In the spring of 2017, BEKK developed a Proof of Concept for the Norwegian Tax Authority (Skatteetaten) that explored the possibility of automatically generating tax deduction cards (skattekort) using Machine Learning. This was followed by another Proof of Concept in the summer of 2017 in which we aimed to automatically detect former residents of Norway who had emigrated without notifying the government. In this talk we will present the knowledge and experiences we gained from these Proofs of Concept. How do you organize the development of a Machine Learning PoC? How does the development process for a Machine Learning application differ from that of an ordinary enterprise application? And what (unexpected) problems did we encounter on the way to two successful Proofs of Concept?