Initial insurance losses are often reported with a textual description of the claim. The claims manager must determine the adequate case reserve for each known claim. In this paper, we present a framework for predicting the amount of loss given a textual description of the claim using a large number of words found in the descriptions. Prior work has focused on classifying insurance claims based on keywords selected by a human expert, whereas in this paper the focus is on loss amount prediction with automatic word selection. In order to transform words into numeric vectors, we use word cosine similarities and word embedding matrices. When we consider all unique words found in the training dataset and impose a generalised additive model to the resulting explanatory variables, the resulting design matrix is high dimensional. For this reason, we use a group lasso penalty to reduce the number of coefficients in the model. The scalable, analytical framework proposed provides for a parsimonious and interpretable model. Finally, we discuss the implications of the analysis, including how the framework may be used by an insurance company and how the interpretation of the covariates can lead to significant policy change. The code can be found in the TAGAM R package (github.com/scottmanski/TAGAM).