Making AI Chatbots:
Based on our 18+ years of experience for designing and developing conversational AI (AI chatbots), here we share a set of design tips for making practical AI chatbots that can truly help human users and personalize their help in real-world applications.
In case you are not sure what an AI chatbot is or why you need one, check out this comparison on AI chatbots vs. Non-AI chatbots.
Support Tasks and Social Chitchat¶
Effective AI chatbots should be able to accomplish specific tasks and deliver delightful social experience at the same time. Use the tips below to create such an AI chatbot with minimum time and effort.
Start w/ Chat Flow Outline¶
Building an AI chatbot is similar to writing a screenplay, which should have a beginning (Welcome), body, and end (Wrap up). Moreover, it should have a clear conversation objective that can guide the chatbot's to drive a dialog and achieve the objective.
Juji provides a set of chatbot templates, each of which always includes a Welcome and Wrap-up. We strongly recommend that you first create a chat outline, which defines a chat flow that your chatbot will use to guide a conversation with a target audience. You can use your favorite text editor, such as Word and Google Doc to edit such an outline in a plain text format. Online editing tools, like Google doc, allow collaborative editing, which let your teammates or clients help you polish the outline.
Below is a sample outline that is intended to create a chatbot
that can chat with gamers about games.
Here is another sample outline that aims at creating a chatbot that helps make restaurant reservations as well as manage reservations.
Draw Conversation Graph¶
As shown in the above outline, chat topics may be conditioned upon
previous chat topics. For example, topics
T4 follow up to
one branch of
New Booking), while topics
follow the other branch of
Manage Booking). Since Juji
AI chatbots support arbitrarily complex conversations that may include
complex depencies, it is always a good idea to draw the underlying
conversation graph to layout various dependencies. Below is the
corresponding conversation graph representing the restaurant
reservation chatbot mentioned above.
Once the outline is ready, you can then mark each item as a chatbot
message (requiring no user input) or chatbot
user input). Below is the above sample outline with markings
highlighted in yellow. This avoids unnecessary
mistakes during a chatbot making
process. Based on the markings, you can then create a
chatbot and add the marked items in the main chat flow.
Prepare Q&A List and Chitchats¶
As shown next to the conversation graph, one can also define a list of Q&As or social chitchat topics independent of the main chat outline. These Q&As and social chitchats can be invoked anytime during a chat to answer user inquiries or handle user comments falling outside the main chat flow. Not only does this capability deliver a superior user experience, but it also makes a conversation more natural and useful (e.g., providing instantaneous responses to user inquiries). Additionally, Juji AI chatbots automatically tracks and manages a conversation context, including topic switches (e.g., switching from a topic in the main outline to a Q&A or social chitchat). They will always bring users back on track (i.e., the topics in the main outline).
- In particular, we recommend that you prepare answers to three types of user questions that can be anticipated.
No matter how smart your chatbot is, it cannot do everything. To make
a conversation more efficient and transparent, we recommend that you
always prepare a
HELP guide to inform users what your chatbot can
do. This will help users figure out what they can or cannot do with
the chatbot. It will also reduce user frustrations and help the
chatbot better guide a user behavior.
You can go to the Q&A dashboard and directly add an entry with "Help"
Question column, and your help guide in the
and then click
Submit. You can also do so by downloading the CSV file on
the Q&A board, filling in the entry related to
Help in the CSV file,
and then uploading the revised CSV file.
As part of the HELP guide, it is also valuable to customize answers to the following two questions:
User request to chat with a human agent. By default, this question is already in a chatbot's knowledge base. You may want to edit the system default response based on your situation. Just download the CSV file from the Q&A board and edit the answer to this question in the CSV file. Then upload.
Users often ask reciprocal questions, such as
what is your favorite
color when asked the same question by a chatbot. One should
anticipate such user behavior and prepare the chatbot to handle such
Users will enjoy interacting with a chatbot more, if the chatbot can answer simple, "common sense" questions related to the duties of the chatbot. For example, if a chatbot is used to greet online customers of an e-commerce business, it should answer questions about the price and availability of the products. Similarly, if a chatbot is used to onboard customers for an application, it should answer questions about the application.
