The Six-Stage Model of Conversational AI Design was first published on June 30, 2021 by Sophie Hundertmark and Prof. Dr. Christian Hildebrand in the St. Gallen Marketing Review. An excerpt of the most important steps can be found in this blog post.
The 6 stages described are intended to support companies in the successful introduction of Conversational AI applications, such as chatbots or voicebots. Sophie and Christian recommend working through all 6 steps step by step to ensure the success of the new AI application.
The term Conversational AI is used below as an umbrella term for all chat and voicebot or robot applications,
Stage 1 – WHAT: Define the User Problem
The first stage involves defining the specific task to be automated by Conversational AI. It is advisable to analyze the entire customer journey and critically evaluate the stage at which Conversational AI applications contribute to existing corporate goals. The recommendation from experts Sophie and Christian here is to first outline a list of specific tasks with high automation potential and then prioritize the selected use cases or tasks in terms of feasibility and business value. Suppose a company has three potential pilot projects where the options for solving the user problem are between (1) automating parts of the customer service process through a chatbot, (2) providing automated decision support by offering additional product information on a website through a chatbot, or (3) actively providing recommendations during the sales process through a chatbot. While the first project is probably very feasible, the expected business value is comparatively low. In contrast, the last project has a comparatively high business value, but a lower feasibility, as the recommendations need to be highly individualized and contextualized. Finally, the clear articulation of the problem to be solved should also be closely linked to the definition of how the company will track progress towards this goal (e.g. the evaluation of customer acquisition costs before and after the pilot or the duration of customer service calls, depending on the main goal of the Conversational AI application).
Bill Price’s Value Irritant Matrix also serves as a further decision-making aid here. It helps to find out which processes and customer touchpoints are most likely to be automated. Namely, precisely those that are of low value to the company, but of high value to the customer.
To summarize, the first phase requires an exclusive focus on the problem, which should definitely be approached from the user or customer perspective.
Stage 2 – WHO: Define the Target User
The second stage serves to define the intended target user. Sophie and Christian recommend using methods that have already been used in previous work on customer journey mapping by defining the “archetypal user persona” (Lemon & Verhoef, 2016). The company should evaluate the predominant user of the chatbot or conversational AI application for the defined problem (stage 1). The intended persona(s) should be as specific as possible and cover both objective characteristics such as demographics (age, gender, marital status) and subjective characteristics such as the expected emotional state or important values of the user. Anticipating a user’s emotional state is crucial, as a negative experience from an already frustrated user can lead to a negative downward spiral and a negative evaluation of the company. For example, a recent study by Hadi (2019) showed that customers who were in a state of anger or frustration before interacting with a customer service chatbot were even more frustrated after the interaction and rated the company more negatively compared to a group of control customers. Companies are therefore advised to carefully define the expected target user and, above all, to take into account the expected emotional state of the user.
Stage 3 – WHERE: Define the Channel Integration Across Touchpoints
The third phase is used to decide how, where and when Conversational AI should be used. This can range from integration on an existing website or within an existing third-party platform (WhatsApp, WeChat, Facebook, etc.) to use as a standalone app. In this phase, a decision must be made as to which core channel the application should be used via and how the channel change should be handled across the touchpoints. For example, a company might decide to implement a chatbot as part of its sales automation processes on the website. Integration and deployment therefore require a decision on where on the website the Conversational AI should be made available and what should happen in the event of a failed intent match (i.e. if the Conversational AI does not know how to answer or process a request from a prospective customer). If intent matching fails, the user should switch seamlessly from one interaction channel to the next. This could mean that the seamless integration is handled by human sales representatives or that the request is forwarded directly as an email to the human customer service team. Failed intent matching and ineffective channel integration are a major source of frustration for users of bots due to the reduced sense of goal achievement (Leung & Chan, 2020). The aim of the third stage is therefore to decide which channel to focus on and to consider possible interactions across all channels and touchpoints. Furthermore, companies should carefully evaluate and optimize failed intent matches not only before or during a pilot phase, but also at regular intervals after the Conversational AI application has gone live. Most Conversational AI interfaces provide identifiers that explicitly flag failed interactions or intentions in the backend, which can (and should) be carefully and systematically analyzed.
