In a broad sense, artificial intelligence (AI) uses computers and machines to simulate human decision-making and thinking. More modern definitions of AI describe it as the ability of a machine to generalize its knowledge and skills to new environments and to efficiently learn new skills or knowledge. Some current applications of AI include online shopping, facial recognition, speech recognition, and autonomous vehicles. This article will focus on conversational AI and the user interface considerations specifically for designing chatbots. A chatbot is an application of AI that simulates a conversation with a user using natural language processing through either text or voice communication. A digital or virtual assistant is a more complex form of a chatbot that can also complete tasks for the user.
AI and Popular Culture
The first thing that comes to mind when people hear the terms artificial intelligence or AI is often related to what they have seen in movies or read in fictional novels, such as the loyal and helpful droids R2D2 and C3PO from Star Wars or the sinister cyborgs from The Terminator. Although AI is portrayed accurately in some popular culture, many movies and books distort our reality about what AI is and its capabilities. These include incorrect assumptions that AI is so advanced that it can do anything a human can or that AI can act autonomously. Depending on which portrayals of AI we adopt can lead us to experience either positive or negative feelings toward this technology and can set unrealistic expectations about what it can currently accomplish.
Historically, a machine was an AI if it could perform a task that previously required human intelligence. This definition, however, was not constrained by how human beings might solve complex problems (for example, we do not consider 100 million possible moves simultaneously) and did not factor in human learning. The early techniques of AI included hard coded-algorithms or fixed rule-based systems. In the article “What Is AI? Here’s Everything You Need to Know about Artificial Intelligence,” Nick Heath suggests that more modern definitions of AI describe it as the ability of a machine to generalize its knowledge and skills to new environments and efficiently gain new skills or knowledge.
Most modern applications, including chatbots and digital assistants (considered AIs by many), fit this narrow definition of the ability to generalize their training to a limited set of tasks, such as understanding speech and recommending products for purchase based on previous purchases. The concept of machine learning, a more recent development in AI, neatly fits this more modern definition of AI. With machine learning, algorithms are trained using large amounts of data without relying on explicit rule-based programming. The algorithms identify patterns to handle more complex problems, such as image recognition or predicting future stock prices. Heath also discusses artificial general intelligence (AGI), whereby a machine can learn and execute a wide variety of tasks and can reason about various topics similar to human ability. This form of intelligence is commonly portrayed in popular culture even though many experts do not believe it yet exists.
Role of the Conversation Designer
The meteoric rise in chatbot use over the last decade has created a new breed of UX practitioner: the conversation designer. Conversation design is equal parts content writing and interaction design. AI has not yet reached a level of maturity where it can spontaneously create chatbot dialogue. Instead, content designers conduct research to
- understand the domain the chatbot is expected to cover,
- determine the purpose of the chatbot, including why users will be interacting with it, and
- create the actual chatbot dialogue so that it is accurate and consistent with the tone and voice of the chatbot.
Anatomy of a Chatbot
Conversation designers create three types of information for a chatbot:
- intents
- entities
- dialogue
Intents represent the actions that users want to take by identifying user intentions or goals. The scope of information that all the intents together cover is known as the knowledge corpus. For each user intent identified, a list of example utterances is generated to represent common ways the user could state their intention. The chatbot will train on the utterances to learn what requests are considered equivalent. For example, if the user intent is “I want to place an order for pizza,” then utterances could include
- I want to order pizza.
- I’d like a pizza.
- I’ll take a pizza.
- I need a pizza.
- I’m ordering a pizza.
Entities are the nouns in the user examples or what the intents will act upon. In our example, the user wants to order a pizza, but there are different terms for pizza (e.g., zaw, pie, deep dish). One approach is to create more examples that use each of these terms so that the chatbot can learn each identified food that users can order. Alternatively, we can create an entity (@food) and add a value (pizza) with the synonyms we identified (pie, zaw, deep dish), and then replace pizza in our user examples with @pizza to represent all variations of the term the chatbot should take into consideration. Likewise, we could add values and synonyms for other types of food the user could order, such as salads, appetizers, and desserts, which would allow us to have a single intent for ordering food:
- I want to order @food.
- I’d like a @food.
- I’ll take a @food.
- I need a @food.
- I’m ordering a @food.