All the Q&As can be entered in a CSV file or directly in the table on
Q&A Board page. Please refer to Customize
Q&A for more details on how to
Balance Business Goals and User Experience¶
Businesses use chatbots to scale out human-human communications and optimize business outcomes (e.g., improving customer satisfaction while reducing cost). To achieve this goal, it is important to design a chatbot that can balance the accomplishment of business tasks and user experience.
Mix Messages and Requests¶
Juji AI chatbots can send two types of messages (check out chatbot design). One type is a plain chatbot message that ignores user input. The other is a chatbot request that waits for user input and responds to it. If a chatbot sends too many messages that ignore user input, it feels like a monologue (or chatbot spam) instead of a dialog. If a chatbot asks too many questions, it feels like an interrogation instead of conversation. Thus, an AI chatbot should support a balanced asking and answering questions, also known as a mixed-initiative conversation. When writing a chat outline, mix the use of chatbot messages (don't require user responses) and requests (requiring user responses).
In addition, if you use multiple chatbot messages in a row, you may
want to add
delay time (you can set it in topic
settings between them to time each message so they
do not rush out too quickly one after another.
IMPORTANT TIP: If you intend to have your chatbot wait for a user input and respond to it before moving on, make sure you choose
Make a Request. Otherwise, your chatbot simply ignores any user input even if the message is worded like a question. As the example shown below, T6 will not wait for a user's input but T5 will. Note the chat icon appearing on T5, indicating T5 is a "true" question.
Mix Different Types of Requests¶
Juji AI chatbots support several types of requests, e.g., choice-based and free-text requests. While choice-based questions are quick and easy for users to answer, they gather limited information for businesses to act upon. Moreover, choice-based answers can be easily "cheated" (e.g., a user simply makes a random choice without even reading the request). On the other hand, free-text questions especially open-ended questions often elicit rich and meaningful responses, but they take more time and effort for users to respond. In addition, Juji has built-in gibberish detection to prevent gibberish responses.
To make users stay engaged without feeling too tired, it is a good
practice to mix choice-based and free-text requests in a chat. From
our experience, maintaining a 2:3 ratio, 2 choice questions and 3
free-text inquiries, normally works well. We also encourage the use of
open-ended chatbot inquiries (e.g.,
How do you feel about the movie)
instead of simple yes-no questions (e.g.,
Do you like the movie). The
former can elicit in-depth often unanticipated responses and make a
conversation more lively, while latter may end a chat prematurely. If
you have to use a yes-no question (e.g., collecting a definite yes-no
answer), you may want to add a follow-up
Why free-text question to
elicit the rationale behind user responses.
Use Proper Juji Built-in Dialog¶
As mentioned in the design section, Juji provides a rich set of built-in, mini conversations. These built-in dialogs automatically handle highly diverse, potentially complex user expressions. Such dialogs deliver great conversation experience without requiring much customization.
Since Juji uses the request label (see how to give a good label) to find the matched built-in dialog, such match is not always correct. For example, your chatbot is supposed to ask a yes-no question, such as
Would you like to take a test drive?. Juji may not always match such a question with the built-in dialog that handles yes-no question.
We thus strongly recommend that you always check the retrieved built-in dialog to see if it is what you want. If the retrieved built-in dialog is incorrect, use the search button to find a better matched Juji built-in dialog. Using a proper built-in dialog not only supports a better conversation experience, but also reduces your effort of customizing a chatbot.
Optimal Chat Length¶
Although Juji AI chatbots can engage users in a very lengthy conversation (e.g., the longest was 3.5 hours), engaging someone in a conversation requires much time and mental effort. It is thus a good practice to keep a chat at a certain length to keep your audience engaged while completing intended tasks.
From our experience, a chat should be kept below 5 minutes if your chatbot does not have much interesting content or topics to discuss. If your chatbot poses questions and engage users in free-text discussions, making a chat between 10-12 minutes enables your chatbot to bond with your audience but without wearing them out. If your chatbot is intended to conduct lengthy interviews, try to keep it within 45 minutes.
If you need to gather certain amount of information from your audience before your chatbot can help them, dividing a long chat into multiple shorter chats is always more effective. For example, instead of chatting with a user 30 minutes at once, see if you can make your chatbot engage with a user 10 minutes every day for 3 days.
Create Natural and Engaging Conversations¶
The tips listed below help power chatbot to deliver a natural conversation experience that can best engage with target audience.
