AbstractThe concept of operational research in practical agriculture is introduced. Three aims of this approach are discussed and illustrated with actual examples in ley‐dairy farming‐to help understand and manage dynamic systems, such as growth‐to explain interactions and make predictions for current management measures, such as harvest time.‐to find alternative input combinations by sensitivity analyses.Prediction of ley growth and change in nutritional value illustrates the first point. It is shown that prediction of growth requires real time weather data but not necessarily weather forecasts. Alternatively, these can be substituted for data on normal weather conditions and some more extrem conditions of known probability.The second point is illustrated by analysing the main interactions that result in the margin above feed costs in the ley‐ruminant system. It is claimed that if the model satisfies an observed farm situation, as in this case, the explanation of the interactions is valid.The third point ‐ sensitivity analyses ‐ is illustrated by the higher margin over feed costs that is predicted in forage‐based feeding plans with 12‐15 kg DM forage, as compared to conventional feeding plans with 8 kg DM forage.The Discussion centres on (1) deterministic and stochastic models, (2) scale and experimental replication, and (3) statistical problems in parameter estimates and model validation:(1) When very dynamic systems are encountered, deterministic models may be difficult to validate. It is then far better to develop less precise stochastic models than to abolish the idea of model development. Such models would still predict risks and uncertainties.(2) With large systems such as production chains, experimental replicates often become too costly. A validated systems model then provides the possibility to explain unreplicated observations and state their generality.(3) Parameter estimates in mechanistic models require in many cases statistical developments because of problems encountered with non‐linear models where parameters may not be normally distributed.Requirements on model validation may differ when the models are used for long‐term or short‐term planning. In the second case, the model may be repeatedly calibrated to a current situation, and validation then includes the entire calibration procedure.