Chatbots have been heralded as great saviors and often fall short of expectations. In our report, you could learn that there is another way and that there are six steps to go through for successful chatbot development.
The euphoria in the entire western technology world was huge when Mark Zuckerberg announced the availability of bots on Facebook Messenger in April 2016. Bots were already known from China, for example at WeChat, but Facebook still brought the necessary reach in the western world from almost one billion users at the time. A technology that seems mature enough, promises companies cost savings and at the same time also serves the increased need of customers for self-service offers. In other words, a "win-win-win" combination.
However, almost three years later, a sobering scenario emerged. Chatbots don't understand customer requests, development costs exceed expectations, and users are stuck in endless loops. What to do to avoid these failures?
The following six steps must be observed for successful development:
Step 1: Find the right technology foundation
When you are initially interested in the topic of chatbots, you will first encounter a rapidly growing market. A large number of technology providers and start-ups appeared, all of which promised the best AI ("Artificial Intelligence") and the best bots. Differences in quality are often difficult to identify at first glance and every demo always works smoothly. Providers usually differ in the technological approach (mathematical or semantic), in hosting (on-premise or cloud), in the business model (license or open-source), and in the size of the company (startup or large corporation).
To objectively select a suitable partner from this, you should better do a small prototype test. To do this, select a provider that covers everything you need. Test the chatbot on the next parameters such as recognition rate (how often the chatbot recognizes the customer's concern) and depth of conversation (how far the chatbot can process the concern).
Step 2: Identify the right use cases
After the right development partner had been found, the task was to define use cases for the bot. Bots can solve such cases that can be delimited. Besides, cases should also be able to be processed end-to-end. For example, if the customer wants to change his address, the change of address is first recognized by the bot and then recorded directly in the system.
A first good starting point is the analysis of previous customer chats for recognition of content patterns and frequency. In addition to the pure statistical analysis, it is imperative to include colleagues from the operational business in the selection and prioritization of use cases. The wealth of experience helps to identify cases that occur frequently, but can be dealt with very quickly and therefore do not pose any problem in day-to-day business. To enable the promising “win-win-win” situation, the focus should be on those cases that create added value for both the customer and the employee.
Step 3: Determine the basic design
An important decision in the development of chatbots is the definition of possible user interaction. It is possible to offer the user a free text field or to guide him through the dialog using a logic tree with questions and buttons. Free text is the much more complicated variant, as the chatbot cannot predict the user's reaction/answer and must understand the language. However, it is precisely the understanding of speech that is the decisive point of the technology and offers the possibility of a real conversation. Alternatively, a decision tree can imitate a conversation well and is very secure compared to free communication, since the chatbot does not have to deal with unknown user reactions.
Of course, new technology cannot function correctly from day 1, but it must also be well tested for a realistic picture. Therefore, there is initially a conflict of objectives between testing the technology on the customer and the high-quality requirements in live operation. A good idea is to opt for a hybrid solution to start with. Hybrid means that the chatbot has been used, but does not communicate directly with the customer, initially supporting the employees in their work. The chatbot answers the customer's question and also writes the message in the customer service representative's answer field, but the employee still has to finally approve this message and send it to the customer. If an answer is not correct, the employee can intervene at any time and take over the chat. This hybrid approach has the advantage that the chatbot can be tested in real dialogue and the quality can be checked at the same time.
Step 4: Convert user concerns into technical intents, entities, and utterances
So-called intents, entities, and utterances are required in the technical development of the algorithm. Intents describe the concerns of a user that happen the most often. For example, a user may have forgotten his password for the login area. “Forgot password” represents an intent in this example. The algorithm now needs as many examples sentences as possible for this intent to be able to reliably recognize it in the future. The sentence “I forgot my password for my account with the email “email@example.com” would be a suitable example here. In operation, the chatbot tries to recognize these example sentences in the customer's concern and thus to identify the intent. If the number of example sentences increases, the chatbot “learns” and is better at recognizing it.
Entities are variables in a matter. The algorithm takes the e-mail address "firstname.lastname@example.org " as the entity from the example. Entities are typically places, addresses, dates of birth, etc. Ultimately, there is also at least one defined utterance for each intent. This represents, the answer, or the next step. A meaningful utterance could be “Thank you for your message. Your password is MrSimpsonQwerty!”. When developing the algorithm, you always have the option of defining the best answer for a specific incident (intent) or several answers, from which the algorithm can choose.
Step 5: Let the bot “learn” and keep it up to date
In the context of bots and AI, one hears the term "self-learning system" again and again. The term is a bit misleading as it often creates the expectation that the algorithm will just magically get better by itself. A system does not learn completely independently, but always needs some kind of feedback. In the hybrid system, the algorithm could learn that its answer was correct if the employee did not have to intervene. If the service employee intervened, the system noted this response as negative and was able to adapt its response. Of course, other forms of feedback are also conceivable, such as a user evaluation at the end of the chat.
In addition to constant feedback, the chatbot, like every employee, must be informed about process changes – in case if, for example, the answers which the bot has to give has changed due to some internal reasons like a new product.
Step 6: Expand the solution step by step
Integrate and expand the solution step by step so it can prove itself over the few months. We are sure that the quality will increase the level almost as good the direct communication with the customer.
Don’t think that chatbots will magically transform everything from the very beginning – as the bot will learn bit by bit and in the end, you would be able to allow the customer to interact with the bot directly. Then the employee will no longer be needed to approve the bot’s answers, but will only take over the chat from the bot if the bot does not know the answer.
Conclusion: is it worth investing in the technology?
It might be the project which could be characterized as a success story in many ways. At the heart of this is a partnership with an ambitious company like ESKA, in which an innovative solution for operational use will be created.
One thing is clear, however: chatbots are not a sure-fire success and a lot of thoughtful decisions are required in development. If implemented correctly, chatbots can take over a relevant part of future communication and thus also do justice to the initial euphoria about their implementation.