Lekta’s Guide to Bot Design: Part 4, Chatbot UX Best Practices
At Lekta AI, we’ve been building chatbots since early 2017. While each project is unique, we found that certain universal principles can be applied regardless of the business domain, target audience or communication channel. Over the years, we’ve developed a battle-tested 4-step process that helps us deliver consistently great conversational experiences for our clients and their customers.
This is the fourth and the last in a series of articles on the topic that can serve as a comprehensive resource for anyone looking to build a bot. If you haven’t read Part 3: Conversational Best Practices, you might want to check it out beforehand.
You’ve now gathered requirements, developed your bot’s personality and written the dialogues. As a next step, we’ll share with you the key UX principles we rely on when building AI-driven solutions for our clients.
We take great pride in our UX expertise, and anyone who ever built a chatbot will know that building one presents multiple UX challenges. There’s a lot of unpredictability to be handled, you need to plan for misunderstandings, misspellings, different accents, and non-standard requests… At the same time, you need to keep your users engaged, conversations natural and customer satisfaction high.
Sounds like shooting for the stars? Don’t worry, the chatbot world ain’t so gloomy. With the right UX process (and technology), it all can be done and has been done before. Without further ado, here are our top principles you can learn and rely on from now on too.
20 UX Principles to Design Better Chatbots
1. Don’t pretend that your bot is a human being.
This principle should be common knowledge by now and most businesses do follow this rule, but let’s just explain once and for all why this is a no-go.
First of all, given the state of AI technology today, there’s no point lying :) Sooner or later, users will find out they’re not talking to a human being. Should you choose to lie, you have to realize that ethically speaking, you’re not being very respectful towards your customers. They may be disappointed with the dishonesty of your brand, and your next PR crisis will be only steps away.
In fact, according to Nielsen Norman Group — a leader in UX research, customers prefer when a business is transparent in this matter, not only for the ethics’ sake but because it helps them calibrate expectations and language. This is beneficial for both sides — the simpler the language your customers use, the better your bot understands them, and everyone wins. Not to mention, that people are more forgiving of any mistakes, if they know they’re talking to a machine.
Lastly, why hide that your bot is a bot if some people actually prefer talking to a chatbot than a human being? We observed this is often the case with debt collection. Talking to a bot (rather than a human being) puts customers at ease, as they’re more willing to discuss their debts with a machine.
2. Put time into developing your bot’s personality.
This principle is so important we wrote an entire article about it (Part 2, Chatbot Personality). The way your bot speaks and how it handles your customers’ issues has a significant impact on user experience. No one wants to talk to a dry, emotionless machine, no matter how efficient it is.
Defining the tone of voice, level of openness, attitudes and vocabulary helps design a consistent, on-brand experience for your users. This is especially important when you have multiple writers working on your bot’s dialogues. A predefined persona helps all writers stay in character and avoid inconsistent responses that may give users an impression they’re talking to a bot with a split personality.
3. Introduce your bot and its capabilities.
If you’re about to add a bot to your communication stack, you’ll need to think about how to introduce it (and its talents) to your users. There are many chatbots of varying quality out there, so it comes as no surprise that your users will have questions: “A new bot?”, “What can it help me with?”, “Can it deal with my return?”.
It is, therefore, a good practice to present the bot’s scope to the user at the beginning of an interaction. This has several goals:
- reducing uncertainty related to new technology
- setting users’ expectations
- avoiding unnecessary frustrations related to the tasks outside of your bot’s scope.
An introduction usually makes sense in two situations: at the very beginning of the interaction or when something goes wrong and we want to lead the user back to the path of effective conversation. Be careful not to overdo it though — too long of an introduction can be overwhelming and cause your users more harm than good.
However, at the initial stage of development, when the bot’s scope is not yet clearly defined, it may be a better choice not to communicate the narrow scope of the bot’s operations. On the one hand, it may cause more situations where your bot will not understand the users, but on the other, it will allow you to collect more data about your clients’ real needs and set a direction for further development.
4. Rely on popular streams to engage users and increase interactions.
If you want to get people to use your bot, automate business processes they use on a regular basis, i.e. mobile top-up, password reset. Frequent interactions with your bot will adapt the users to this new communication channel. Once they learn how to interact with your bot to solve their standard problems, they’ll be more willing to interact with it in other instances too.
This strategy will work far better than deploying a bot in popular streams but used by users usually once, i.e. filing a claim. Therefore, we always recommend checking if a business process you’re about to automate is popular because many people use it just once, or because the same people use it now and again.
5. Use different message components wisely.
Many communication platforms offer multiple conversation components. For example, Facebook Messenger provides buttons, quick replies, images or message templates, in addition to text. Making good use of these additional options can enrich your users’ experience and guide them at the same time (see example). Not to mention, that they’re good for business as well, as they allow you to control the narrative, reduce misunderstandings and speed things up, as users don’t have to type.
