Artificial Intelligence (AI) changes the way data can be produced, processed, and collected. Not only does AI change the way data is processed in a vehicle: AI does not only consume data, it also continuously generates new, derived or synthetic data whose availability and semantics depend on the resultant model itself. Dynamic data preprocessing already challenges the notion of a static data catalogue for an automotive telematics system. AI further challenges this notion: changing an AI model can directly influence which data exists and can be collected. While the Vehicle Signal Specification (VSS) has a notion of derived data, its ability to describe how data is derived is limited. Due to their opaque and non-deterministic nature, AI models make the task of describing how derivation takes place nearly impossible. Add in the need for more complex data such as pictures, video, and lidar data, and it becomes clear that not all data can be transmitted all the time. This talk will discuss these challenges and present use cases that motivate them. Using concrete use cases, it demonstrates why VSS only scratches the surface of the problem and why selective, purpose-driven data collection is essential. Building and maintaining AI models require vast amount of data, both for accomplishing its tasks, as well as for retraining. However, training, inference, validation, and retraining have fundamentally different data requirements. Determining what data is needed where and how often is key to success. It is also key to a real Software Defined Vehicle. Takeaways: Data collection use cases with AI. Challenges of data collection in conjunction with AI. Why selective data collection is an essential element of any software defined vehicle. How collecting data selectively can improve support for model maintenance and scalability.