This paper is a case study of the validity of different data sources and modelling approaches commonly used to build biokinetic models for radionuclides. The paper examines the basis and apparent predictive accuracy of each of the biokinetic models for Pu used over the years by the International Commission on Radiological Protection (ICRP), in view of recent improvements in the database. The results of this study and similar retrospective studies for other elements suggest the following five points. (1) Extrapolation of biokinetic data from laboratory animals to man is particularly uncertain for the liver due to qualitative differences among species in the handling of many elements by this organ. (2) As a 'default' approach to biokinetic modelling or model assessment, biokinetic data for unhealthy human subjects should be given higher weight than information extrapolated across animal species or chemical families, but there are counter-examples. (3) Little confidence can be placed in biokinetic model predictions for long times after exposure based solely on curve fits to short-term biokinetic data. (4) Bioassay and dosimetry models for a radionuclide should not be developed separately. (5) Where feasible, the systemic biokinetic model for a radionuclide should be developed within a physiologically realistic model structure, because this allows biokinetic data from experimental studies to be supplemented with physiological information, provides a basis for extrapolation of data across animals species or chemical families, results in models that can be used for both bioassay interpretation and dosimetry, and provides a logical basis for extrapolation of data to subgroups of the population (e.g., various ages) or to times outside the period of observation.