There are three things to keep in mind though: make sure that your chatbot understands text messages that say the same thing as your buttons. Sometimes people choose to type anyway, and it would be a shame if your bot had to answer “I don’t understand”. Keep in mind that each platform is different and what may work on Facebook might not necessarily be a good idea on your mobile app, for instance. Finally, rely on common sense — don’t use different message components just because they exist.
6. Get ready for chit chat.
Conversational AI is still a relatively new technology. No wonder that more tech-savvy users will want to test your bot inside out and see where its boundaries lie. You need to do better than a constant stream of “I don’t knows”. Wittiness, charm, and cleverness are usually your weapons here. From our experience users usually ask simple questions you can easily prepare for. Here are the most often asked questions from one of our recent deployments.
- How are you?
- What’s your name?
- What can you do?
- What’s the weather like?
- Are you single?
- What day of the week is it today?
7. Get ready for swear words.
How your chatbot will react to swear words is another situation you need to prepare for. Abusive or angry users are not uncommon, however, our experience shows that most of the cursing occurs while testing the bot and its capabilities. In fact, a huge chunk of testing is checking how the bot reacts to swear words or even sexual abuse. You need to prepare for these situations in advance and handle them accordingly.
So what should your bot do? Accept the abuse? Teach users good manners? Fight back? You may think it’s best to avoid conflict, acknowledge the swearing, yet remain polite and reply something along the lines of “Ouch, I’m sorry I didn’t meet your expectations”. No matter what you decide on, don’t forget that being too forgiving when it comes to swearing (especially sexual harassment) can get you bad rep too. Siri and Alexa are a prime example of that — they learned their lesson the hard way.
8. Avoid conversation loops.
There are a few simple tricks you can use to prevent your users from getting stuck in error loops. All you need to do is drive the conversation forward, even if your bot fails to understand a customer. Rather than saying “I don’t know what you mean”, suggest ways and actions a user could take to be understood. Here are a few examples:
- “Sorry, why don’t you ask me again using different words?”
- “Could you put it more simply?”
- “Could you rephrase that?”
- “Could you spell that?”
- “Could you type it instead?”
Naturally, you shouldn’t be asking these questions with no end in sight — this would lead to a conversation loop itself. Once you exhaust your options to understand the user, it’s worth adding a “submit error” option, buttons to click on, or an escape hatch to one of your consultants.
9. Provide an escape hatch.
You’re likely building a chatbot to lighten the load of your customer support team but don’t forget that your chatbot is there to help your customers first and foremost. This means that you need to do everything in your power to help them solve their issues. And even if you build the best chatbot on the market, there will be unforeseeable circumstances and times where the bot fails to help a customer.
In those times, you need to offer an escape hatch and connect the customer with a human consultant — you can’t leave them hanging. There’s nothing worse than a feeling of powerlessness and inability to solve a problem. Relying solely on automation will surely push your customers away.
10. Handle over-information.
In a natural conversation, we often provide a few pieces of information in one sentence, i.e. “I’d like to book a table for 4 people for Friday night”. Your bot has to detect and handle more than one intent at once. There’s nothing more that speaks you’re talking to a machine than a bot asking you about the information you already provided… Take a look at our example to get a taste of the frustration a dialogue like this may cause.
11. Confirm the user’s input when needed.
There will be times where a customer will answer in an ambiguous or complicated manner, and your bot will need to confirm whether it got everything right. Some situations, such as bank transfers always require confirmation, as the risk associated with them is much higher than with a product return, for instance.
There are two ways a bot can go about confirming information:
- By stating what was understood + leading the conversation forward, i.e. “I understand you’d like to book a dentist’s appointment. What time?”. At the end of the sentence, we’re asking a new question to advance the user journey, and treat lack of objection as confirmation.
- By confirming everything at the end of the process rather than in the middle of the dialogue, i.e. “To confirm, you’d like to send $100 to John?” — this is a good option if a misunderstanding would lead to serious trouble, in this case sending money to the wrong person.
12. Manage change of mind.
Change of mind management distinguishes winning chatbots from those that don’t live up to the task. People make mistakes or they change their minds, and your bot needs to be clever enough to manage these with grace. Imagine your customer changes one detail (i.e. visit time) when he practically provided all the necessary information to book a doctor’s appointment. Starting the process all over again would be a major setback and poor UX (see example). It should be possible to make changes at every step of the dialogue, just like in a natural conversation — even if it’s an immediate correction, i.e. “Yes, that works well. Sorry! I just remember I’m busy then.”
13. Master the art of negotiation.
Negotiating times and dates is no tea party, especially when it comes to voice-powered chatbots. Text-based conversational interfaces allow you to display all slots available on screen, but how do you go about designing this feature in voice?
