Conversational Interfaces

Firstly, what are LLMs and ChatGPT? This is not an article about what LLMs (or large language models) and ChatGPT are. If you have been living under a rock and are unfamiliar with these names and terminology then this article written by ChatGPT explaining itself should be a good starting point.

We have been receiving questions from – and participated in many discussions with – our customers and peers about this exciting new tech and wanted to clarify our stance on where we see the opportunities and weaknesses at the current stage, as well as looking forward to a potential hybridized future. The biggest talking point has been the need for conversation design in an increasingly automated and generative world.

From our perspective as experts on conversational interfaces and conversation design we see predominantly two paths that this technology and trend will continue to develop on: the path of consumer-facing applications and the path of the technology as a tool and force multiplier. Neither of which will be eliminating the need for humans behind the wheel, steering the technology, anytime soon.

Hopping on the LLM bandwagon

Broadly speaking, this technology and its implications are spreading at breakneck speed. Many platforms are currently aiming at capitalizing on this goldrush-like state. You may have heard of Microsoft implementing ChatGPT in Bing and Google looking at fusing their proprietary equivalent LaMDa with their own search engine. These search engines follow a trend that companies such as SoundHound have been pursuing for a while, responding to users not in lists of search results, but in concrete answers in the form of natural language.

Other examples of quick wins in this brand new space are bot platforms such as Voiceflow and Cognigy.AI. Here the same purpose of applying LLMs to dynamically generate the system responses or predictable training data for intent training is being used heavily. Some platforms, like Cognigy.AI, are also considering going a step further and looking into the empowerment of conversation designers by allowing the creation of flows and elements through natural language prompts, speeding up the process of setting up new conversations greatly and thus contributing to rapid prototyping capabilities of these low-code platforms. Will these features collate into conversations that are production-ready, about to be rolled out to millions of users, out-of-the-box? Of course not. But they provide a good first framework to expand upon.

Trust in the system and the tech is dwindling

Widely broadcasted anecdotes of tech journalists and influencers, as well as hear-say from colleagues and friends have recently lead to a lot of skepticism when it comes to the current state of the technology. Articles quoting the unsettling feeling, individual erroneous responses and behavioral patterns reinforce negative connotations when it comes to LLMs in todays world. This obviously has a huge negative impact on consumer-facing applications.

Finding an appropriate place for LLMs should not be difficult

Focusing on this new technology as a force multiplies and enablement tool, is therefore the more stable path from our perspective. At least while the technology matures and new, more refreshing experiences for consumer-facing applications improve the publics perception in the mid-term.

On a more immediate and applied note, ChatGPT and LLMs are a great vehicle for innovation and a popular driver for change, but they are tools and will not replace human experts in conversation design. It is a good gap-filler and repetitive tasks but it will not provide the confidence and accuracy of dialogues designed by humans for a while.

The conversation designer is still the agent of change for this new tech

Our workflows in the future could consist of conversation designers laying down the structure of a dialogue, such as the starting point, the goal of the conversation and some checkpoints along the way, with the generative AI or LLM filling the gaps.

In an ideal world we would provide the AI with a purpose and a personality, but no actual dialogue would need to be written by humans. The conversation designer would be focused entirely on the strategic purpose of the interface and the decision on a vector of the personality and tone of voice of the bot.
Paul Krizsan, Director Conversational AI

So while remaining up to date with the current developments of this exciting new technology is vital, we do not share the current ubiquitous sentiment that users are ready for unfettered access to potentially image-harming experiences without having some of the kinks of current LLMs ironed out over the course of 2023.

Are you interested in talking about conversational interfaces, LLMs and how to design for conversations? Talk to us!

This past year, we have designed and refined a wide range of chatbots. One project was particularly challenging because it required a chatbot with more than 400 intents, multiple audiences, multiple languages, and multiple touch points.

In the process, we had the opportunity to try out and learn a lot, which we would like to share with you in this post. You will find out what’s involved in developing chatbots and why it’s so important to keep enhancing them.

Get to Know The Business, Brand, and Users

When starting to work with a customer, it is essential to get to know and understand them: Who is the customer? What does their brand stand for and what problem needs to be solved? What is the business model? What do we want and what can we achieve with a chatbot? And for whom? Which challenges are significant for which target groups?

Any such questions help to gain a profound understanding of the users’ needs and the customer’s demands. In the process, we analyze possible interactions and synergies. The goal is to define what the chatbot should be capable of and how it can make a contribution.

