Ethan Mollick's Co-Intelligence argues that large language models are best understood not as conventional software but as unpredictable collaborators, and that working with them productively requires four practical principles together with a Centaur and Cyborg division of labour. This review evaluates the book through an organizational-practice and action-research lens, attentive to the tension between its evidentiary rigour and its central methodological choice. The principal strength is its grounding in original field experiments, notably the Jagged Frontier study and the companion recruiter study, rather than speculation alone. The principal unresolved issue is that the book instructs readers to anthropomorphize AI for pragmatic benefit while itself documenting the ethical harms of doing so, a tension it flags but does not resolve into operational guidance. The book offers a portable practitioner framework for organizations, action researchers, and educators grappling with AI adoption, though its evidence is scoped mainly to elite United States business and education contexts.