Suggesting the earliest slot available is a good practice, but what happens when it’s no fit for your customer? Your bot should do better than let the user propose the time, and respond with “This slot is unavailable. Please choose a different one”. With a customer having no idea what slots are available, this exchange could go on forever (see an example on the left). Reading all options out loud would also take too long, and it’s not something practiced in human-human interactions either. Here’s where your bot’s negotiation skills come in, and where it can shine by narrowing down and filtering available slots most efficiently (see an example on the right).
14. Don’t let side-topics divert your bot away from the task at hand.
In a natural conversation, it’s quite common to make a digression or divert from the main topic sometimes. This also happens in human-chatbot interactions, and it’s usually related to testing, i.e. asking the bot if it likes Star Wars in the middle of a mobile top-up. However, side questions related to the main task can also sometimes come up. In both scenarios, your bot needs to be prepared for them: not only with an answer but also with a quick return to the task at hand (see example).
15. Prepare multiple responses.
We covered this principle briefly in our previous article, as it’s both related to a good conversation and UX. You don’t want your chatbot to sound too repetitive when it has to say the same thing more than once. A row of identical responses will leave your bot sounding like a broken record, and that’s just poor UX. You can easily prevent that by creating different variants of the same response, i.e. “sure”, “all right”, “very well then”, “okay”, “OK” and so on.
16. Prepare your bot for emoji.
Emojis have changed the way we communicate with each other. According to Emojipedia, 5 billion emojis are sent on Facebook Messenger every day. They’re part of our language now, and your bot needs to understand them and identify moods and sentiments that come with them. Some would go so far as to refer to an old adage and say that emojis are worth a thousand words. See for yourself what a conversation with a non-fluent in emoji bot could look like.
17. Pace out the answers.
Long answers sent in a matter of seconds won’t resemble a natural conversation. They will leave your users with the impression they’re talking to a robot and not a high-class virtual assistant. It’s a good practice to make use of typing indicator bubbles that are commonly used in human-human communication channels. Using them will make your bot look like it’s typing a message. Pacing out messages is also important when you’re bot is sending a few in a row. You wouldn’t want them to come all at once — not only it wouldn’t look natural, it would also be difficult for your users to follow.
Make sure not to overdo it though — you’re dealing with a bot, after all, so you wouldn’t want a user to wait 30 seconds for a reply, even though that’s how long it would take a human to write that.
18. Take into account any marketing campaigns.
Make sure your bot is up to date with any seasonal campaigns that might be going on. If you put posters all around the city with your new “buy 1, get 1 free” offer, rest assured your users will enquire about it, and your bot needs to be in the know. You wouldn’t want to end up in this situation…
19. Keep in mind the differences between voice and text.
If you’re developing both a text and voice bot, make sure that you thoroughly test and adjust the dialogues for the voice version too. There’s a difference between spoken and written language, and what may work in text might not necessarily be the best solution for voice. It’s common practice for voice dialogues to be shorter and more concise, as users don’t have the option to easily come back to what was said before.
Voice bots require an entirely different set of testing too: you’ll need to test for different accents or dialects, and for how your bot pronounces its dialogues. It might be the case that you’ll have to misspell some words, so your bot pronounces them more naturally, especially when your bot speaks in a different language than English. In our case, we had to spell “transakcje” with a “z” instead” of “s” to improve the bot’s pronunciation.
Lastly, bear in mind that different rules apply to voice-and-text bots too. Those projects open up an array of UX possibilities, as your bot can both speak and display the conversation on screen at the same time. For example, it can ask the user to confirm the bank account number and showcase it in the app instead of reading it out loud. Those chatbots are quite interesting from the UX perspective, as they allow you to use voice and text to their advantage and create truly seamless experiences.
20. Test and iterate, at all times.
Chatbot development, being quite a challenging endeavor, requires testing and iterating at every step of the process. Test your dialogues, new bot features and all new business processes you introduce to quickly find out about any problems or bugs. Test different scenarios: change your mind, negotiate, divert from the main topic and see if your bot preserves context.
Test by yourself, with your colleagues, friends & family, and with your customers. Finally, when you release your bot, roll it out slowly, first to a selected group of users, and only when that works well — to the entire database. This way, you’ll catch potential mistakes early, and will be able to iterate as you go.
Final Thoughts
That’s it! We hope you enjoyed the article and learned plenty today! Equipped with this knowledge, you’re more than prepared to create beyond-compare conversational experiences. We’ve also prepared a visual summary of all the principles listed in this article, should you wish to print it out and use it as a cheat sheet.
We hope you enjoyed our Bot Design Guide, and if you haven’t yet, feel free to check out its previous parts:
