This thesis reports on aspects of language use and discourse in staff replies to patient feedback posted on the health service review part of the website NHS Choices. The overall aim of the thesis is to contribute to understanding about how NHS staff use language when communicating with patients in a feedback context, and the reasons for the particular linguistic choices they make. The study uses a corpus-assisted discourse studies (CADS) approach to examine linguistic patterns in datasets based on three staff reply text types derived from an 11.5-million-word corpus of NHS replies. The three datasets are ‘stock replies’ (texts completely or mostly reused in full), ‘unique replies’ (texts that are likely to have been individually written for one-time use) and ‘mixed replies’ (texts that consist of a mix of reused and non-reused elements). This study finds that, while there are linguistic differences between reply types – for example, a greater tendency for those based on text reuse to be more formulaic, and unique replies to entail more variation – these do not predict the interpersonal aspects of replies. Staff replies can be more or less impersonal/personalised irrespective of reply type. In its examination of unique replies, the study highlights a number of patterns contrary to expectations that individualised replies are more personalised, including evidence of indirect criticism and use of discrediting strategies against patients. The latter is a feature of marketised discourse, evidence of which is found across all three reply types. In addition to findings about the language use and discourse of NHS staff, this thesis also presents an original method for using CADS to analyse a corpus containing a high amount of text reuse.