In large-eddy simulations (LESs) of atmospheric reacting flows, homogeneous and instantaneous mixing of reactants within a grid-cell is usually assumed. However, highly reactive species are often segregated or pre-mixed at small scales. In this paper, we propose a parameterization to account for the effect of the unresolved scales on the chemical transformations. Its formulation relies on the description of the subgrid unresolved reactant covariance as a function of the resolved covariance by using scale-similarity arguments. A dynamic procedure is used to compute the model coefficient from the resolved reactant concentration fields, therefore not requiring any parameter specification or tuning. In simulations of a convective boundary layer with a fast second-order reaction, using the new model is found to perform better than ignoring subgrid chemistry effects.