Real world data (RWD) has become an important tool in pharmaceutical research and development. Generated every time patients interact with the healthcare system when diagnoses are developed and medical interventions are selected, RWD are massive and in many regards typical big data. The use of artificial intelligence (AI) to analyze RWD seems an obvious choice. It promises new insights into medical need, drivers of diseases, and new opportunities for pharmacological interventions. When put into practice RWD analyses are challenging. The distributed generation of data, under sub-optimally standardized conditions in a patient-oriented but not information maximizing healthcare transaction, leads to a high level of sparseness and uncontrolled biases. We discuss why this needs to be addressed independent of the type of analysis approach. While classical statistical analysis and modeling approaches provide a rigorous framework for the handling of bias and sparseness, AI methods are not necessarily suited when applied naively. Special precautions need to be taken from choice of method until interpretation of results to prevent potentially harmful fallacies. The conscious use of prior medical subject matter expertise may also be required. Based on typical application examples we illustrate challenges and methodological considerations.