Get in touch
Welcome to the second edition our series on technology case studies from our portfolio companies. Check out the first edition in case you have missed it!
This post was also published on Medium.
Around 16 million visits a month make Quoka – a member of the Russmedia Group – one of Germany’s most successful classifieds portals. Being an online marketplace, all of our potential user actions such as account creation, ad placement, and messaging are designed to be highly user-friendly and intuitive. Yet, users may stumble upon uncertainties and require more information on a specific topic from time to time. In many cases, advice can be found in our frequently asked questions (FAQs). But let’s face it: Have you ever enjoyed browsing through this rather dry section of a website? In today’s rapidly progressing digital world, users expect to spend the least possible effort until having a question solved. They want their requests to be answered with an individually tailored response. And ideally, they receive the needed information immediately. To provide this kind of high-level support, a whole lot of qualified customer service assistants seemed to be necessary. But what if we could meet our users’ needs and manage their requests with just one single agent?
With Rasa, an open source machine learning framework, we started building a chatbot, designed to handle most of our customer support requests on its own. We never considered it as a static problem-solver but rather as a social partner providing personal care within a natural conversation. To ensure a positive user experience, we set focus on both, the user-goal as well as the interactional journey until reaching it.
In this article, we will share our past experiences and give you guidance on what it takes to design an exciting while at the same time effective customer service chatbot from a conceptional perspective.
Many humans are fascinated when thinking about talking to computers. In our minds, vivid images have emerged over the last decades, mostly evoked by depictions of futuristic, speech-enabled computers in the media. However, surreal machines such as HAL9000 or KITT have been superseded by genuine voice-controlled assistants like Apple’s Siri or Amazon’s Alexa. With speech-based systems actually being able to produce and understand natural language, computers are no longer seen as bare zeros and ones, but are in fact perceived as social actors within a conversation. Pioneered by findings of Byron Reeves and Clifford Nass , a series of studies in the 1990s has shown the impact of certain contextual cues on the ‘socialisation’ of technology: The computer’s ability to process speech, to cope with interactional reference or its occupation of human roles (e.g. as assistant, mate, or therapist) usually leads to the expectation of a social exchange . Humans count on the machine to act in a socially compatible way and at the same time show social behaviour themselves when reacting towards it. This may manifest itself in several social (inter)actions such as applying gender stereotypes to a computer , handling it with concepts of politeness  or even disclosing one’s secrets in an act of reciprocity .
Taking these insights into account, designing a chatbot seems to be more than just creating a faceless answering-machine. Instead, users may – in some cases unknowingly – expect to engage in a social interaction. To meet this kind of requirement, a holistic design approach is necessary, starting with the appearance of the bot itself and carried forward by a smooth and elaborate conversation flow.
Every conversational agent should be based on a well-conceived persona. To get an idea of this concept, imagine a fictional character from a novel or a movie. In order to create a relatable figure, it was given a name, personality traits, and even a distinct life story – and so should your bot. In case you are working with visual elements, take into consideration that there needs to be a “physical” appearance of your virtual assistant as well. If you ever wanted to be a Hollywood screenwriter, this is your chance to shine.
When designing your persona, there is one key factor to be taken into account right at the beginning: Your character should fit to the image of your brand or product. It will be irritating if your bot tries to sell investment funds with a hang-loose-attitude or on the contrary speaks in a highbrow, formal tone to customers of your local surf shop. Try to capture and adapt to the communication strategy of your established channels to create a feeling of consistency within your overall brand representation. On top of that, be aware that every little detail of your persona design will evoke a set of expectations in your users’ minds. The choice of your virtual assistant’s name for example is one of the most crucial challenges. As mentioned in the last section, a perceived gender – usually triggered by the bot’s name or look – can have a huge impact on stereotypical expectations and behaviour. Therefore, a clever pick may have a positive or negative influence on your chatbot’s success. Moreover, your naming may not only affect an ascribed gender, but (at least partly) decide over your chatbot being construed as a rather machine- or human-like dialogue partner. This goes along with certain consequences: While a more human agent may build up trust due to its familiarity , the machine may be treated with more acceptance concerning its limited comprehension in case of conversational failure .
While planning your persona design keep in mind that there is no universal blueprint for all possible use cases. Instead, you should aim at creating your own custom-fitted character which – based on its story, ideas and look – resembles your brand without even talking about it.
We all know these situations from our daily lives: We meet someone for the first time – let’s say a new colleague or a friend of a friend – and only need a few moments to decide over the chances of getting along with each other. A similar process takes place when your users start reading your bot’s initial lines. This first impression serves as a prime determinator regarding whether your users will accept your chatbot as a new communication channel or not. Furthermore, the tone and meaning of your introductory words will strongly influence the bot’s perceived persona. Therefore, you should wisely choose on how to start your conversation.
