Data Science in the Age of GenAI
Interview with Dr. Timo Klerx
"GenAI makes AI more tangible.
But real data science begins where data turns into better decisions."
Five practical questions for Timo Klerx
Few topics are currently being discussed as intensely as artificial intelligence. With the rise of GenAI, interest in data science, machine learning, and data-driven applications has grown even further. At the same time, new expectations, new misconceptions, and many practical questions are emerging in everyday business contexts.
We spoke with Dr. Timo Klerx, Co-Founder and Chief Data Scientist, who has been working on data science projects in practice for many years. In the interview, he explains what data science really means today, why GenAI changes many things but does not replace everything, and where AI projects in companies often fail.
What do we actually mean by data science today?
Data Science means turning data into usable knowledge and better decisions. This includes collecting, cleaning, analysing, and modelling data using a wide range of methods — from simple rule-based systems to machine learning and GenAI — and integrating the results into real systems and processes.
One example: A retailer uses data science with our forecasting platform prognotix, to predict demand more accurately. This helps plan inventory more effectively, reduce overstock, and avoid out-of-stock situations. The real value, therefore, does not come only from a more accurate forecast, but above all from better operational decisions
Is data science being replaced by GenAI, or is it being redefined?
Neither. GenAI rather makes more visible what data science has actually always been about.
For a long time, there has been a misconception that data science is mainly about building models. This usually refers to the technical part: selecting algorithms, training machine learning models, testing and optimising them. In practice, however, that is only one part of the work.
A much larger part is understanding the data: What data is available? Is it complete? Which patterns are relevant? What question are we actually trying to answer? This also requires domain knowledge, critical thinking, and the ability to interpret results correctly.
What GenAI really changes is that AI becomes more tangible. Many people in companies have experienced for the first time what AI can do themselves, for example through ChatGPT, generated texts, or support with programming. As a result, the barrier to engaging with AI has become much lower. At the same time, expectations have also increased significantly.
Customers often ask: “If AI is so good at writing texts and even coding complex programs, why can’t it also do predictive maintenance?”
That question is completely understandable. AI can support . Predictive Maintenance and, in many cases, make it possible. But it works differently there than it does with text or code.
A large language model mainly works with language, meaning patterns that are available in huge volumes, for example in text and audio. Predictive maintenance, on the other hand, is based on highly specific (usually tabular) machine, sensor, and process data. Every machine has its own operating conditions, its own data structures, and its own patterns when it comes to wear or failures.
This requires clean historical data, such as sensor data, domain knowledge, and an understanding of failure patterns. None of that is solved by a clever prompt. GenAI can support the process, for example by analysing documentation, helping with code creation, or explaining model results. But the actual work remains data science: understanding the data, identifying signals, developing and validating models, and integrating them into real maintenance processes.
Why do many AI projects fail to make it into production?
Many projects fail before the first model is even built. Companies often start with a concrete AI idea, but the actual business problem has not yet been clearly defined. Sometimes they ask for A, while B is the real issue behind it — and C might even create more value as a first step. This is exactly where good data science begins: not with the model, but with the right question. Which decision should be improved? What data is available for this? What quality does that data have? And how will success be measured later?
The second typical mistake is that the project remains a data science experiment for too long. The model works in a notebook on a one-time data export, but it is not clear how it will be supplied with new data on a regular basis, who will use the results, and where exactly they will flow into a real business process. This is often where the move into production fails. A model alone is not yet a solution. It needs stable data pipelines, interfaces to existing systems, monitoring, clear responsibilities, and people who can work with the results. If these things are not considered from the beginning, a good prototype often remains just a demo.
And then there is another point that is often underestimated: IT. If the IT department is only brought in at the end, they have neither helped shape the project nor built trust in it. At the same time, management commitment is needed so that a prototype can become a prioritized, productive solution. Data science should therefore be considered early together with the relevant business departments, IT, and management, so that business needs, technical feasibility, and strategic responsibility come together from the start.
What advice would you give companies that want to get started with AI?
Don’t plan for too long — start. But start with a clearly defined problem.
We Germans tend to over-engineer: the perfect architecture, the perfect data foundation, the perfect use case. But in reality, companies learn the most when they start with a concrete, manageable use case and then develop it step by step into a robust solution.
What matters is not to begin with the biggest and most complex case right away. It may promise the highest value, but it also comes with the greatest uncertainty, from a business, technical, and organisational perspective. A better starting point is a use case that has a clear business case, but can be validated quickly.
From a data science perspective, this means: start small, check the data quality, build a first model, validate the results with the business teams, and think early about how the solution will be used in production later. This creates not only a proof of concept, but also trust in the data, the models, and the processes. Because the first case has to work. If it fails, people are quick to say: “It was obvious that AI wouldn’t work.” If it succeeds, it creates the foundation for larger and more complex projects.
Conclusion
GenAI has changed the way we look at AI. Many things have become more tangible, and many things more accessible. At the same time, it is becoming clearer than ever how important solid data science work remains.
Because between an impressive demo and a productive, reliable solution, there are usually exactly the steps that are often overlooked in public discussions: understanding the data, bringing in domain knowledge, integrating results into real processes, and setting the right priorities.
Companies that want to create real value with AI therefore need more than technological possibilities. Above all, they need a clear view of the specific problem they want to solve. That is exactly where good data science begins.