We present an experiment where natural language processing tools are used to automatically identify potential constructions in a corpus. The experiment was conducted as part of the ongoing efforts to develop a Swedish constructicon. Using an automatic method to suggest constructions has advantages not only for efficiency but also methodologically: it forces the analyst to look more objec-tively at the constructions actually occurring in corpora, as opposed to focusing on “interesting” constructions only. As a heuristic for identifying potential constructions, the method has proved successful, yielding about 200 (out of 1,200) highly relevant construction candidates.