9 Tips to Design Conversational Style for Your Bot

Greg Bennett
Salesforce Designer
7 min readOct 6, 2017


Leveraging linguistics to design conversation with personality

To design a bot means to design conversation.

Bots are having more and more conversations with humans online. And at a quick glance, it may seem straightforward to design a conversation for one. It’s easy to think that because we can have a conversation, that we’re great conversationalists. But, the truth is there is no such thing as a great conversationalist, let alone a perfect one. The way I talk, text, or chat could delight one person and thoroughly irritate another. It all comes down to conversational style.

It’s far too easy to let your own conversational style shape how your bot has conversation. But, one way of having conversation doesn’t work for all users. The challenge with designing bot conversation that is not only functional for the user, but pleasurable, is to do so in a way that engages users with varying beliefs about how to have conversation. Linguistics can help transform the design of conversation from a process often made based on assumption into one that is strategic and data-driven.

Linguistics is the study of how language and its structure work, and linguists have much to offer in the way of demystifying conversation systematically across cultures. Here, I present some foundational principles from linguistics as a set of tips for strategically designing bot conversation that has style, which you can then adjust to your customers’ and your business’ needs as you see fit.

1. Decide, Detect, Align, and Deviate.

Everyone has a way of having conversation. People develop beliefs about how conversation should be had based on their prior experiences and cultural upbringing. Whether that’s growing up in another country, state, city, or even working on a different team within an office, every social group has a way of talking, listening, and showing listenership. The way you talk in conversation is called conversational style (Tannen 1981; 1984; 1986).

When it comes to designing conversational style for your bot, you need to accomplish four things:

  1. Decide what kind of conversational style your bot should have, and in which contexts. Consider the voice and tone of your company’s brand. Is it enthusiastic? Considerate? Pick a baseline style that fits your business needs. When is it appropriate for your bot to stick to the style that matches your company’s brand voice? When does it more sense to break with that pattern?
  2. Detect the user’s conversational style. Do they use emojis? Do they leave the period off the end of their sentences? Recognize which language patterns correlate to a particular conversational style such as enthusiasm or considerateness, almost like personas of conversation.
  3. Align to the user’s style. If the user fires off emojis, the bot should respond with emojis. If the user leaves off the period at the end of their sentence, so should your bot. Set a conversational norm with the user.
  4. Deviate from the conversational norm with strategy. If the norm is to use emojis with the user, and your brand voice and tone call for high energy, having the bot insert more emojis into the chat can show that additional enthusiasm. Maybe the bot dials back on emojis because the subject matter of conversation is sensitive, serious, or inappropriate. Based on the context, when the bot deviates from the norm, it can get closer to or create more distance between itself and the user.
Joy, a Facebook Messenger bot designed to help users with mental health, has a cute avatar, but refrains from sending many emojis in first chat encounters with users (Davidson 2016), making it not overly enthusiastic or distant.

2. Speaking of context…remember it.

Remember what your users have said throughout the conversation, and be able refer to it at any point later on. The bot should be able to hold multiple topics in memory and use referents like “it” and “that” to create cohesion between utterances (e.g., upgraded Internet provider bot: Oh yeah, and how about that other thing you asked? The blinking light on your modem? Tell me more about that.). When your bot reroutes to a service agent, stories and other important details from the conversation should be passed along, too.

Internet provider bots could use words like “that” to indicate where subjects and references are located in space and time, putting the bot on a more level footing with a human.

3. Tell, listen to, and understand stories.

Human relationships are built on storytelling. It is an excellent way to build common ground and rapport with the other speaker (Tannen 1984). Bots should be able to listen to and remember users’ stories, especially in service encounters, and ideally, tell stories of their own in return to create rapport. Consider pre-writing stories for your bot to tell your users if it fits your business and interactional goals.

Microsoft’s Cortana has the beginnings of storytelling in its framework, and could tell more robust narratives about its character as it evolves its rapport strategies with users.

