Implementation science asks questions about the delivery of interventions to those who need them. Technologies such as drugs, vaccines, or surgical procedures may be the ‘direct mechanism’ of the interventions in question. But implementation science trials answer questions about how best to implement these, about the practice of their delivery, or about the systems through which they are delivered. As such, these trials are inevitably more about human behaviour and system function than they are about the biological efficacy of the drug, vaccine or procedure. This presentation will discuss four implications of this perspective for the design of implementation science trials, illustrating key points with examples. First, such trials are usually most appropriately conducted as cluster-level trials since this is the level at which delivery packages and strategies operate. Second they should be ‘pragmatic’, as defined in the CONSORT guidance for pragmatic trials. One implication of this is that it is important that they enrol study participants who are representative of the intended target population, rather than opportunistically recruiting selected populations as may be appropriate for individual-level efficacy trials. A further implication of ‘pragmatism’ is that, intervention delivery should be monitored and supervised in a realistic way, not closely controlled to ensure ideal implementation as in efficacy trials. Third, integrated process evaluation, ideally guided by the UK MRC Process Evaluation framework for complex interventions, is an essential component of such trials. This framework emphasises the importance of documenting the implementation of interventions, analysing pathways of change and identifying contextual factors relevant to future policy recommendations about scale-up to other settings. Fourth and finally, owing to lack of investigator control, ethical or other reasons, it may not be feasible to undertake randomisation in some implementation science situations. While a full discussion of the range of design alternatives is beyond the scope of this presentation, I will briefly discuss a framework we have developed to map the most rigorous non-randomised impact evaluation designs against constraints to randomisation often described by practitioners in implementation settings.