Within “word-based”, “paradigm-based” or “abstractive” models of inflectional systems ( Blevins 2006 , 2016 ), only full inflected wordforms are considered primitives; subword strings are treated not as distinct entities, but as abstract generalisations inferred by speakers across multiple inflected forms. These models stand in contrast to “constructive” approaches, which proceed from individual, distinct subword units to full words. An argument consistently adduced in favour of abstractive approaches is that they afford a descriptive advantage regarding “fusional” systems characterised by pervasive non-canonical exponence, such as cumulative exponence, extended exponence, and morphomic structure ( Stump 2016 : 17–30). Via an exploration of inflectional phenomena including non-canonical exponence and arbitrary distributional regularities in the verb inflection of standard Swahili, a language usually described as exemplifying “agglutinating” inflection and amenable to constructive, morpheme-based, analyses, this paper will argue that abstractive systems are equally applicable to “agglutinating” inflection, offering greater empirical plausibility and in some cases descriptive advantage.