The Gap Between Answering a Question and Solving a Ticket

A customer arrives on your website at midnight because they have a problem that requires an immediate change to their account. They do not want to read a help center article. They want a specific outcome, like a refund or an upgrade. They will leave if they are told to wait until Monday morning.

Most automated systems fail here because they are built only to retrieve information. They can summarize a document or crawl a website to explain a policy, but they cannot execute a command. This limitation represents the tool call threshold where an agent stops being useful and starts becoming a bottleneck.

Moving past this ceiling requires giving the agent access to the same tools a human representative uses. In Duvi, this happens through tool calls to HTTP APIs or systems like Salesforce. Instead of just confirming that an order exists, the agent hits an endpoint to retrieve the shipping status. If the business logic allows, it triggers a return label immediately.

The same logic must apply across every channel. A customer might start an interaction on a web widget and then continue it over a phone call. If the agent on the phone has access to the same knowledge base and the same API tools, the transition is invisible to the user. This consistency prevents the action gap crisis where tasks remain unfinished because data is trapped in separate silos.

Voice adds a layer of technical complexity that text does not. A person on a phone call will not wait five seconds for a model to think. Duvi addresses this by allowing builders to choose specific transcribers and models based on latency estimates. When an agent calls an external API during a live conversation, the speed of that external system becomes the bottleneck. Engineering the agent to handle these wait times with verbal fillers or status updates prevents the user from hanging up. This approach helps resolve the phone line bottleneck that typically slows down resolution times.

Control over the agent behavior comes from a system prompt structured into specific blocks. One block defines the role. Another defines the constraints. Keeping the core facts in the knowledge section prevents the system prompt from becoming a list of rules. This separation ensures the agent stays focused. It also ensures that your knowledge base remains the primary source of truth for the AI.

There are still moments where an AI should step back. If a customer becomes frustrated or a request falls outside the defined API capabilities, the conversation needs to move to a person. Duvi manages this through a handover system that transfers the live chat to a human agent. For voice, the system records the call and provides analysis so the follow up starts with full context. This prevents the friction of repeating information.

Deploying these agents does not require an engineering team to write custom logic for every new conversation. The agent learns from existing material and acts through existing infrastructure. Usage is managed through credits. This setup allows a business to scale automation based on actual volume. The focus moves from managing a queue to refining the API tools the agent uses for ticket resolution.