How artificial intelligence helps to process customer complaints

Mr. Behrendt, you are working on how Porsche Consulting can use artificial intelligence and data analysis methods to help companies evaluate data in order to gain useful insights. First of all, what kind of data are you actually working with?

Behrendt: It's impossible to say in general terms; it always depends on the specific project. In addition to structured data, which is available in tabular form and often contains numerical values, there are also more and more projects that involve unstructured data in the form of text. Completely different methods, especially from the field of Natural Language Processing (NLP), are used here. In general, data availability and quality is often the first major obstacle in projects. Even the best models unfortunately bring little or no added value without a clean database.

When you think of data, you usually think of Excel spreadsheets. But what kind of data is there in text form?

Behrendt: The majority of company data is actually in text form. For example, we once looked at a company's customer complaints that were collected in the form of e-mails. With a critical mass of such customer complaints, it is very time-consuming for customer service employees to process the customer complaints promptly. This is where NLP methods can help to automatically classify emails and route them to the right places so that customer service can do its job even more efficiently and the company can ultimately increase customer satisfaction. In addition to the right methodology, the second step is to ensure that the solution can be operationalized - in other words, that it can be used in day-to-day operations. Without anyone having to laboriously execute individual scripts with code. This is definitely a change when you come from a scientific background, where you often pay less attention to such things.

Can you explain to me in more detail what exactly happens in natural language processing? Does the artificial intelligence understand what is written in the emails?

Behrendt: No, the AI doesn't understand it in a cognitive sense. You can think of it as performing various mathematical operations based on the available data. In the classification of complaint emails, this means in concrete terms that the texts have to be converted into a numerical representation in the first step - in the end, a given text should be expressed in the form of a vector whose entries are numbers. This is also referred to as a semantic representation. Texts - expressed as vectors - should be close to each other in the semantic space if they are similar in meaning. If we imagine, for example, customer complaints about an airline, two complaints about lost luggage would be closer together than a complaint about lost luggage and another complaint about a flight delay. Such semantic representations can then be used to classify emails fairly accurately.
NLP methods can also be used to generate new data that can be used for other analyses. For example, you can analyze the sentiment of investors via social media posts and then integrate it into forecasting models - that was still part of my doctoral thesis at the time. NLP methods therefore always pave the way for the other areas that our team deals with: Causal analysis, predictive models and mathematical optimization.

NLP methods are therefore good for analysing and categorizing data. What about the next step: responding to customer complaints? With corporations, you often get the feeling that you are not dealing with people at the beginning, but with machines that provide you with ready-made answers. Can possible answers also be linked to inquiries in the background via the vectors?

Behrendt: Exactly, you can of course also use NLP methods in this case to generate answers to customer complaints. The system has to be trained on complaint-response pairs written by humans. Above all, however, it is important to ensure that the answers are followed by appropriate action - e.g. in the case of recompensation payments - and that the answers are factually correct. It would be unfavorable, for example, if the correct telephone number was not given by the service center.

If we disregard customer management: In which other areas can artificial intelligence be used to analyze data?

Behrendt: Anywhere where sufficient data is generated. From a company's point of view, however, it is always crucial to consider a very specific use case that also generates a clear benefit. On the one hand, you should avoid implementing a use case just because the term "artificial intelligence" is mentioned in the title without being able to highlight the exact benefits. On the other hand, however, it is also important to overcome the blockade that arises from the fact that it is often not at all clear what is behind the "artificial intelligence" that is supposed to make a use case so special. A realistic assessment is important here in order to be able to weigh up the opportunities and challenges.

As already mentioned, however, the availability and quality of data must first be ensured - which is easier said than done, as this is not just about technical aspects, but also governance aspects. Depending on the company, a wide variety of use cases can then be derived from a solid data basis: from forecasting product demand to the optimal utilization of a transport network to checking financial transactions for suspected money laundering.