In many languages, what looks like a verb meaning 'say' can also be used alongside another verb in an attitude report. A common analysis of this 'say'-complementation is that it reflects a diachronic link between 'say' verbs and complementizers (e.g., Heine & Kuteva 2002). However, recent work questions this analysis, proposing instead that 'say'-complementation actually involves a verbal form of 'say' (e.g., Major 2021; Driemel & Kouneli 2022; a.o.). For instance, Driemel & Kouneli (2022) argue that the morpheme le in Kipsigis—which means 'say' in matrix contexts and appears in complementation structures—is always a verb, with verbal syntax and semantics across all of its uses. In this paper, however, I show that there are systematic differences across instances of complementation with le, which make the uniformly verbal analysis in Driemel & Kouneli (2022) untenable. Specifically, le only shows verbal syntactic behaviors when it transparently means 'say' and lacks these verbal behaviors in the absence of speech semantics. As a result, I propose that le represents two distinct forms: in some cases, it is the verb 'say,' but in others, it is a complementizer devoid of speech semantics and only linked to the verb diachronically.