4. A question is always more than a question.

Asking questions in conversation can achieve many different social goals: you can try to solve a problem, get someone to open up more, encourage someone to ask you questions, show your enthusiasm for the conversation at hand, etc. (Tannen 1984). Bots should be able to assess how many questions the user is asking it and how quickly the user is asking questions, then gauge what kind of social goal a user is trying to achieve. Depending on the strategy you’ve given it to adhere to or abandon the voice and tone of your brand, your bot should route to dialogs that either turn up the heat on questions and answers, or cool down.

Steve Worswick’s Mitsuku has extensive question-and-answer capabilities which achieve different social goals in chat. In Image A above, Mitsuku engages me in asking questions to build rapport, but in Image B, the extra questions make it difficult for me to quickly problem solve.

5. Repair, repair, repair.

There is no perfection in conversation, only variation. Misunderstandings will always occur. We tend to forget that our everyday conversations are riddled with misunderstandings because we have built-in conversational techniques for clarifying them. Don’t be afraid to make mistakes in bot-human conversation. That’s inevitable. The key to a good conversation is in how you fix your mistakes. There are strategic ways to do this (Sacks, Schegloff, & Jefferson 1974). Click here for examples of repair in conversation (Nordquist 2017a). Ask followup questions, expand on what users say, give bots options for rephrasing. Teach your bot how to repair by building it into your dialogs. If you happen to be working with a data scientist to make your bot, they know this process as “disambiguation,” and it is used to train the bot.

LINE’s LINE USA bot can only pump static messages to the user and doesn’t handle trouble in conversation when it arises.

6. Embrace words like ‘hmm’ and ‘oh.’

Words like oh, well, and y’know — which we dismiss as unnecessary “filler words” — actually do things in conversation. They’re called discourse markers and they are crucial to keep a conversation going. They indicate how bits of information and references in talk relate to one another (Schiffrin 1987). You should build these into the way your bot converses to make it feel less rigid. Want to help your bot transition between dialogs or topics in the same conversation? Discourse markers can help you do that. Click here for your cheat sheet (Nordquist 2017b).

7. Pause with purpose.

Taking pauses/creating silence in conversation has meaning, a real function. People use it purposefully — whether that’s to create emphasis, show respect/politeness, etc. (Tannen 1984). It’s key to co-creating a conversation with other speakers. Write dialogs for your bot that use silences strategically. If the user is vehemently complaining and they’re still typing, give them the space to rant in the conversation. Sometimes, people just need to vent.

8. Strategically use openings and closings.

Everyone has a general sense of how to start and end conversation. But, linguistics can help you design your bot to do so strategically. It’s called openings and closings. Most bots already have phrases such as Hi, I’m here to help you and Thanks, and have a good day pretty well baked into their dialogs. Maybe you change Hi into How’s it going, and right before Thanks, you add in, Well, I think that just about covers it depending on the context. Maybe you decide not to because it doesn’t fit the conversational style you intend your bot to have, nor the voice and tone of your brand. Maybe the user expressed serious dissatisfaction in the conversation and Awesome! Have a great day! isn’t the best closing for that context. Fit your openings and closings to the context at hand.

9. Observe your surroundings.

Learning how to teach your bot to analyze and accomplish pleasurable conversation with your users actually teaches you more about yourself in the process. Ask yourself: Why do I talk/chat the way that I do? Why does it irritate me when my partner says/texts this way or that? Why does my coworker not understand what I’ve been repeating over and over for the past 20 minutes? Observe yourself in the wild. Observe others in the throes of conversation. Take notes, write everything down, and discover. Use this opportunity to reflect and connect more deeply with others — all while building a great bot that could evolve what it means to have conversation.

Final Thoughts

Designing pleasurable conversation for different users is tough. But, by relying on linguistics, that challenge gets a lot less fuzzy and a lot more purposeful. As you blaze your own trail and design your own cutting-edge bot, use this research to your advantage. I look forward to experiencing the conversations you design. 🤖 🙌

Click here to access works referenced in this post.

Thoughts? Questions? Comment below!



Greg Bennett
Salesforce Designer

Director of Conversation Design at Salesforce