The paper describes ongoing empirical research into a fundamental problem of linguistics, viz. the architecture of grammar, or the division of labor between lexicon and grammar. We try to find an answer to this question by investigating which part of the utterances in a recent corpus of spontaneous spoken Dutch consists of "Extended Lexical Units" (ELU's), hypothesized to be stored in the lexicon, rather than new syntactic constructs creatively generated from lexical atoms. We describe some problems involved in the identification of these ELUs and in the implementation of them in an NLP system. For the latter, we assess the usability of basic assumptions from frameworks such as Construction Grammar and Head-driven Phrase Structure Grammar.