For this purpose, we usually conduct workshops with the customer. Such joint development is important not only to clarify expectations, but above all to see what is feasible. In addition to defining the goals of the chatbot development, this serves as the basis for later collaboration. For the aforementioned chatbot, this specifically meant to:minimize the volume of inquiries to the service centerlive up to the promise of high-quality serviceengage with customers at all timealign the user experience with the brand and its values

In addition, it is useful to develop a common basic understanding of the chatbot personality because, in our experience, it is essential for further decisions and developments.

Develop a Distinct Personality for the Bot

It’s much like us humans: the more unique a bot is, the more memorable it is.

For chatbots, the brand personality equals the user experience. We therefore attach particular importance to designing the chatbot to match the brand. If the behavior of the chatbot does not correspond to the expected brand behavior, it does not contribute to an authentic brand perception.

To understand the brand of our customer described here, we first developed a brand personality from which we later derived important attributes for the chatbot. We start by asking what type of person the brand might be. What makes them tick? What are their specific characteristics, interests, and traits? How and where do they live? It’s okay to be creative! At the same time, it helps to also keep in mind what you spontaneously associate with the brand.

Later, when the bot personality is created, an array of associations with the brand are incorporated, helping to create a strong, distinct character that responds appropriately to its users and — while potentially different from the brand personality — is still consistent with the brand. Just as the service staff at Apple stores represent the Apple brand, but are still distinct personalities.

Our sample bot personality has the following core attributes that make it stand out as a character, being:humanfocusedvisionary

It may help to match such attributes with so-called archetypes in order to gain an even clearer understanding of the personality. These archetypes are models from psychology that categorize certain personality attributes and ur-forms of human beings. We use the 12 archetypes by Carol S. Pearson as a guide, with our sample bot corresponding to the protector and the hero. Accordingly, it strives to support its users, to help them, but also to act competently and courageously as a “rock in the surf”.

From this preliminary work we later deduced design principles, which we transferred to the look and feel of the chatbot, its tone of voice, as well as its user guidance and navigation.

Defining the Tone of Voice

As a next step, the bot personality is to be experienced linguistically. This involves defining the way the chatbot writes or speaks. The tone of voice (read more about it in our article), also called text tonality or brand voice, is crucial for this.

It should always be clearly communicated to the user that he or she is talking to a chatbot — and not a human being. Doing so avoids irritation, creates transparency, and sets a clear expectation. The chatbot from our example clarifies this by introducing itself as a “virtual assistant”. Still, a name for a chatbot can make perfect sense because it can strengthen the connection to the brand and also underline its individual personality.

Designing a Chatbot-Human Dialog Structure

In order to simulate the possible conversation flows, it helps to script ideal dialog sequences and to map them in flows. Using methods from UX design, a good user experience with the chatbot is laid out. Decision trees are used to outline how the dialog between user and chatbot might unfold. What we know is that a natural conversation usually does not run in a linear fashion.

The chatbot from our example was to provide a wide range of information about our customer’s product. It was supposed to know the answer to a large number of questions from different user groups. However, depending on each user group, the contexts of the conversation were completely different. For example, a salesperson has different questions than a consumer and a press representative has different questions than a sponsor. To adapt the UX of the chatbot to the needs of the different groups, we optimized the start of the conversation with an initial query.

In addition, we have anticipated other likely dialogs that can be summarized as “small talk”. Indeed, our experience shows that people also use empty phrases or ask follow-up questions about the weather when talking to a chatbot. Be it to challenge the bot or because certain manners in natural language are standardized and run partly subconsciously.

In any case, it is important to present as many conversation scenarios as possible. The more the bot knows and the more versatile it reacts, the greater the added value for the brand and the user.

Embracing the Chatbot as a Living Brand Experience

Chatbots are living creatures that can and even must adapt to user and brand needs. Only with real user input can the bot personality and user guidance, so carefully crafted beforehand, be tested. Does the chatbot correctly recognize user intents, does it provide the appropriate answers, is it used actively and on a recurring basis, or does it often get stuck?

Based on our projects in the areas of branded conversations and conversational AI, we know that users sometimes find it difficult to articulate questions and concerns. In such cases, the potential and benefits of the bot can only be exploited to a limited extent. In this project, too, such a trend became apparent quite quickly in the monitoring. Therefore, we adjusted our user guidance concept: We clustered all topics and knowledge domains to transfer them into a scheme for suggestions. Depending on the previous answer, the chatbot would propose further questions. By doing so, our chatbot was able to live up to the “hero” archetype and provide competent assistance to those seeking information. We quickly saw that this customization brought significant benefits to the chatbot’s users.

Intrigued?

Interested in learning more about chatbots? Feel free to visit our website or contact us at hi@thinkmoto.de. We are happy to help you!

Tag: Conversational Interfaces

What ChatGPT and LLMs Mean for How We Build Conversational Interfaces for the Future