Should I base my customer service Chatbot on Machine Learning?

Should I base my customer service Chatbot on Machine Learning?

A few weeks ago I decided to follow an online course taught by Amazon Web Services for free. I was curious to know more about a concept that is on everyone’s lips lately, especially when it comes to talking about Artificial Intelligence or digital transformation: Machine Learning. From the introductory video, I have been able to build this definition of Machine Learning:

Machine Learning consists of processing and analyzing an immense volume of data in order to make predictions of behavior, especially for use cases related to sales conversion and web.


In fact, this definition fully coincides with an Amazon practice, this time as a global product sales website, which we all have experienced: under the article we are about to buy, the section introduced by the phrase “Customers also shopped for”, which provides other articles of possible interest. The introductory phrase of this section perfectly reflects the procedure carried out: Amazon collects millions of behaviors of its users in terms of purchasing processes, analyzes them in order to predict the future behavior of their customers, and with high probability manages to increase sales on the web.


There is no doubt that this methodology turns out to be an extremely powerful weapon when it comes to doing directed marketing, forecasting demand or of course turning visits to the web into sales. But what happens if the same strategy is applied to customer support?

In the context of a customer service channel, such as a Chatbot, it is important to stop and think about the concept of predictions, which is not compatible with the accuracy or precision required by the support reality. While a recommended sale prediction does not entail major risks — if the sale converts, that’s fine, if it does not convert, that’s fine too -, the answers given to users through an official support channel of the company can be highly sensitive: they need to be reliable, accurate and contextualized, both legally and commercially.


The Chatbot is nothing else than an additional channel to solve doubts about airline tickets, check contractual aspects of a bank account, download manuals of household appliances, consult data of a hotel reservation or the balance of a telephone line … The knowledge about the company provided through the Chatbot must not differ from the rest of the website or from the other customer service channels, whether these channels are automated or managed by human agents.


A wrong answer, especially in reply to a delicate question, could have disastrous consequences. Imagine that a chatbot providing support to the policyholders on the website of an insurance company, in reply to a question about breast leads the user to answers about plastic surgery, because they are more common according to Machine Learning, instead of offering one or more answers about pathologies, gynecology, etc. The image of this insurance company would be damaged due to the low reliability of the Chatbot, more precisely due to the misuse of a predictive methodology, such as Machine Learning, with the purpose of customer care.

In addition, to fulfill this purpose, the essential requirement of Machine Learning is to have a huge volume of historical data to process and analyze, from which to build predictive behavior models. Nowadays, some macro-companies have these kinds of volumes of data, after taking years monitoring their customer support channels and processing the information manually in order to rationalize and classify it. However, the truth is that most companies do not have these traces of activity from their clients, therefore they can not benefit from Machine Learning in a simple way.


The alternative that remains is to create data manually, as if they were traces collected from users, to “train” an ad-hoc Machine Learning model. This training process is highly promoted by numerous chatbot providers, to the point that it has created the false belief that the only way to implement a chatbot is to train the system manually, for months, with as many utterances as possible — considering as utterances the many different ways of expressing the same intention. This myth is fed by those who do not have Natural Language Processing technology to relieve the lack of previous data. Likewise, the manual creation of data to fit with Machine Learning models implies a colossal effort for companies, an effort which translates into both a prolongation of the project’s time-to-market and disenchantment with Artificial Intelligence.


On the contrary, a provider capable of using the Natural Language allows solving the two problems that Machine Learning entails for a Chatbot customer service: on the one hand, it does not require extensive volumes of data, on the other it is based on linguistic engineering for offer semantically relevant, contextualized and highly reliable answers.

This does not mean that Machine Learning is an evil approach but rather that it is not always the most appropriate approach depending on the use case. While to boost sales on the web with recommendations it seems to be the most appropriate methodology — and with demonstrable success — for something as sensitive as a customer service Chatbot, it is likely to be incorrect and lack of precision, disadvantages that end up hurting the image of the companies.


Betting on Natural Language Processing, on the other hand, allows to guarantee the reliability of the Chatbot, thus the reliability of the company, as well as favoring an agile time-to-market and trust towards Artificial Intelligence.


Looking at this panorama, the ideal approach could be to bring together the best of each world: fully rely on the Natural Language Processing as the main technology of the Chatbot for customer support, and at the same time resort to some features of Machine Learning to complete certain purposes, such as analyzing customer interactions to suggest new contents for the knowledge base, or using clickthrough statistics to adapt the interface and improve the user experience.

Article Source:

Leave a Comment