Dialogue is what your users will ultimately see and interact with based on the AI’s interpretation of user goals. Given the user’s input, how should the chatbot respond when a specified pattern of intents and/or entities is recognized? Conversely, how should the chatbot respond when it does not understand the user’s input? For example, the chatbot could offer options that help move the conversation along rather than simply, “I don’t understand.”
Conversation Design: Considerations
There are a lot of decisions that need to be made and actions taken before a single word of dialogue is written:
- Determine the chatbot’s purpose.
- Conduct research to understand the domain.
- Understand the user goals for interacting with the chatbot.
- Identify the intents and entities.
- Select the tone and voice of the chatbot.
- Map the conversational branches in a flow.
- Write dialogue.
Identifying the chatbot’s purpose is critical to understanding if a chatbot is the best solution to the understood problem.
Research is critical in chatbot design. Unless the conversation designer is a subject matter expert (SME), the designer will need to talk to the SMEs and evaluate any available information to determine what user goals the chatbot should handle versus a human agent. Research with users is necessary to understand their expectations for the chatbot. Determining what users want to do and how they might state their goals are key to building a successful dialogue interaction.
Before writing dialogue, consider the tone and voice of the chatbot, ensuring it is consistent with the existing brand and other available materials. For example, if your company sells pet toys, then having a happy dog personality for your chatbot may very well match your brand. But if the chatbot is for a bank, a dog-like chatbot might feel off or out of place.
If the conversation flow can include branching, it is helpful to map out the dialogue for an intent using a flow diagram so that the user does not accidentally end up in an unintentional dead end or loop with the chatbot.
Conversation Design: Best Practices
The goal of conversation design is to create an experience that feels natural while giving proper attention to grammar, spelling, and formatting to make the text easy to read:
- Use contractions.
- Avoid “yes” or “no” responses from the chatbot.
- Move the conversation forward with each response.
- Attempt to include the solution in the response.
- Provide links to videos or more detailed textual explanations, as needed.
- Give users buttons or text links to clarify options and reduce misunderstandings.
- Update intents, entities, and dialogue to reflect the dynamic nature of content.
Interaction Design: Best Practices
In “Designing for AI—A UX Approach,” Marielle Lexow mentions that in addition to considering what the chatbot will say, conversation designers need to also consider the overall interaction with the user:
- Set user expectations early.
- Use a common language.
- Enable users to have flexibility in interaction.
- Design to create trust, transparency, and explanation.
- Enable users to have control and provide feedback.
- Test your chatbot to ensure it’s working as designed.
Chatbot Design: Real-World Examples
We incorporated many of these best practices in our designs of a voice-enabled chatbot for technical support and a text-based chatbot for employee-related questions.
One key to success is to set expectations regarding how the AI can assist and how the user can interact with it from the start of the interaction. In designing the voice-enabled chatbot, we communicated to the user from the start the specific technical issues it could address. It was also essential to explain early on how the user could interact with and navigate in the application by noting basic voice commands (e.g., main menu, repeat, and agent to escape the AI and speak to a human agent). By setting user expectations, we saw increased user satisfaction and positive feedback, plus fewer user errors.
A voice-enabled chatbot should also mimic human conversational speech patterns by pausing between sentences when it picks up speech and allowing the user to interrupt with a spoken command or intent. A successful conversation with a person or a machine is dependent on a common language that both parties understand. Sajid Saiyed, in “Design Considerations for Conversational UX,” notes that the voice-enabled chatbot should learn our language and understand our intents, not vice-versa. Unfortunately, in our voice-enabled chatbot, user utterances were not always correctly understood, which required some users to repeat their intents, eventually routing them back to the main menu, leading to increased frustration and decreased satisfaction.
When users have a higher expectation of AI applications than is warranted by the current level of technology, they often experience disappointment if their expectations go unrealized. Because AI systems use specific algorithms and models to analyze and interpret data autonomously, it is critical that users feel in control and develop a certain level of trust when using AI applications.
In the article “UX of AI,” Lennart Ziburski mentions that one way to facilitate trust in AI is to make users aware of how the AI system came to a decision or recommendation rather than acting as a black box experience. Users must be able to trust the provided answers and understand how they were determined. In addition, when users provide sensitive personal information (e.g., user IDs, passwords), they need to trust the AI to keep the information confidential. This trust extends to any navigational links provided by your AI.
