The purpose of the present study was to investigate a new procedure for establishing accurate discriminations of delayed acts of saying, of doing and discriminations of say–do correspondence and non-correspondence with three developmentally delayed subjects. A corrective feedback procedure for incorrect discrimination responses, that involved multiple-exemplars, was initially employed, but failed to establish most of the target discriminations for all three subjects. A near-errorless training intervention was subsequently employed that also involved multiple-exemplars. This training used two referents (one for acts of saying and another one for acts of doing) as prompts to produce accurate delayed reports of what they promise to do presently, and accurate delayed reports of what they did. Prompts were also used to indicate whether the say–do relations were the ‘same’ in terms of correspondence or not the same in terms of non-correspondence. Prompts were subsequently eliminated. All three subjects demonstrated highly accurate reports of delayed saying, doing and say–do correspondence or non-correspondence discriminations in the absence of prompts and with novel stimuli. The results are discussed in terms of generalized classes of behavior. The implications of these findings for the use of the errorless learning paradigm as a means of establishing complex behavior are also discussed.