Digitalhaus Franken: Using AI to predict subscriber churn

Project: PULSE

Newsroom size: 51 - 100

Solution: An AI-powered tool that predicts and prevents subscriber churn to improve retention and strengthen audience loyalty.

Digitalhaus Franken, a German digital media house, faced a challenge familiar to many publishers navigating the shift from print to digital: while they successfully attracted new digital subscribers, keeping them proved difficult. This led to Project PULSE, an AI-powered initiative to predict and prevent subscriber churn.

The problem: High acquisition, poor retention

As digital subscriptions became increasingly central to their business model, Digitalhaus Franken discovered a troubling pattern. "The beginning of our funnel worked quite well – we had lots of conversions," explains Lisa Riech, the project lead. "But we struggled to keep subscribers long-term."

The company recognised a fundamental truth about subscription businesses: retaining existing customers costs significantly less than acquiring new ones. They'd addressed retention before, but never with AI's predictive capabilities. "This was our first time using prediction models," Riech notes. "We wanted to identify customers at risk of churning before they even realised they wanted to cancel."

Building the solution: From consensus to action

Unlike many organisations hesitant about AI adoption, Digitalhaus Franken found immediate buy-in across the company. The proliferation of various generative AI tools had already familiarised staff with AI's potential.

"Everyone was on board," recalls Riech. "We all had the main aim of delivering an MVP, but equally important was learning how to manage machine learning projects."

The team distinguished early between what they called "classic AI" (machine learning and deep learning) and generative AI, helping set appropriate expectations for this prediction-focused project.

The new workflow

Project PULSE short for Predictive User Loyalty System for Engagement, unfolded across several carefully planned milestones. The first phase, lasting until mid-April, focused entirely on data exploration and integration. The team needed to understand what data they had, which systems housed it, and what would actually improve their model's predictions.

"We learned that throwing all available data into a machine learning model doesn't necessarily produce better results," Riech explains. "You need to experiment with different data combinations."

The team tested three machine learning models: XGBoost, Random Forest, and Support Vector Machine. After extensive experimentation, they selected XGBoost for its superior explainability. "With XGBoost, we can examine predictions for each customer ID (anonymised) and see how changing one value affects their churn risk," Riech notes.

To make the predictive insights actionable, the team decided to visualise both the input data and the calculated churn probabilities in a Power BI dashboard. This approach allows the entire customer base to be segmented according to their churn risk, ranging from "very low" to "very high." In addition to current predictions, the dashboard also presents historical data, illustrating how customer engagement – such as pages read or time spent – along with churn probability, has evolved over time. This perspective is crucial for evaluating the effectiveness of marketing initiatives aimed at reducing churn, enabling the team to measure impact and adjust strategies based on real results.

The technical team remained deliberately small: a project lead, data manager, two data engineers, and the head of AI, supplemented by external partners who provided two data scientists and one data engineer. This external expertise proved essential, handling approximately 50 - 60% of the technical work, particularly the model selection and experimentation phases.

What worked well

Communication, often a stumbling block in technical projects, functioned smoothly. Weekly meetings with external partners created accountability and momentum. "They asked questions every week, and we needed good answers," Riech recalls. "This pushed us to stay prepared and move quickly."

The decision to maintain a small, technically focused core team proved crucial. Initially, more people from different departments were included, but the team soon recognised the efficiency gains from keeping the group homogeneous. "When a project is predominantly technical, ensuring clear communication with non-technical colleagues involves considerable effort in translating technical concepts. Working with a fully technical team has allowed for faster execution and more efficient workflows," Riech observes.

Data quality, an anticipated challenge, proved sufficient despite initial concerns about not having "millions of customer records”.

Building sustainable capabilities

Crucially, the external partners built a foundation the internal team could maintain independently. "They established the model, and afterwards we set up the initial clusters and values. Now we just feed in new data – it’s a ten-minute process," explains Riech. "We only need external help if we want to fundamentally change the model."

This sustainability extends to the technical infrastructure. The internal data engineering team handled all data formatting and provision throughout the project, ensuring they could continue operations post-launch

The opportunities: A roadmap for media organisations

Project PULSE succeeded in creating a functional churn prediction system, but as Riech emphasises: "The project provides a good foundation, but now the real work with customers begins."

The marketing team must now leverage these predictions to craft targeted retention campaigns. 

For media organisations considering similar projects, Digitalhaus Franken's experience offers an encouraging message: with clear objectives, the right expertise, and strong project management, AI-powered subscriber retention is achievable – even without millions in customer data or massive internal AI teams.

Lessons for newsrooms

The project revealed three key insights for Digitalhaus Franken and other media organisations considering similar initiatives:

• Need dedicated data science capabilities: "We have plenty of people with data engineering and management backgrounds, but we're actually missing the data science component," Riech acknowledges. "If you want to pursue prediction AI projects on a large scale, you need this expertise in-house."

• Tap 'classic AI' for prediction: Machine learning offers untapped potential beyond generative AI's current dominance. "There's significant value in 'classic AI' for prediction models that we're not fully utilising yet," notes Riech.

• Prioritise exceptional project management: Clear roles, defined timelines, and focused teams matter more than having cutting-edge technology.