The ability to predict the consequences of an atmospheric release of a toxic material accurately, as opposed to conservatively, is of increasing importance to decision-makers who need to plan for or respond to such incidents. A conservative estimate of casualties, for example, may involve multiplying the average population density by the area over which a conservatively-chosen threshold atmospheric concentration is exceeded on average. We have begun to investigate the requirements for producing accurate consequence estimates using modern atmospheric dispersion models, including the capabilities and limitations of different end-to-end consequence modelling approaches. Some issues of concern include the use of ensemble-average concentration predictions in conjunction with toxicity models, especially advanced models involving the toxic load; the fact that different methods exist for calculating the toxic load from time-varying exposures but none have been experimentally validated; and the potentially inapt comparison of atmospheric dispersion model predictions to the consequences of real-world historical incidents. The work presented here focuses primarily on the problems associated with the choice of toxic load model and the potential use of individual realizations of a release in place of ensemble-average concentration predictions for consequence estimation. Our study of individual releases from the FUSION Field Trial 2007 (FFT07) experiment and from VTHREAT- generated predictions of a small-scale chemical attack indicate that casualty estimates can have a substantial dependence on the choice of toxicity model. We have also probed the limitations of using ensemble-average concentration predictions for casualty estimation and the potential use of low-order concentration moments to improve those estimates.