This paper analyses data from a curriculum intervention project designed to introduce new forms of discussion, seen as educationally effective, into the primary classroom. While the introduction of talk as an aid to learning is premised on a social approach to learning, such interventions are often evaluated in terms of cognitive benefits and gains in subject knowledge, and considered independently of social and interpersonal factors that affect the conduct of interaction. The study discussed in this paper is intended to complement such cognitively-based approaches. I argue that it is difficult to ‘bracket off’ official educational activity within interactions, and that research needs to take into account how talk is used to a range of effects in specific interactions, as well as differences between pupils in their take up of the communicative strategies they have been taught. I suggest also that the context-dependency, variability and potential ambiguity of utterance meaning raises important issues for the implementation and evaluation of projects designed to introduce particular ways of speaking within education. These points are illustrated by the analysis of transcripts of small-group discussions between pupils taking part in a recent intervention project.