This article describes some dilemmas and problems encountered in a Delphi study of general practitioners’ (GPs) information requirements. The research involved a three-round Delphi administered to an expert panel of 50 GPs in one Welsh health authority area. The Delphi is generally perceived as a qualitative method. However, we argue that the requirements of the technique make it difficult to sustain the kind of inductive analysis – faithfully reflecting respondents’ perspectives – that is axiomatic to many of the theoretical approaches that underpin qualitative inquiry. We describe how our attempts to incorporate respondents’ views in near-verbatim form in the first round were undermined by the need to classify and reduce statements for later rounds, and to impose judgments about what should count as consensus. The iterative ‘consensus-building’ process, that is so central to the Delphi, was difficult to take forward without active intervention by the research team, but this involved a re-ordering and reduction of the data, which moved the statements included in later rounds further and further from the verbatim responses on which they were based. Whilst the findings of the study were useful on one level, the final consensus statements took a general, virtually context-free form, which contained few references to background preoccupations (largely concerned with the recent NHS reforms) that were exercising GPs at this time. The method also coped badly with the polarized opinions that existed regarding the development of information systems for commissioning. We conclude that the Delphi is best used for large-scale research in areas where opinions are well established, where the problems and choices confronting the study group are well known, and where a major programme of organizational reform is not already underway.