Replacing the Phone Tree with Contextual Intent

Most customers hang up when they realize they are trapped in a numeric phone menu. They want a specific answer about a part number or an appointment time. Traditional systems only understand single digits. This friction creates a queue of frustrated callers. It results in a backlog of missed leads that never call back.

Traditional Interactive Voice Response (IVR) systems rely on rigid logic. They follow a hardcoded tree that cannot adapt to how people speak. A standard system fails if a caller asks if a specific branch has a red mountain bike in stock: it can only offer to transfer them to sales. The caller then waits on hold until a human can check the inventory. This is a common version of the specific question trap where high intent leads are lost to generic responses.

An AI voice agent maps spoken requests directly to business data. The caller interacts with a natural language interface that has already read the company website and technical manuals. This approach addresses the contact us friction where static knowledge is hidden from the customer.

Moving knowledge into the voice channel

A Duvi agent does not rely on a person to write every possible question and answer. It uses a vector store to hold information gathered from a website crawl or uploaded PDFs. This process is detailed in our overview of building agents with natural language. When a customer calls a Duvi connected number via Twilio, the agent pulls relevant facts from this indexed content to form a response.

This indexing happens in four states: pending, crawling, indexing, and completed. Once the state is completed, the agent can answer technical questions without a developer writing new code. If the website says the service center closes at 6 PM on Fridays, the agent knows this. If the pricing page updates, the agent knows that too.

The same logic applies to WhatsApp. A business can connect its Meta WhatsApp Business account and use the same knowledge base. A customer might start a conversation on the website chat before leaving their house. They then follow up via a phone call. Because the agent draws from a single source of truth, the answers remain consistent across every channel.

Action over information

Answering questions is only half of the work. A voice agent becomes a functional part of the team when it can perform tasks. This is handled through tool calls. Duvi agents trigger HTTP APIs or connect to Salesforce to update a record or check a database. Connecting agents to live data avoids the tool call threshold where agents that only provide information increase the manual workload for support teams.

If a caller wants to reschedule a delivery, the agent does not just explain the policy. It uses a tool to check the current order status and another to write the new date to the business system. This happens while the caller is still on the line. The agent identifies the intent, selects the correct tool, and executes the call.

Managing these interactions requires precise control over the agent behavior. In the Duvi builder, this is done through Developer mode. This allows for a system prompt structured into six blocks. These blocks define the persona and the constraints. Rules are paired with specific reasons to help the model understand why it must verify a phone number before sharing order details. This prevents the triage trap where expert staff spend hours acting as a search bar for customers.

Latency and the human experience

Voice interactions are sensitive to delays. A two second pause in a chat widget is acceptable. A two second silence on a phone call feels like a dropped connection. To manage this, the Duvi dashboard provides a combined latency estimate based on the chosen transcriber and the model. The selected voice also changes this calculation.

Choosing a faster transcriber and model reduces the time between a customer finishing a sentence and the agent starting a reply. The goal is to hit latency targets that keep the conversation fluid. High latency can cause failures that lead to the voicemail abandonment leak where leads are lost to passive recording systems. If the conversation becomes too complex, the agent can hand the call over to a human staff member.

Deploying this does not require an engineering team. A business connects their preferred number, writes the prompt in natural language, and indexes their existing files. The system runs on credits, allowing for scaling based on actual call volume. Every completed conversation is then analyzed to provide insights into what customers are asking and where the agent might need better knowledge or a new tool.