In another chatbot created for employee-related questions, many internal links identified the information source, which helped users determine if they had already visited that link. Some of the links, however, were external to the company and were not recognized by users, making them feel less trusting of where the links would take them.
Chatbots are living applications that require ongoing maintenance to thrive and consistently provide a positive user experience. Part of this process is having the right tools available to gather user feedback, particularly for questions or intents where the AI was incorrect. Even better, the tool should automatically capture user behavior to determine which intents are working well and which ones are not. The chatbot application we designed did collect basic user sentiment about its responses in the form of thumbs up and thumbs down, but this required people to analyze the data and take manual corrective action. However, in a beta version of the chatbot, we collected a complex set of field metrics for each intent that the team could examine for potential improvements.
Allowing for personalization of the content by users provided them with a greater sense of control and let them ensure the answers and interaction met their needs. In the voice-enabled chatbot we developed for technical support, users had the option to allow the AI to intervene early on if it had a solution or wait until after they had proceeded through the menu options. Users could choose to have their answers delivered via a voice response on the phone or by text in an email. For the text-enabled chatbot, users could choose to either type their answer or use the provided links. Providing users with more personalization and control resulted in more positive feedback and greater satisfaction with the chatbots.
UX Considerations to Think About Before You Build
When many people think of chatbots, a 100% conversational interface comes to mind. The chatbot gives a greeting and asks what the user wants, then the user responds in a text format. Conversation then continues in turns back and forth until the user either gets an answer or gives up in frustration. Such an open-ended interaction assumes a mature chatbot with a well-defined knowledge corpus. If your chatbot, however, is still in the building stage or not meeting user expectations, you should consider some other approaches.
Consider adding menu-driven prompts. If your chatbot can only respond to a narrow set of topics, embrace a more closed-ended approach by leveraging menus in place of natural language. Your users should not have to guess what the chatbot is good for—spell it out. Your dependency on this approach can diminish as your knowledge corpus improves.
Leverage personalization. If someone signs into a chatbot, only offer options that apply to that user based on their account details. It is a frustrating experience when the user needs to supply the same information multiple times or information that the chatbot can access from their records.
Consider alternative solutions. While chatbots are ubiquitous and every business seems to want one, consider if a chatbot is the most appropriate interaction to meet your users’ goals and needs. Sometimes a search option or an FAQ is the best format instead of a chatbot. Users want the information quickly and may not be impressed with newfangled technology.
Chatbots perform better when
- the domain area is limited in scope,
- user goals are known and well-defined, and
- machine learning or rule-based queries are used to improve the discoverability of high-value information.
If your chatbot does not meet these criteria, consider integrating your chatbot into your existing search experience. As noted in Temple et al.’s paper, “Not Your Average Chatbot: Using Cognitive Intercept to Improve Information Discovery,” cognitive intercept is a concept created at IBM whereby the chatbot runs in the background as the user searches in a standard-looking search box. If the chatbot has a high confidence match to the query, it displays its match in addition to the search results. Otherwise, it remains in the background, and the user continues with a standard search. This combined approach allows users to use a single, shared interface with lower cognitive overhead while tempering expectations for your chatbot. You can only make a first impression once.
At every step of the chatbot creation process, be curious. Ask if a decision adds value to the process and if it helps to meet the goals of the business or user. If the answer to any of these is no, then re-evaluate your approach.
Conclusion
Chatbots have found their way into our homes to help with home automation and into our daily lives as information devices that can answer questions or solve real problems. While machine learning will continue to improve all aspects of the chatbot experience, UX researchers and designers will continue to have critical roles in delivering an experience that anticipates user intents and responds in a conversational style that feels familiar.
Jason Telner, PhD, is a senior user experience researcher within IBM’s CIO design team. Jason has over a decade of UX research and design experience. He has several years of experience working specifically on conversational AI applications including digital assistants and voice interfaces specifically for customer support. Twitter: @jtelner
Dabby Phipps is a UX Designer/Researcher for CIO Design at IBM. She’s worked in various domains including healthcare, military, sales, IT support, and chatbots. Dabby has many years of experience conducting research in support of chatbots and has spent several years writing cognitive intents for an internal IT support chatbot. Twitter: @lbw8lrrh
Jon Temple, PhD, is a Design Principal at IBM with over 25 years of experience in user-centered design and research. His work includes several years of integrating conversational chatbots into other solutions such as search and online chat. Twitter: @jongtemple