Tired of listening to NLP talks from people working at cool startups, FAANG, or contributing to sexy open source projects? Curious about NLP in a more traditional setting, where data is scarce, domain knowledge is king, and the stakes are high? Then you have come to the right place! In this talk, we present our experience with developing and deploying NLP systems at Fremtind, one of Norway's largest insurance corporations (we know, right? Insurance!). We will discuss challenges such as data scarcity, domain specificity and ad-hoc evaluation, and how we addressed them using different NLP techniques. We walk you through some of our projects, such as data augmentation using weak supervision, customer feedback classification and ranking, and zero-shot classification for dynamically changing label-sets. Finally, we will reflect on the nuances of real-life system evaluation, and how some of our solutions did or didn't change with the increasing ubiquity of LLMs. This talk should peak the interest of bbuzz participants who work in more traditional industries (we exist!), by providing tips and tricks to address concrete and highly relatable NLP challenges. It will also be highly relevant to tool and model developers, as it will provide insights on the actual challenges we face in a setting that is traditionally underrepresented in ""cool"" tech conferences.