The ability of 'digital agriculture' to support on-farm decision making is predicated on the real-time combination of observations and prior knowledge into an integrated digital environment. The mathematical discipline that seeks to provide this integration is known as model data assimilation (DA), with demonstrated benefits including improved predictive reliability, and the capacity to identify unexpected changes in field conditions and potential measurement errors. Despite routine adoption in other fields, the delayed adoption of DA in agriculture is due to the need to express end-of-season outcomes such as yield, update forecasts of these outcomes throughout the growing season as data become available, and enhance forecast reliability. To overcome these challenges, three guiding principles are introduced, providing a means to operationalize crop model DA for robust on-farm decision support. We apply the guiding principles using a South Australian viticulture case study. Our case study involves application of an iterative form of a widely used DA algorithm (ensemble Kalman filter) to dynamically update both static parameters and states associated with a grapevine simulation model. Daily weather data as well as fortnightly ground-based leaf area index (LAI) data are used for assimilation. It is shown how crop model DA can lead to not only significant improvements in forecasts of LAI but also to forecasts of end-of-season yield. The guiding principles also enable observations of greatest value to be identified throughout the season. This study highlights the role that formal crop model DA can play in agricultural decision support through enhancing situational awareness in real time.