Why dialog trees are not a good foundation for powerful AI assistants
For e-commerce store operators, high customer satisfaction is the key to success. Using chatbots to communicate with customers promises fast response times and thus higher customer satisfaction. There are a variety of bot providers on the market. Their natural language understanding techniques are often similar. The main difference between bot developer companies is the type of dialog management applied. Many store providers have had bad experiences with bots that are too impersonal or inflexible. This has a negative impact on customer satisfaction and leads to aborted interactions. In case of doubt, users look for another store. This can be avoided with good dialog management.
For artificial intelligence assistants such as Alexa, but also for conversational bots, it is not enough to understand the meaning of every sentence spoken (see our blog post on training bots). They need to have a good strategy to manage the entire dialog, react flexibly to a user's input and thus avoid frustration and abandonment. In fact, for the customer experience with a bot, dialog management is much more important than the ability to recognize individual sentences. This is about optimizing bot behavior along the entire interaction with the user. What should the bot say or ask next? If there are different possible interpretations, which is the right one in the context of the overall conversation? Is it important to ask for more details from the user or show results quickly? Is it better to display an image or text as an answer? How should topic changes by the user be handled?
There is a lot that can be done wrong when it comes to dialog management. Different approaches have clear advantages and disadvantages. A conventional approach encountered by many conversational bot vendors is the use of dialog trees. Dialog trees fix the sequence in which a bot asks for information from the user, often in the form of rules such as "If the user has chosen the size of the T-shirt, ask him for the color." Once the size is fixed, there is no escape. Only after the user has entered the color can the dialog continue. This seems inflexible to many users and leads to frustration. Dialogs that are conducted by working through a dialog tree are far apart from the kind of interaction we humans are used to. We have the flexibility to adapt to what the other side is saying.
The fact that a dialog tree is not a good basis for powerful chatbots can easily be seen in the following analogy. Imagine you are applying for a new job. You lay out a script for each interview in advance, which you strictly follow during each interview, and thus do not react to the flow of the conversation or even the questions of the other person. For the same reason that these interviews won't go well (and you wouldn't get the job), neither can a good flow in chatbots be reached with dialog trees; the dialog just feels "scripted." Such a bot loses the feel of the flow of the conversation, bluntly responding to each sentence and working off a script. The bot literally "can't see the forest for the trees" here.
Mercury.ai bots, on the other hand, implement an innovative way of conducting dialog, in which the bot calculates the next action based on the course of the conversation. This means that changing topics, for example, is not a problem. A user can search for a product, interrupt the search at any time to ask when a shipment for a product already ordered will arrive, and then resume shopping.
Mercury.ai bots achieve a more flexible dialog design through so-called games. In other words, the bot engages in a "game" with the user. Each game represents a type of interaction, for example product search, shipment tracking or return of goods, and uses the dialog strategy that is best suited for this subgoal of the interaction. With each message from the user, the system calculates which game is responsible for the user's concern. Thus, the user can jump to any other topic with each message. Within the games, flexibility is achieved by reacting dynamically and contextually. This means that there is always a new sequence within a game, which makes the dialog interesting and less predictable. In addition, the games remember the previous state and context, so that an earlier topic can be taken up again at exactly the point where it was last left. This creates a high-quality interaction that responds to the user's needs and enhances customer experience.
With dialog trees, this kind of flexibility can only be achieved through a large number of different paths and branches, which makes further developments and extensions increasingly laborious, time-consuming and therefore cost-intensive.
What are the most important advantages of a flexible approach to dialog management? Quite simple: a better customer experience, fewer abandonments and thus higher conversion rates. Customers reach their destination faster, which leads to higher customer satisfaction.
Are you already leveraging the benefits of customer service automation via channels such as WhatsApp or Facebook Messenger to increase your shop's customer satisfaction?