Conversational UI needs AI

It will take some time before we have a chatbot capable of reacting like a real human being. Therefore, this can’t be the objective of deploying a chatbot. However, a chatbot is very well-suited to have simple dialogues that follow from not too complex, frequently asked questions.

In case the questions are complex, it’s sensible to integrate a switch to a human channel in the process. However, before a switch to another channel is necessary, a lot can still be improved on a chatbot by using a system working on the basis of artificial intelligence (AI).

The most important added value of AI is its continuous improvement in recognizing the input of users. Suppose you begin with a chatbot and you put a lot of answers in it to provide to users when they pose a related question. You can claim that an answer is never wrong: it has been written with a specific situation or customer journey and maybe even a persona in mind. And so it provides the perfect answer. But to what question?

Users of chatbots can (and will) pose questions in endless ways. Sometimes as a coherent sentence, sometimes as separate words, sometimes as randomly chosen words, typing errors fully included. The chatbot must recognize all these variations to be able to provide that perfect answer.

It has been our experience with chatbots that to train the chatbot manually (puppy learning) – even with proper dashboards – in many cases proves to be such a big task that the quality of a bot doesn’t live up to its potential.

To deploy AI in this context, a chatbot can learn from user input which isn’t recognized yet and add it. In this way, a smart algorithm can learn from questions which are posed slightly different, but do have a recognizable structure. Also, these algorithms can learn from user feedback on a given answer to determine if something needs to be adjusted in the recognition.

The selection of the proper answer is another application of AI. After recognizing parts of the question these have to be matched with answers to determine what is probably the best answer. A consideration of many chatbots is that an answer is always selected, even when the ‘matching confidence’ (the confidence that the answer suits the question) is very low. In cases of low confidence, it might be better to switch to a human channel where there’s a person who does understand the question.

An AI system which pays attention during the use of the chatbot can contribute to the selection of the proper answer. On the basis of context, the history of the dialogue, the behavior of the user (for example, think about the moment somebody does or doesn’t exit) and by its explicit or implicit feedback on the appropriateness of the answer, the system can determine if an answer matches the question. Just do this for a while and you will know which answers are and which aren’t appreciated by customers.

In this way AI brings a chatbot one step closer to intelligent interaction: a system learning from the dialogues and improving all the time.

If by now you are convinced that a chatbot with AI is the way to go, how do you approach this? Before you start searching for chatbot vendors who can provide you with the technology, spend a little time on first taking some smart design steps.

1. Determine where a chatbot provides added value to your company. Is it sales or rather service? As a conversational partner for product selection, in a customer process or as partner for after sales? Especially do not start too big and select an application that has strong customer value and at the same time provides cost reduction in an internal process.
2. Create design sketches of the conversational interface on your website. Determine the location of the chatbot and where customers might land. How’s the chatbot going to appear, what is its size, what is its relation with the other parts of the site? And what is the relation with search, navigation, and other channels?
3. Create a design of the content of the dialogues. Determine which topics the chatbot covers, think out the questions, counter questions, and answers, but also the tone of voice and maybe even the personality of the service.
4. Create a design of the feedback loops which are the basis for the AI which brings your service further. Are you going to request feedback from users or are you going to interpret their behavior? Requested feedback doesn’t always provide a representative point of view because users are inclined to react more often when they are dissatisfied. However, feedback can contribute to quality improvements. NB: AI is no silver bullet which can overtake the management of your chatbot completely. The algorithms of AI first need to learn before they can perform. It can be useful to implement the chatbot first without AI, in order to just collect customer questions and answers. They can serve as the training sets for your algorithms.
5. Take the prototype from the first four steps with you in your search for a company which can provide the proper chatbot technology with which you can implement your conversational UI. Ask them to provide a demo with your prototype as a reference, start a proof of concept, and experience how AI can really get your conversational UI to start talking.