Paraphrase Messages and Requests¶
To make a chatbot sound more natural, define paraphrases for a chatbot message or a request. For example, if your chatbot says hello to your audience everyday, you want your chatbot to say something differnt to avoid repeatitiveness. This can be easily done by adding paraphrases to the hello message. Below indicates the use of the green "+" to add paraphrases to a chatbot message:
Similarly, a chatbot may need to repeat a question/request if a user does not comply to it. In such a case, you want to add paraphrases of the request, so your chatbot does not repeat the request using the same phrase. Moreover, when giving a request first time, the chatbot should give more information, such as the rationale of the request. When repeating this request, the chatbot however should not repeat everything to sound robotic.
Below is an example showing the initial phrase of a chatbot request. Since this message is long, it would not be used by the chatbot to repeat the question:
Here is a paraphrase (shorter version) of the same message above and will be used by the chatbot to repeat the question if needed:
Below shows another example with a question expression for the first-time use and a list of re-asking expressions. Note the checkbox "Reasking Message". You can use this checkbox to control whether a question expression should be used for asking the question first time or re-asking. If it is unchecked, a question expression is used to ask a user first time when such a question is posed.
If it is checked, it means that the associated expression will ONLY be used when a chatbot re-asks the question but not the first time.
If the box is not checked for any of the question expressions, then every expression can be used for initial asking and re-asking. So it is the best practice to define a list of expressions for re-asking (normally abbreviated, shorter expressions) by checking the box but leave certain expressions (complete, longer expressions) for initial asking.
The screenshot below shows how question paraphrases are used in a chat.
Personalize Messages and Requests¶
Personalized messages make users feel being heard and more willing to be engaged or stay engaged. There are two simple ways to make a chatbot message more personable:
- Address users by their first name occasionally during a chat. The function
(user-first-name)can be inserted into any chatbot message to make a user feel that the chatbot is paying attention to him/her.
Repeat what a user says in a chatbot response. The example below shows that the chatbot repeats the user's word and makes the user feel being heard.
STORE MATCHED INPUTinto a custom attribute (e.g.,
Insert the attribute into chatbot response using a function (e.g.,
Below is how the chat is like:
Echo a user's feelings in a chatbot response. The example below shows how the chatbot acknowledges the user's feelings.
Create a corresponding chatbot response based on the detected sentiment
Here is how the chat is like:
Determine Chatbot Default Response¶
No chatbot is perfect and can understand every user input. To
cope with unrecognizable user input, Juji provides many built-in
dialogs. However, these built-in chatbot responses may not be suitable
for your chatbot applications. For example, if a user asks about the
weather of a location, such as
What's the weather like in San Jose. The default,
built-in chatbot response would be like:
I am not a weather
bot. You may want to check with weather.com. However, if you are
building a weather chatbot, such a response might not be proper.
To override Juji built-in chatbot responses, you can define your own
default chatbot responses to anything it cannot recognize. You can go
Chatbot Settings page (under
Design menu) to define this defaul response:
In the above example, the default response that you entered will then be used instead of Juji built-in default responses.
Periodically Refresh Chatbot¶
Just like a person, your chatbot may not want to engage with anyone in a never-ending conversation. For example, if a chatbot is intended to conduct an interview, it should end after the interview is done. Similarly, if a chatbot is intended to help users in an e-commerce situation, it may want to be refreshed from time to time so users can be helped from the start (like re-entering a store).
By default, a Juji chatbot will be refreshed (restarted) every 60
minutes after it chats with a person. However, this refresh rate may
be different depending on your chatbot tasks. For example, if your
chatbot is conducting an interview that normally lasts for 45 minutes,
we recommend that you set the refresh rate to every 180 minutes
because your audience may not finish their interviews and want to
continue after 60 minutes. They certainly do not want to start it over
again. To set the refresh rate, you can go to the
Ensure Conversation Quality¶
Like in a natural conversation, humans rarely follow a pre-defined conversation flow because humans are creative and spontaneous. To engage users in a quality conversation, a smart chatbot should be able to anticipate user digressions and handle such digressions accordingly.
From millions of human-chatbot conversations, we have identified a number of user digressions, especially when open-ended questions are involved in a conversation. Below is a list of examples.
- Gibberish user input (e.g., "aasfa asfs fa")
- "I don't know" input
- User excuses (e.g., "this question is too hard for me")
- User clarification (e.g., "What do you mean?")
