A chatbot that makes someone repeat a claim-status request, answer unnecessary questions, and then call support has not saved them time. Conversational AI should help people reach a useful answer or complete a task through written or spoken natural language—with as little friction as possible. section 1.1 section 1.1.1
The assistant’s job shapes the conversation:
- A question-answering bot responds directly, often without a follow-up. section 1.1
- A process-oriented bot guides a user through steps such as checking a balance or booking an appointment; it may complete the transaction or collect information for someone else to act on. section 1.1
- A routing agent identifies where to send the user, whether to a specialist bot or a person. One assistant can combine all three jobs. section 1.1
A practical mental model is: determine what the user wants, gather only the missing information, then fulfill the request. The user’s utterance is what they say; an intent is the goal it expresses. For example, a request about being locked out may mean “reset password” even if it never uses those words. A classifier can identify that intent; a dialogue engine then manages follow-up questions, conversation state, and any needed connection to other systems. section 1.1.2 section 1.1.3
Figure 1.3 Flow diagram for conversational AI. In many use cases, “additional information” includes user profile data.
The flow diagram shows where understanding the request, gathering additional information—including possible user-profile data—and fulfilling the request fit together. section 1.1.2
Ask less when context already supplies the answer
For a claim-status request, a bot could make the user navigate several questions before providing any value. If the user is logged in, it can instead use that context and the reasonable assumption that they mean their own claim to reach an answer sooner. Context does not remove the need to clarify when an assumption is uncertain; it helps avoid asking for information the system already has. section 1.1.1 section 1.1.2
Design starts with users’ goals: what they need to do, what information they have, and how they are likely to ask. Keep responses concise, especially in voice interactions where users cannot scan ahead. Handle misunderstandings and offer a route to a human rather than trapping someone in repeated rephrasing. Fulfillment must also respect legal and ethical boundaries, including secure handling of sensitive information and not pretending the assistant is human. section 1.1.3 section 9.2.4 section 2.1.1 section 1.1.2
Generative AI can help draft clearer messages, improve intent handling, or answer questions using retrieved information. It can also misunderstand users or produce wrong answers, so it complements rather than automatically replaces traditional techniques. Improve the whole experience over time: track where users disengage, are misunderstood, or fail to get what they need, then check whether changes improve outcomes such as time to resolution, satisfaction, and calls completed without human involvement. section 1.2 section 1.3.1 section 1.3.3
