7 Steps to Well-crafted Dialogue for Conversational AI
Updated: Oct 8
The business case for prioritizing dialogue
The dialogue a system presents is perhaps the most important element for success. It guides input for ML language models and recovers the interaction if input is misunderstood. Unfortunately, dialogue creation is frequently underestimated, under-crafted, and rushed during Conversational AI projects. Teams often treat dialogue creation as a distant second to models and coding when prioritizing activity.
Low-quality dialogue can doom a system's success faster than any other element. Mediocrity or worse when it comes to dialog will lead to confusion about where the performance problems are and where the opportunities for improvement exist.
The key to achieving your business goals
After over 20 years of deploying Conversational AI, we can tell you that well-designed dialogue is the key to achieving your business goals. A great language model can't overcome dialogue that causes problems, but great dialogue will:
Make the interaction clear
Improve input data quality
Be the most direct path to good performance from the get-go
Enable recovery from lackluster model output
Crafting dialogue is a challenging skill that requires a deep appreciation of language and what humans need to interact easily with AI systems. As you might expect, the emphasis we place on this raises a question: "If it matters so much, how do we learn to create effective dialogue and how do we know when it’s good?"
It turns out we have an answer!
There are several important steps in the conversation design process to achieve well-crafted and effective text, voice, and visual output. We built the steps by combining robust design practices with a lovely writing framework developed for the television series "Mr. Rogers' Neighborhood" and documented in this article from The Atlantic.
Below is our set of practices, which use that framework. With it, you will generate and revise lines of dialogue using multiple iterations. Each step has an instruction and then examples for voice and chat interaction. We will cover visuals such as buttons and images in a separate post. As you will see, this can be time- and focus-intensive. It isn’t meant for every single piece of dialogue, but certainly for the key elements such as intros, task instructions, and delivery of important information. Successful and seasoned designers do this by habit, so as you begin to practice, press on even when it feels long and repetitive.
Follow these steps to design key dialogue
1. Don't worry about the exact wording, order, or length of what you write in this step. Consider this as if you are sketching visual ideas. Refinement and editing will come later. Capture V1 of the dialogue making sure to include the important details and necessary functions. Be thorough about the context of use, the task step, and any system requirements. Use clear and precise language and connect the piece of dialogue conceptually to what has already happened and what will happen in the future.
Voice: To complete the transfer transaction, you must enter all ten digits of the account number because the interactive voice application needs to access your accounts. Use your voice to speak the ten digits clearly or use the buttons on the phone keypad to type it.
Chat: To complete the transfer transaction, you must enter all ten digits of the account number, because the chat application needs to access your accounts. Type the ten digits in using your keyboard.
2. Replace any “machine speak” or jargon with clear and plain language. This reduces the risk of misunderstanding due to unfamiliar terms. Don't be tempted to look at 1 and 2 and think 'no one would write these' because people do sometimes. Clarify the language and ensure the dialogue is friendly to all your potential users.
Voice: To perform the transfer, the system needs the ten digits of your account number to look up your account, so please enter them using your voice to speak naturally or by typing them in.
Chat: To perform the transfer, the system needs the ten digits of your account number to look up your account, so please enter them.
3. Read what you have so far out loud. Would a person speak this way? Remove anything that could be interpreted in a way that seems like it’s a machine telling a human what to do. This creates the feeling that the system is serving the person instead of the other way around.
Voice: So the system can look up your account and perform the transfer, would you please say or type in the ten digits of your account number?
Chat: So the system can look up your account and perform the transfer, would you please type in the ten digits of your account number?
4. In this revision, add priming and motivation followed by how they will do it: "Why" + "Action" + "Method". Use word order and structures that are more like everyday human speech. This makes the interaction more natural and less formal.
Voice: To pull up your account and make that transfer, can I get your ten-digit account number? You can use your voice to say it or just type it.
Chat: To pull up your account and make that transfer, can I get your ten-digit account number? You can just type it.
5. Now it's time to remove extraneous words and add natural contractions. Rely on context, implicature, and anaphora to simplify and lighten the weight of the interaction. Make it feel nice.
Voice: To do that, what’s your ten-digit account number? Just say it or type it.
Chat: To do that, what’s your ten-digit account number?
NOTE: It's very possible that after Step 5 you will think you've created dialogue that is ready to be used in your design. However, reflect on the emotional elements in 6 and 7. They may enhance the customer experience and even add to success. Notice that the emotional elements might override a strict focus on brevity. Balance is good.
6. Earn trust. Enable feelings of connection, prediction, and confidence. It's often assumed that customers will understand that the system was designed to provide assistance, but due to the reputation of many software applications, you should remind your users the system is here to help.
Voice: I can help you with that. What’s your ten-digit account number?
Chat: I can help you with that. What’s your ten-digit account number?
7. Now point to a positive outcome, such as aiming for success and respect for the customer's time.
Voice: I can take care of that for you. It'll be quick. Just tell me your ten-digit account number.
Chat: I can take care of that for you. It'll be quick. Just type in your ten-digit account number.
Conversational designers who put in the effort to craft superior dialogue will see fewer abandons and higher resolution. If you answer yes to any of the following questions, you'll know you should bookmark this page for the future:
Does the business performance of your application depend on well-crafted dialogue?
Are you trying to create a sophisticated multi-turn AI application?
Are you seeing mediocre performance?
Are you struggling to break out of old-school IVR dialogue that sounds stilted and robotic?
Looking to dive deeper with your team on voice, chat, or multimodal dialogue design? CCAI offers a 6 hour workshop for teams looking to accelerate conversational AI.