BLOG

The AI experience: Combatting churn

New customer acquisition is expensive and the longer the customer stays with you, the higher lifetime value you can expect with them. When managing your customer turnover is a daunting task, it is time to turn to AI for help. Churn modeling aims to predict the probability of your existing customers leaving you, and possibly also deducing the most likely reason for that.

There is no arguing that returning and loyal customers are highly important for most companies. Whether it is a subscription-based service or fast-moving consumer goods, persistent customers bring revenue without notable marketing costs and possibly even spread the word about your excellent product. So, in the end, churn modeling is about better customer experience and customer service.

Turning to AI and machine learning for help comes into question when your customer base is large, and you do not have a strong in-person connection with your customers. As a rule of thumb: if you do not know why your customers leave and do not have a straightforward way of finding out, it most likely requires at least some machine learning to dig up the answers. Once you have a solid customer churn model, you have the information you need to take action and start making a difference.

Getting started

This is true in many AI endeavors, but especially with churn modeling: the first step is to define the goal you want to achieve. What actions are you planning to take? What do you need to know to get there? This is especially important, as the end goal might even govern choosing the tools and models used, but at the very least it keeps the objective clear and boosts your chances for success. The two most common aims for churn modeling are:

Recognizing the most probable churners in your client base
Identifying the key reasons why customers are leaving you

Most often the goal is a combination of these two, with a slight emphasis on either one. For example: “What is the probability of this customer stopping use of my service within the next two months? What is the most probable reason for that?” are excellent questions to include in a machine learning solution.

Just as important as figuring out what you want to know, is what you are planning on doing with the information. Perhaps you wish to better allocate your customer service resources to hesitant customers, give gifts or discounts to nourish the customer relationship, or make your product better by focusing on the underlying issues that have been recognized to cause churn. Whatever the actions are, having them clear in your mind from the beginning makes the whole modeling process a lot smoother.

The art of choosing the right methodology

As said, with churn models the goal often is two-fold: predicting who is about to opt-out but also understanding the reasons why. Models primarily focusing on the first task are predictive, whereas the latter task is often better tackled with descriptive models. Sometimes both needs can be fulfilled with a single model, but more often than not choosing between predictive power and descriptive capabilities is a balancing act and you have to give on one end to gain more on the other. In other words: the most accurate predictions often come from black-box models that are next to impossible to read into. On the other hand, descriptive models are easy for humans to interpret, but they lose out in the accuracy competition. Luckily, there are more and more descriptive methodologies emerging that can be used to decipher black-box prediction models, thus providing you with the best of both worlds.

Dodging some common pitfalls and managing expectations

To be blunt, churn modeling is one of the more difficult AI solutions that you may need. Just like any AI solution, a large challenge lies in the data. Unlike with a recommendation engine or customer base clustering, you may have plenty of high-quality data from your customers, but still, your data might not even remotely reflect the reasons why customers are leaving you. This wreaks havoc on your prediction power. The upside is that complex AI solutions can find complex connections and subtle hints completely invisible to us. So even if we think the data we have cannot possibly predict churn, it very well might. Here, data quantity is of the essence, as the danger of overfitting is always lurking just around the corner: in the end, a churn model is only as good as it is in predicting future churners. A good dose of realism comes in handy with churn models: do not expect to get it perfect on the first try and remember that even a low accuracy model can be many times better than guessing.

  • There is only so much you can do with “data scientist magic”: keep your eyes open for new data sources. New data is often one of the best ways to notably lift model performance.

  • Churn modeling does not have to be exceptionally accurate to be beneficial: even a low accuracy model can be multitudes better than nothing.

  • Before diving into the deep end of sophisticated predictive AI models, conducting a customer base analysis might come in handy: a deeper understanding of your customer base gives valuable input for example regarding the type of data your predictive churn model might need to succeed.

Wrap-up with few key takeaways

From defining the goals to finding appropriate data and selecting the most suitable model, building a working churn model requires clear objectives and dedication. However, getting the model right is only half the work. Even the most accurate predictions are worth nothing, if there are no actions taken or if the actions work against customer retention. Knowing your customers is key when choosing the actions; after all, you probably do not want to annoy them into canceling their subscription by bombarding them with irrelevant emails.

Where churn modeling can be an invaluable tool in your toolbox, 1+1 equals more than two when it comes to building exceptional customer experience with advanced analytics and machine learning solutions. For example, functioning and continuous customer clustering helps you understand the probable churners a lot better than mere churn prediction, which in turn helps to find smart actions. There are always lost causes and low-value customers and recognizing them enables redirecting scarce resources to where they bring the most value. Customer clustering can also help you identify and understand the reasons for churn in situations where sufficient descriptive models are not possible, or the necessary data is not available.

Liked this post? Share it with your friends
Share on facebook
Share on twitter
Share on linkedin
Tiina Kangasniemi
January 21, 2021

I personally like to say that “don’t choose an AI solution, choose the BEST solution. Sometimes that best solution is an AI solution”, meaning that however nice and shiny AI can be, they are not always the smartest thing to do. But sometimes they are 😏.