While recent years have seen a substantial increase of studies investigating vocal iconicity in the lexicon of spoken languages, its presence in grammatical structures is poorly understood. This study investigates the presence of vocal iconicity in nominal classification systems by collecting nominal classification devices from the two main system types: 210 non-agreeing languages (126 families) and 151 agreeing languages (123 families). To detect overrepresentations of sound types in class meanings, the nominal classification devices were grouped according to comparable semantic categories, transcribed using comparable phonetic system, and analyzed through Bayesian mixed models. The strongest results were found for associations between nominal classification devices denoting flat and low, front, unrounded vowels, along with several weak associations relating to shape/size/quantity, function, humanness/animacy, and sex. These associations mostly correlate with previous vocal iconicity findings, but crucially, the involved nominal classification devices are mostly semantically typical for non-agreeing, for example, classifier, systems. These findings were attributed to structural differences between nominal classification system types, which result from grammaticalization processes, for example, phonetic erosion and semantic bleaching. Thus, increased formal predictability through grammatical agreement comes at a cost of semantic transparency which, in turn, dismantles the semantic prerequisites needed for vocal iconic associations to be operational.