead of reading ten available appointment slots, the agent should ask for a preferred timeframe. This ensures Why Customers Repeat Themselves Across Web and Voice Channels is a problem of the past by maintaining a single source of truth.

Structuring the Six Block System Prompt

A vague prompt leads to unpredictable behavior. Duvi agents use a system prompt divided into six specific blocks. This structure prevents the model from hallucinating or losing track of its primary objective. Each block serves a distinct purpose, from defining the persona to listing the tools available through HTTP APIs or Salesforce. Telling an agent to be brief is less effective than explaining that callers have limited attention spans. This logic helps the underlying model understand the constraints of the voice channel. Facts are kept out of the prompt and stored in the knowledge base. This separation ensures the agent refers to crawled website data or uploaded files instead of relying on generic training data.

Managing Latency Targets

Latency is the primary killer of voice AI adoption. A delay of two seconds feels like an eternity in a live conversation. This delay is the sum of transcription, model inference, and the time required for speech synthesis. In the Duvi dashboard, you can see a combined latency estimate as you build. Choosing a faster model or a specific transcriber directly impacts this number. The goal is to hit latency targets that match human speech patterns. If the agent needs to call an external API or an MCP server to check an order status, that action must happen fast enough to maintain the flow. If the process takes too long, the agent should be prompted to provide a verbal filler. This prevents The Action Latency Friction where answering a question without updating the database creates a secondary support ticket.

Knowledge Parity Across Surfaces

A common failure mode in customer support is the information gap between channels. A customer might see a promotion on a website but find the phone agent has no record of it. Because Duvi uses the same knowledge base for every surface, this gap disappears. The vector store created from your files and website URLs serves the web widget, the WhatsApp business account, and the phone line simultaneously. This approach unifies business logic across voice and text channels. The agent always replies on the channel the customer used. If a customer sends a message on WhatsApp, the reply stays there. If they call the connected Twilio number, they get a voice response. The logic remains identical because the agent is the same. The medium of delivery is the only difference. This prevents The Fragmented Front-Door Tax where leads are tired of introducing themselves to the company multiple times.

Handling the Handoff

Not every conversation can be solved by an AI. A complex technical failure or an upset customer might require a human. Duvi allows live chats to hand over to a human agent when the AI reaches its limit. For phone calls, the system replaces the phone tree with contextual intent. It provides recordings and analysis of completed conversations. This allows managers to review how the agent handled specific queries and adjust the knowledge base or the system prompt accordingly. Launching an agent requires no engineering work. You connect your own number or WhatsApp account through Meta or Twilio, paste a script tag for the website widget, and the agent is live. Usage runs on credits, which makes it easy to scale based on actual conversation volume rather than seat licenses. This setup allows a business to deploy a voice and text agent that understands its specific business data in a single afternoon.