Skip to main contentIBM Design for AI

Conversation overview

Now that machines, powered by analytics and cognitive capabilities, can technically understand, reason, learn and interact using natural language, what will they say when designers give them a voice?


Whereas traditional speech recognition systems understand what people say, today’s sophisticated natural language systems understand what people mean. And yet, the experience must walk a fine line: the bot must act like a human, with very human attributes (such as empathy, curiosity, humor, compassion, and patience) while maintaining the transparency of being a robot.

Chatbot and agent uses

For the sake of simplicity, we will be referring to both chatbots and conversational agents as “bots” henceforth. Successful bots must augment capabilities and expertise, while also being an extension of the brands they represent. Bots have some clear uses including:

Natural language and understanding

Bots have the ability to engage in real conversation with users. The bot will be able to understand the user’s messages and context to provide a response that is relevant and useful.

Engagement and personalization

Bots provide personalization at scale. They can help companies automate one-on-one communications with their customers and personalize those communications using customer information and by asking meaningful questions to understand the customer. Bots with personality will build emotional connections between customers and brands to increase engagement.

Integration and growth

Bots are omni-channel. Successful bots will not be standalone applications, but rather a set of common tools that operate like a central cognitive brain. These can be deployed across all of the channels consumers use – messaging, mobile, phone systems, web, chat applications and social media. Bots do not have to roll out entirely new versions in order to constantly update the content and they can be trained on the fly based off real user data.

Scalability and consistency

A bot allows you to own the conversation and the dialog for every engagement and make that interaction available 24/7. You can offload questions that don’t require human interference for an instant response.

Education and entertainment

Bots are great for educating users on topics that are relevant to a brand, product or conversation topic. Bots also are diverse enough to entertain the user with games, natural conversation, or other forms of interaction.

Intelligent authentication and security

Voice biometrics allows consumers to easily and naturally authenticate their identity without having to type in a password or PIN by simply speaking a short passphrase. Voice biometrics significantly improves security over legacy authentication methods and prevents fraud.

What is conversation?

Since conversation is the bedrock foundation of meaningful relationships, a bot must be capable of holding an intelligent, two- way conversation. As social beings, we converse every day without giving it a second thought and our discourse is natural and autonomic. We tend to take that natural flow for granted. To utilize conversational technology to its full, game-changing potential, we must be consciously aware of how we communicate.

If we can understand how we communicate with each other we can begin to replicate this with a machine. For our intents and purposes, conversation is the meaningful exchange of ideas and information between two or more individuals. The requisite parts of a conversation are topics, exchanges, and utterances.


Topics provide context. They are the high-level subjects of a conversation at any given point.


Exchanges communicate information. An exchange consists of two or more utterances. Everything said within an exchange is relevant to either the topic or previous messaging. Without this correlation, there is no basis for understanding one another in conversation.


Utterances are the individual statements articulated in an exchange and the building blocks for all conversation. These expressions are the atomic, single turns within an exchange. Here is an example conversation that we can identify topics, exchanges, and utterances within.

Example Conversation

Example conversation

Crafting effective conversation

Keep in mind that a bot will only provide one half of the conversation. It could be initiating an exchange, or providing a response. You won’t be able to control the user’s end of a conversation. Your job, as a designer, is to provide a delightful conversational experience to the user using a bot as the medium. The key is developing your bot in a way that, no matter the utterance, the bot sounds natural and provides a believable response.

Preferred Responses

To provide a realistic conversation for your users, your bot must be relevant. A tacit expectation exists at the heart of every conversation. When we speak we are expecting a response that is relevant to the topic at hand, whether its good or bad. Yet, it goes a bit deeper than that. We are unknowingly hoping for a specific type of response in conversation. This is what we call a preferred response. The same goes for responses we hope we don’t receive, aka non-preferred responses.

Understanding what your users may view as preferred responses, then maximizing preferred responses in conversation is a key to natural, positive conversations. Another key is to develop satisfying, informative non-preferred responses that don’t come across as negative to the user. These are aspects of the conversations that we as humans find to be the most rewarding. Many small, rewarding interactions like these can build relationships over time with the bot.


In general, we expect a response of some kind when making an utterance. At the bare minimum we expect something back that is relevant to our initial utterance. This is how we anchor our conversations; we aren’t just shouting random utterances at each other.

It’s most thrilling when we feel, just as in human-human conversation, that a bot “understands” us. In a bot’s case, that means being stateful and contextually aware of the topic at hand. It’s critical for your bot to make the user feel understood while also maintaining relevancy.


Bots should be able to elegantly fail in nearly all situations. It’s crucial for a bot to provide authentic and relevant acknowledgement to a user when failure occurs. It’s okay for the bot to be wrong, but it’s not okay for it to be wrong and irrelevant. This will immediately take users out of the moment and will degrade perception of the bot’s abilities and comes across as unnatural. The ability for a bot to jump across multiple topics of discussion, handle harassment, recognize when an utterance is irrelevant or nonsense, or just get back on topic will be critical. Part of the designer’s job is to identify where and when conversation could get messy and account for it beforehand.

How does a bot interpret conversation?

A bot interprets our conversation digitally by identifying the intents and entities in our utterances, which in turn will give it the information it needs to craft a proper response.


Intent = The action or purpose behind their utterance. These are generally verbs.

For example, in the utterance: “I’d like to order some tacos and beer,” the intent would be to order. In the utterance: “Where is the bathroom?” the intent is to find a location.


Entities = An object that was mentioned in the utterance that is relevant to the user’s purpose. These are generally nouns. By recognizing the entities that are mentioned in the user’s input, the bot can choose the proper response.

For example, in the utterance: “I’d like to order some tacos and beer”. The entities would be:

Tacos = food

Beer = drink