Stage 4 – HOW: Design the Appearance, Tonality, and Personality of the AI
The fourth stage comprises all design-related decisions and includes the visual appearance and the conversational design (the structural and semantic characteristics of how the conversation between a human user and the AI is regulated).
The visual design in the context of chatbots includes, for example, the company or brand-congruent design of the digital avatar, the name of the avatar, the color scheme and the font. The selection of the avatar and the appropriate naming are critical design decisions (Miao et al., 2021), as they directly reflect the brand. Conversation design captures the structural and semantic properties of the conversation. The structural dimensions include, for example, the frequency or extent of turn-taking (i.e. whether conversational AI actively promotes back-and-forth communication as in human-to-human communication). Recent studies have shown that a greater degree of turn-taking promotes trust and a more positive evaluation of the brand (Hildebrand & Bergner, 2020). The semantic dimension captures the tonality of the chatbot and can range from a more formal to an informal communication tone (e.g. through the use of affective language or emojis). The combination of visual and content design ultimately defines the type of personality that the user ascribes to the AI. For example, more affective language can be deliberately used to create a more extraverted “personality” of the AI. Such personality attributions can even be elicited by more subtle cues, such as longer pauses to signal greater thoughtfulness of the AI, or increasing the variability of voice frequency to signal excitement (Hildebrand et al., 2020). In short, human users tend to ascribe different personalities to the chatbot, and the visual and conversational design of the system are key factors in developing the intended personality profiles of the company or brand (Nass & Moon, 2000; Nass et al., 1994).
The results of the second phase, which among other things deal with the expected user emotions, certainly serve as a decision-making aid in this phase.
Stage 5 – WITH WHOM: Define the Extended Project Team & Stakeholders
The fifth phase focuses on the definition of the extended project team and the active onboarding of internal stakeholders.
This phase is essential to gain buy-in from the entire organization and the extended project team (all internal stakeholders who are either directly or indirectly involved in the project, such as the internal IT department). The fact that Conversational AI applications are based on relatively new technological developments can lead to internal resistance that needs to be actively managed. This phase is crucial to avoid false expectations and at the same time to identify strong internal promoters for the pilot project. It is important to anticipate potential internal resistance and react to it. For example, it is advisable to actively communicate how the success of the pilot will be measured (e.g. as part of a sales automation project; this may include the amount of traffic to key landing pages, the actual conversion rates on these landing pages or the number of pages visited before the key landing page).
Stage 6 – WITH WHAT: Define the Technology Stack
The sixth stage focuses on selecting the most suitable technology stack with regard to all previous stages. The decision on the technology stack is located in the last stage of the model for two main reasons. Firstly, the technology stack should be selected according to the specific use case, regardless of the internal processes and infrastructure already in place. This sequence is intended to avoid narrowing down the options and distracting from the central problem of the end user as opposed to the technology. Secondly, focusing on the technology according to the requirements for conversational AI helps to critically evaluate whether the targeted use case requires a complex natural language processing engine or whether a simple rule-based conversational agent is sufficient.
However, Sophie and Christian point out that organizational requirements, such as technologies already in use, existing data protection regulations and other project resources, must also be taken into account at this stage. The model presented can be seen as an ideal archetype, and the actual implementation of a project may require several iterations going back from the technology to earlier stages. As the number of providers of Conversational AI solutions continues to grow (from Amazon, Google, Microsoft and IBM to smaller, specialized solution providers), Sophie and Christian recommend keeping a regular eye on developments on the technology side. As with any enterprise computing project, large companies tend to offer highly scalable but more standardized solutions, while smaller providers are often more actively involved and offer better customization.
The following figure shows the “Six-Stage Model of Conversational AI Design” developed by Sophie Hundertmark and Prof. Dr. Christian Hildebrand.
The paper on the “Six-Stage Model of Conversational AI Design” was first published by Sophie Hundertmark and Prof. Dr. Christian Hildebrand in the St. Gallen Marketing Review. The entire journal can be found here.