- User asks to alter chat flow (e.g., "skip")
- User asks an irrelevant question
- User comments on the question (e.g., "this is a strange question")
- Thin input (e.g., "good")
Not only must a chatbot respond to each type of user digression properly, but it must also decide how to continue a conversation. For example, if a chatbot asks a non-required question, it should not re-ask the question if a user asks to skip the question.
To help chatbot designers handle diverse user digressions, Juji offers a rich, built-in fallback conversation library that automatically detects many types of user digression and figures out how to handle each type properly. Chatbot designers can leverage the fallback library directly but still have the flexibility to turn on/off specific digression handlers using the chatbot settings as shown below.
Next we use more concrete examples to explain how Juji handles several common types of user digressions to ensure conversation quality.
Handling Gibberish and "I don't know" User Input¶
Users may intentionally or unintentionally test a chatbot by feeding
the chatbot with non-sense or gibberish input. Juji has built-in
functions to automatically detect such input. Moreover, users may take
an easy way out by responding with
I don't know. In such a case, a
chatbot can be configured to accept or not accept such user
responses. For example, if a discussion is about a user's knowledge,
such a response should be acceptable. On the other hand, if the
discussion is about a user's opinion, such a response may not be
acceptable since s/he can always come up with an
You can customize gibberish detection or decide whether to permit user
I don't know response in a topic setting as shown below.
As shown above, you can also control the required length of a user response and also indicate whether a chatbot request is required to answer or not.
Handling User Excuses to Open-Ended Questions¶
Users may also give "excuses" or intentionally dodge a question. Assume that a chatbot asks a user "What's the top challenge you face?". One user may respond "I don't really know since I have many challenges." while another user may state "This is too hard for me to answer."
To simplify a conversation designer's task of anticipating diverse user excuses, Juji has a built-in library that captures a number of common user excuse expressions. During a conversation, Juji automatically detects a user excuse expression and respond to it accordingly. As part of excuse handing, if a question is optional, Juji will also inform a user that the question is optional and the user can skip it if s/he wishes to.
Below shows Juji's handling of several user excuses.
Example 1: Too personal
Example 2: No matched answer
Example 3: Too hard
A chatbot designer can enable/disable Juji's auto excuse handling by going to the "Chatbot Settings" tab, and check on/off the box labeled "Handles a user's various excuses not answering a question" under "Fallback handling".
Handling User Clarification Questions¶
Just like in any conversations, a user might not fully understand a chatbot's question or find the question unclear. When this occurs, the human user may ask a clarification. As shown below, a user asks a clarification question per the chatbot's asking "Could you introduce yourself in 2-3 sentences?"
Here is another example, a chatbot asks "What's the top challenge you face?" A user may ask a clarification question "What kind of challenges are you referring to?" or "what do you mean".
Since it is hard to anticipate what a user's clarification question might be and pre-train a chatbot to handle all possible clarification expressions, one of the easiest ways for a chatbot to handle user clarification questions is to paraphrase a question in multiple ways. This would help a user better understand the question.
Using the first example mentioned above, the chatbot may paraphrase the question as following:
- Could you describe yourself in three key phrases?
- Could you say a bit about the type of work you do?
- Could you say a bit about yourself, e.g., the kind of gamer you are?
Make the paraphrases more specific and the specifics can be determined by the conversation context (e.g., a conversation with job candidates vs. employees vs. gamers). Should the chatbot just start with a more specific question? Our tip would be keeping the initial asking broad because you never know what kind of answers people may come up with. You can always design paraphrases to be more specific to handle user clarification questions.
Juji enables you to paraphrase a question easily. As a chatbot designer, you just need to design a chatbot question with several alternative question expressions (see paraphrase a chatbot question). Juji will then take care of the rest. This is because Juji has a built-in library that captures the most common user clarification questions. During a conversation, if Juji automatically detects a user clarification expression, it will then rephrase the question using an alternative question expression. This way, a user gets a chance to re-interpret the question.
Preview Chatbot Often¶
Since an AI chatbot can exhibit complex conversation behavior, we
strongly recommend that you preview your AI chatbot frequently during
the customization process. This will also help you revert your
customizations if needed before going too far, since
not supported at this moment.
If you need to undo the designs made, you may want to use the
function to clone a chatbot first before undoing the designs. This way
you will always have a copy of what you have made.