How much AI does a chatbot need?

The question of how much AI a chatbot needs cannot be answered unequivocally, because it depends on the use case. There are chatbot use cases that do not require the use of artificial intelligence (AI) at all. For other companies, it is essential that the chatbot uses the possibilities of AI - or Artificial Intelligence.

In general, it can be said that simple use cases in which the dialogs always proceed in the same order can usually manage without the use of AI. We are also talking here about rule-based chatbots or so-called click bots.

We know this when chatbot users are guided through the conversation by means of buttons. In this case, the chatbot designers can predict the course of the conversation very well, define buttons and texts and guide the user through a predefined dialog, so to speak.

If the conversation is not so easy to predict or the answers are highly context-dependent, then it is advisable to use a chatbot that works using Natural Language Understanding.

What follows is a brief digression on the different ways chatbots work:

Rule-based chatbots

Rule-based chatbots work, as the name suggests, with rules that are defined in advance - as in a tree diagram. Depending on which path the user selects, the bot responds according to the specifications.

If the chatbot is created for a very specific case, where the process and the user's questions always come in the same order, it is recommended to opt for this simple variant. For example, if you want to ask for feedback from your customers, it may make sense to use a conversational interface, i.e. a chatbot, instead of a form. The chatbot asks its users questions one after the other and always provides them with certain answer options in the form of buttons.

AI-based chatbots

In many cases, guiding users through a predefined dialog is not enough, because users ask their questions in an uncoordinated order and the questions and answers are much broader. In this case, the chatbot needs a "Natural Language Understanding (NLU)" component. This means that the chatbot uses some kind of artificial intelligence to understand the user.

Users ask all their questions to the chatbot and the chatbot tries to understand the intention of the user in order to give him an appropriate answer.

Example sentences are used to simplify the process of intent detection. During the development phase, a variety of questions that the user might ask are defined. Accordingly, the questions are divided into different intent categories. For each intent, there is an answer or a solution in the database. The chatbot then compares the user question with one of its sample sentences and tries to understand the user's intent.

It is recommended to define at least 15 possible questions per intent, the more the better. These question catalogs are then infinitely expanded in further development phases. The more users use the chatbot, the higher the variation of questions. Developers track all the questions that users ask the chatbot and can then maintain them in the database.

Chatbot hybrid

A mixed form is also possible. Here, the chatbot uses NLP and users can enter their questions via free text, but at the same time users always have the chance to choose between predefined answers.

When users use predefined answers, usually in the form of buttons, the chatbot acts more like the tree diagram described above. The advantage here is that there can be no misunderstandings - the chatbot knows an answer to every predefined question.

At the same time, however, the user also has the option of entering a text himself in this hybrid form. The chatbot will then use NLP to try to understand what the user wanted to say and answer as appropriately as possible. If the bot does not understand the user's free text, the user has the option of selecting one of the predefined answers in this hybrid form.

This concludes the digression on rule-based and AI-based chatbots.

The implementation of a chatbot with NLP, i.e. an AI-based chatbot, is usually much more complex than the introduction of a so-called click bot. Furthermore, misunderstandings between chatbot and user occur again and again with bots with NLP. This cannot happen with a simple rule-based chatbot.

Companies should therefore carefully consider in advance whether the effort is justified or whether a simple rule-based chatbot would also suffice. If it is only about generating new leads on the website or achieving conversions through a chatbot, a simple click bot is sufficient in many cases.

One example of this is the chatbot on the Website of the "AI for Business" conference. This should only generate new registrations for the newsletter. This does not require a great AI component, but only a simple chatbot that motivates the user to register by means of cleverly defined buttons.

Another example of a simple chatbot is a digital assistant on the website that is only supposed to answer the most frequent questions. Provided that the most frequent questions of the website visitors are already clear anyway, the bot can ask its user specifically on which topic he has a question and then guide him again to the appropriate answer by means of buttons. An example of this is the Raiffeisen Aarburg chatbot. Here, the chatbot suggests possible topics to the user directly at the first use and the user can then select his topic area by means of buttons and gets further information.

Another chatbot example is Ava from Air Asia. Ava is one of the more clever chatbots and answers questions using NLP. In practice, this means that users type their question into the chatbot via an input field and Ava then provides the appropriate answer. The questions can be very broad. Whether questions about the food on board, the ongoing flu, flight bookings, reservations or more - Ava understands the user's intent in most cases and also gives helpful answers.

But of course there are always cases when chatbots like Ava do not understand their users. Here, the developers have to build in a so-called fall-back.

This means that the developers must define in advance how the chatbot should react if it has not understood its user. It is common for the bot to first ask the user to repeat the questions in other words. However, if this does not change the bot's understanding either, the programmers have the following options:

• The chat dialog goes without delay to a customer service representative who answers in chat or on the phone (hybrid chat).
• The chatbot writes to the user that it did not understand him and records his contact data. The contact data, including the user's request, is then forwarded to the internal service team, which then contacts the user personally.
• If no connection to human employees is desired, the chatbot can only point to further help pages, community forums or similar.

Before companies start implementing their chatbot and choosing the right chatbot software, they should always think about the complexity and goals of the chatbot. If the chatbot is to become a comprehensive digital advisor in the future, it is probably worth investing in an AI-based chatbot.

However, if a company would rather gain initial experience with the new technology and the use case or the user questions are initially still highly predictable, a simple click bot may be sufficient. In the meantime, there is even a Swiss providerwhich, as a software-as-service provider, hosts all data in Switzerland.

The Chatbot Canvas is a possible tool to find out which way to go. It serves as a template for creating a chatbot concept. In doing so, you then also have to decide how much AI should be in your chatbot.