As you would probably do in a human-human-interaction, begin with a brief introduction of your bot. Let it state its name and also its purpose. Keep in mind that many users may have never interacted with a chatbot before. Hence, it is also necessary to proactively instruct on how to talk to this new kind of dialogue partner. Make clear if your bot favours to process language in natural forms (e.g. “I would like to create an ad.”) or has a preference for robotic imperative (e.g. “Create ad!”). In addition, give concrete examples on what to ask the bot. Some users may have a question in mind, but others will stumble upon your chatbot by accident. Furthermore, many machine learning frameworks for building chatbots will include the option to work with intent-buttons. These offer predefined utterances to the user and overrule the necessity of natural language input. Therewith, misinterpretations are avoided and a successful interaction can be provided right at the beginning. Take advantage of this positive experience by praising the user and encouraging him or her to continue the conversation.
After your bot has smoothly mastered the start of the conversation, it needs to be able to solve the actual question of your customer. Therefore, it is necessary to define a substantial amount of possible interactions around your product in advance. Surely, you can think of some examples out of the blue or just stick to your already composed collection of FAQs. A better way is to rely on real conversational data. Unfortunately, no existent corpora will provide you with the information that perfectly fit to your use case. Instead, you need to collect the data by yourself – by chatting with your customers human-to-human over a longer period of time. There are several tools on the market which allow you to integrate a chat-interface into your website. These will usually store the conversations so you can analyse and classify your data later on. You may think of this as a Wizard-of-Oz-like experiment where you disguise your human nature under the veil of a digital assistant to provoke authentic communication. In case you do so, keep in mind to stick to research ethics and always disclose your actual identity at some point within the conversation.
After you have gained first insights on what your customers might ask in a support chat, you can start building concrete dialogue units for solving specific issues. In case your bot is mainly serving as an assistant for requests in the style of frequently asked questions, usually two dialogue turns (the user asking and the bot answering) will be enough to handle an issue. The more turns it takes to solve a request, the more complex your underlying design needs to be – and the more misunderstandings may happen. In many scenarios you can prevent uncontrolled dialogue prolongations by anticipating your user’s next action. Provide a few additional facts in your bot’s answer (but still try to keep it short) or offer to share a specific set of related information that you have prepared on an appendant dialogue path (e.g. via intent-buttons). This last approach is also advisable, when your conversational agent is uncertain about a user’s intent. Openly admit that your bot may have reached its current limits instead of presenting an incongruous reaction. As a last resort you can still redirect the user to a human support agent.
In general, don’t be in fear of interactional errors, but rather think of your chatbot design as an ongoing process. Dependent on the versatility of your product, it may take months or even years to create a largely competent conversational partner. Therefore, you will need to improve your dialogue flows by continuously analysing your bot’s conversational success and failure.
Remember that every speech-based human-machine-interaction carries a social component. Answering your customers’ questions in a neutral tone will solve their current issue but does not necessarily lead to a memorable positive experience. To have your users enjoy the interaction with your chatbot (and make them return to this channel and your website), you need to upgrade the conversation on a social level. This starts with general conventions such as greeting, giving thanks, or saying goodbye to the user. With regards to your persona design you may also emotionalise your chatbot by making it more empathetic. Express regret in case a user is having troubles or show joy when he or she has solved a problem with your help. Furthermore, prepare your bot for so-called chit-chat. Users will love to challenge your bot with statements and questions that are not related to your product. These may concern your persona (e.g. “How old are you”; “Are you male or female?”) but also more general topics (e.g. “What day is today?”; “What is the meaning of life?”). Users may even try to tease your bot by asking questions about your competitors (e.g. “How do you feel about eBay?”). Moreover, your bot will be confronted with positive and negative evaluations. Be prepared for praise (e.g. “You are a great help!”) on the one side and harsh affronts (“You are a stupid bot!”) on the other. When planning your reactions to these kinds of utterances, try to turn every exchange into a positive user experience. Always stick to your overall communication style and keep in mind that this will shape your bot’s persona. As it will be impossible to anticipate all varieties of chit-chat-scenarios, a constant evaluation and training process is necessary.
The above considerations should prepare you for a good start when designing your first (or next) conversational agent. Keep in mind, that no chatbot is like the other and every use case is unique. Therefore, always be adventurous and try to make your own experiences. Nevertheless, the following thoughts may help you on your journey:
 Reeves, Byron & Clifford Nass (1996): The media equation. How people treat computers, television, and new media like real people and places. New York (NY): Cambridge University Press.
 Nass, Clifford & Youngme Moon (2000): Machines and mindlessness: social responses to computers. Journal of Social Issues 56 (1), 81–103.
 Nass, Clifford, Youngme Moon & Nancy Green (1997): Are computers gender-neutral? Gender stereotypic responses to computers. Journal of Applied Social Psychology 27 (10), 864–876.
 Nass, Clifford, Youngme Moon & Paul Carney (1999): Are people polite to computers? Responses to computer- based interviewing systems. Journal of Applied Social Psychology 29 (5), 1093–1109.
 Moon, Youngme (2000): Intimate exchanges: using computers to elicit self‐disclosure from consumers. Journal of Consumer Research 26 (4), 323–339.
 Komiak, Sherrie Y. & Izak Benbasat (2006): The effects of personalization and familiarity on trust and adoption of recommendation agents. Management Information Systems (MIS) Quarterly 30, 941–960.
 Duffy, Brian R. (2003): Anthropomorphism and the social robot. Robotics and Autonomous Systems 42, 177–190.
Tech Case Study 5
You’re next Let’s go