What is chatbot design?

A sophisticated chatbot design process within the enterprise context also incorporates business process management and process mining to identify where and how chatbot implementations can improve user experience and business outcomes, mapping specific actions to be taken during or following chatbot interactions.

In chatbot design, as in any other user-oriented design discipline, UI and UX design are two distinct, albeit interconnected, concepts.

UI design refers to how things look: tangible visual elements like layouts, buttons, toggles, colors, text fields and fonts—the aspects of a product, app or website that the user most directly interacts (or “interfaces”) with. Chatbot UI design informs decisions like where a user types text input or the size and location of the chatbot window.

UX design refers to how things work: strategic and logistical concerns like which actions can be taken at each step, what information is provided to or gathered from the user and how the ideal user journey unfolds. Chatbot UX considerations include what questions a chatbot will ask, how it responds to specific inputs or when to escalate cases to a human agent.

In essence, UI design puts UX design in motion. What a chatbot says (and why) is UX design, but how that chatbot dialogue is displayed to users is UI design; the information a chatbot requests at a given step is UX design, but whether users type their answer or select it from a dropdown is UI design.

While the fine details of your own chatbot’s user interface may vary based on the unique nature of your brand, users and use cases, some UI design considerations are fairly universal.

In all contexts, your chatbot UI should be:

• Easy to use: A good chatbot is an intuitive chatbot. It should be obvious where to look, what to click and how to proceed.
• Responsive: Your chatbot UI should offer a consistent experience across all relevant devices, whether on a large desktop monitor or the small screen of a mobile app.
• Engaging: Everything from buttons to animations to copy (or “microcopy”) on labels and dropdowns should be thoughtfully constructed for form and function.
• A reflection of your brand: From colors and fonts to the avatar you assign your chatbot’s persona, your chatbot should treated like an extension of your brand.

For some chatbot implementations, such as integrations into third party messaging apps like Slack, WhatsApp or Facebook Messenger, the conversational interface cannot be customized. Such fixed UI elements should factor into UX planning.

For many businesses, especially those without resources to develop a bespoke UI from the ground up, it’s most efficient to use a chatbot builder with templates and drag-and-drop workflows that streamline UI decisions. Leading chatbot providers offer opportunities to customize stylistic elements to suit your branding, but adhering to proven UI design patterns lets you focus on your organization’s unique UX priorities.

In further exploring chatbot UX design, we’ll use certain terms with specific meanings in this context.

• Intent: Intent is a user’s purpose for interacting with your chatbot—the specific goal or problem to address, like paying a bill or getting a question answered.
• Utterance: An utterance is any individual statement made during an exchange, like “hello!” or “I’d like to pay my bill” or “yes.”
• Exchange: An exchange consists of two or more utterances. Essentially, the term refers holistically to the back-and-forth conversation between user and chatbot.
• Contact: Each instance of a user engaging a chatbot is a contact. Contact is not entirely coterminous with exchange: if a user initiates a contact by opening the chatbot window but doesn’t respond to the chatbot’s greeting, no exchange has occurred; if a user restarts a conversation, that contact now includes multiple exchanges.
• Domain: Domain is a broad description of your chatbot’s scope, like customer support or human resources. Each domain comprises multiple topics. Chatbots designed to converse on any subject whatsoever, like ChatGPT, are called open-domain chatbots.
• Topic: A topic is a specific subject matter or set of tasks within a domain. For example, the domain of customer support might contain topics like bill payment, store hours or returns. Each topic maps to specific user intents.
• Entity: An entity is a noun relevant to user intent—like a product, document, or service—mentioned in an utterance. Proper identification of entities is an important element of natural language processing (NLP).
• Escalation: Escalation is a handoff from chatbot to human agent. This can be a “planned” escalation, or a “fallback” escalation (when a bot cannot recognize or resolve the user’s intent).
• Step: Each back-and-forth interaction between chatbot and user is a step. Steps might include greetings, clarification questions, actions, handoffs, or even small talk.
• Flow logic: Flow logic governs how bots react to each utterance and proceed to the next step. This may be involve simple if-else statements, decision trees, complex algorithms or machine learning-driven probabilistic logic.
• Preferred response: An utterance that positively moves the exchange toward resolving user intent.
• Non-preferred response: An utterance that does not progress the exchange toward intent resolution.

A great chatbot experience requires deep understanding of what end users need and which of those needs are best addressed with a conversational experience. Employ chatbots not just because you can, but because you’re confident a chatbot will provide the best possible user experience.

Your FAQs are an excellent knowledge bases for queries, tasks and problems that surface frequently and predictably. Your customer service teams are likewise an important source of insight. Robust business process management can further identify opportunities and inefficiencies, as well as help delineate the different knowledge centers, communication channels and levels of complexity, security and privacy germane to each domain.

Chatbots offer the most value when two-way conversation is needed or when a bot can accomplish something faster, more easily or more often than traditional means. Some domains might be better served by help articles or setup wizards. Others, like those requiring highly technical assistance or sensitive personal information, might be better left to a real person.

For your first chatbot, it’s wise to walk before trying to run. The less data you have, the less confidently you can make predictions: companies that spend months building an inaugural chatbot spanning many topics often learn (after launching) that key assumptions about user behavior were wrong—and have to practically start again from scratch. Effectively addressing a shorter list of topics and intents yields a better user experience than providing inconsistent results across a wider domain.

Having said that, choose a domain with growth potential. Truly successful chatbot strategy yields not standalone solutions, but conversational tools deployed across all relevant channels—websites, messaging apps, phone systems—that enrich each other by generating shared data for training and optimization.

Broadly speaking, chatbot offerings fall into two categories: rule-based chatbots and AI chatbots.

Rule-based chatbots are simple and economical. They operate on if-then-else rules: each step (or branch in a decision tree) is assigned specific inputs the chatbot can recognize, each matched to a scripted response. Lacking natural language processing (NLP), rule-based bots must restrict user utterances to simple phrases or pre-written options. This may limit success unless your users’ needs are highly predictable, repetitive and straightforward—and will stay that way as you scale up.

AI chatbots are more robust, versatile and scalable. Artificial intelligence capabilities like conversational AI empower such chatbots to interpret unique utterances from users and accurately identify user intent therein. Machine learning can supplement or replace rules-based programming, learning over time which utterances are most likely to yield preferred responses. Generative AI, trained on past and sample utterances, can author bot responses in real time. Virtual agents are AI chatbots capable of robotic process automation (RPA), further enhancing their utility.

Many situations benefit from a hybrid approach, and most AI bots are also capable of rule-based programming.

Before designing the fine details of your customer experience, plan the foundation of your chatbot.

• Determine starting channels: The channels where users interact with your bot should naturally align with functions it serves. Each different channel affects how users articulate themselves, how your bot should respond, and which systems are available for integration. Once you’ve picked a domain, make sure you can work with relevant channels. For example, don’t automate payment questions if the team running the payments page of your website won’t let you add the chatbot client to it.

• Identify primary topics: The fundamental goal in chatbot planning is to determine your bot’s “minimum viable knowledge,” or MVK. MVK is the minimum topic coverage the bot must be capable of in order to fulfill its purpose. This entails both topic breadth—the full range of different topics to cover—and topic depth: how thoroughly each topic must be covered. Start with topic breadth: list all the possible topics relevant to the chosen domain, then prioritize. Focused topic breadth facilitates greater topic depth, increasing your chance of success.

• Aggregate all relevant knowledge bases: Your bot must be able to access any info necessary to understand and address user intents that fall within its scope. That info might not be all in one place: it’s often spread out across disparate sources like webpages, databases, documents, FAQs, CRM platforms and online transactional processing (OLTP) systems. Intelligent search is the ideal way to aggregate all relevant data sources and streamline information retrieval.

Users unconsciously, automatically infer a character behind your bot. It should convey the positive characteristics we seek in human conversation—empathy, curiosity, patience, affability—while maintaining the transparency of being a robot. The latter is essential to both managing user expectations and avoiding the “uncanny valley” effect: the strange uneasiness provoked by humanoid things that are not-quite-right. This can be most easily achieved by thoughtful choice of name, avatar and greeting.

Your chatbot’s personality impacts most elements of conversation design. It should reflect your brand and be appropriate for its intended users and function: a fitness assistant bot should use active language; a healthcare diagnosis app should avoid jokes.

Start by considering where your chatbot falls on various spectrums:

• Funny vs. Serious
• Enthusiastic vs. Calm
• Formal vs. Informal
• Warm vs. Cool
• Authoritative vs. Relatable

A chatbot provides only half of a conversation. You can’t control or fully predict the user’s half. Strong conversation design ensures a positive user experience by approaching conversation flow in a way that, no matter the user’s utterance, the chatbot’s response feels natural, believable and productive.

True coverage of a topic requires not only designing ideal conversational paths, but envisioning all unique paths a conversation might follow, including potential confusions, detours and dead ends. You might program your scheduling bot to recognize “I want to change my appointment,” but a user might say, “I can’t do Tuesday anymore.” You might have an optimal path, but is there a plan B if plan A fails? If plan B fails, can your bot explain the problem to the user? If the user doesn’t understand a request, can the bot phrase the utterance differently?

Even if your flow logic is flawless, mistakes happen—but minor imperfections should not derail an exchange. Here again, AI chatbots have a major advantage: instead of manually predicting and planning for every single typo to avoid interruptions, artificial intelligence can make educated assumptions and keep things moving. For example, IBM watsonx Assistant features autocorrect for mispellings, as well as fuzzy logic to aid recognition of intents and entities. Likewise, AI bots with speech-to-text can be trained to properly interpret accents, mispronunciations and jargon in voice inputs.

As in regular human-human conversation, users want to feel understood. Chatbot design can achieve this by ensuring that all bot responses, even non-preferred responses, are informative and relevant to the user’s utterance. When copywriting chatbot dialogue, aim to acknowledge what the user has said and avoid blunt changes of subject, random leaps in conversation, or “forgetting” information the user provided earlier in the contact.

Chatbots have limitations. The capacity to fail elegantly and provide routes to repair the conversation is essential: it’s okay for a bot to be wrong, but being wrong and irrelevant may doom the exchange and deplete trust in the chatbot. Bots must be designed to gracefully handle harassment, recognize nonsense or irrelevant utterances, react to topic shifts and get the conversation back on track.

Always decrease the user’s burden.

• Ask for less: Instead of a long order number, would the last four digits suffice? Would their name and an additional piece of simple information—say, the order date—eliminate the need for the number altogether?
• Clear choices: Word every question carefully. A user might reply to “would you like a Wednesday or Thursday appointment?” with “yes”—a non-preferred response. But there are only two responses to “please choose Wednesday or Thursday.”
• Buttons: When there is a small list of potential options, choosing a pre-written option—or, for phone systems, a keypad number—makes things easier and eliminates the possibility of a non-preferred response.

Clear, concise copy reduces friction and demonstrates respect for the user’s time. Reconsider your conversation flow if it requires lengthy instructions.

Effective chatbot design involves a continuous cycle of testing, deployment and improvement. Individuals may behave unpredictably, but analyzing data from past contacts can reveal broken flows and opportunities to improve and expand your conversation design.

Chatbots rely on, generate, and analyze a great deal of user data. That data must be handled carefully. Failure to do so has not only ethical consequences, but potentially legal and financial consequences.

It may also impact adoption of your chatbot: per Pew Research¹, more than half of Americans have decided against using a product over concerns about how (and how much) it collects personal data.

• Users’ data should be protected from theft, misuse, or data corruption.
• Policies should comply with EU’s General Data Protection Regulation (link resides outside ibm.com) (and comparable regulations like CCPA).
• Privacy settings and permissions should be clear, findable, and adjustable.
• Users should be apprised and in control of what data is being used, and in what context.