Conversational design style and tone of voice

Conversational Design

To ensure that our system is not only efficient and safe in providing accurate responses it needs to have a helpful, empathetic personality that makes humans want to interact with it. To achieve this, the interactions between the patient and our system should mirror real-world conversations.

We felt it was important to design a system that was polite, but not too verbose or too terse. This is the difference between having a conversation with someone who replies in just “Yes” or “No” statements and another one with someone who goes on a long monologue with unnecessary detail. We needed to find the happy medium for our system to follow to continue the conversation with a user and to keep them engaged.

Introductions

A pleasant introductory question is good but it also makes a big difference to make it clear to the user what purpose the virtual assistant serves. For example “How can I help you is good” but is vague and doesn’t give enough information about how the GP AI Receptionist can help. Our introduction “Hello my name is AiMe, I am a Virtual GP Receptionist. You can speak to a human at any time. How can I help you” is pleasant and informative leaving the door open for the patient to pursue their request further. If the virtual assistant is unfamiliar with the user’s choice or it fails to understand the intent, we designed fallback messages to help the patient. A simple “I’m an AI GP Receptionist and I’m afraid I can’t answer medical questions. Let me transfer you to a live person” helps to close the conversation politely in this instance.

Natural Multi-turn Conversations

If you think about the conversations we have with other humans every day, you will find that there are some quirks at play aside from the words and sounds we utter. The vocal tone, tilt of the head and body language are important non-verbal cues that supplement the words we speak and provide extra context.

The idea of turn-taking is a good example of this. When we engage in conversations, we often adjust our pitch or turn our heads to let others know that it is their turn to speak. Think of the tone you use to end a question and wait for an answer. How do we apply this trait to virtual assistants when there are no visual or other non-verbal queues available from the user?

This is where ending prompts with a question makes a big difference in moving the conversation forward. Consider the following exchange.

Patient: I’d like to make an appointment.

Practice:

Ending the reply with [give example] will prompt the user to continue the conversation and give the information needed.

Implicit and Explicit Confirmations

An explicit confirmation occurs when the AI GP Receptionist asks the patient to confirm by repeating parts of the query explicitly. It is useful for situations when the virtual assistant’s confidence in recognising the intent of the patient isn’t clear or in instances where clarification of the facts is critical in progressing a request. For example, if the task involves asking for the patient ID it helps for the virtual assistant to make certain that the date of birth information is correct by checking again to make sure there isn’t any confusion. If we think about the following exchange:

Caller: I’d like to make an appointment for my grandmother
System Can you confirm the appointment is for your grandmother?
Caller: Yes, I can confirm that the appointment is for my grandmother.

This clarifies that the appointment is for the caller’s grandmother rather than the caller. It represents a specific confirmation step to which the caller must respond to move forward toward task completion. The key advantage of explicit confirmation is its clarity – callers know how to respond to the prompt whether or not the system recognised all inputs correctly.

An implicit confirmation does not require a confirmation from the user, but also leaves the option open for the user to confirm or deny. It makes the conversation a lot more natural, and closer to how humans talk with each other. The key advantage of implicit over explicit confirmation is that when it’s right, it’s faster. When it’s wrong, however, it isn’t always clear to callers what they need to do to recover from the error.

When it’s wrong, however, it isn’t always clear to callers what they need to do to recover from the error. That’s why it’s important to use implicit confirmation only when it has a high likelihood of being right.

To address this weakness, we endeavour to include a quick instruction as part of the implicit confirmation prompting, for example:

System: Thank you for calling The XYZ Surgery, How can I help you?
Caller: I’d like to make an appointment for my grandmother.
System: In order to make an appointment for your grandmother I need to ask your date of birth, your grandmothers’ date and birth and the reason for your grandmothers’ appointment.

Including this type of message to overcome the weakness of implicit confirmation greatly reduces its speed advantage, but maintains the possibility of allowing the caller to proceed without requiring explicit confirmation.

Empathy

Empathy is about showing that you understand the user’s situation. Virtual assistants can never be as empathetic as humans but their conversations can be designed in such a way that users feel that they are heard and empathised with. Words and phrases that show apology and gratitude are useful ways to display empathy.

Caller: I’d like to make an appointment
System: To make an appointment for you Can you confirm the appointment is for your grandmother?
Caller: Yes, I can confirm that the appointment is for my grandmoth