Underneath the surface of managing AI

AI has rapidly emerged as a transformative sociotechnical phenomenon that challenges long-standing assumptions about how organizations manage technology. Whereas earlier technologies could be understood as tools to be implemented and controlled, contemporary AI is characterized by autonomy, learning capacity, and inscrutability, complicating traditional managerial models. Building on sociotechnical perspectives, scholars increasingly emphasize the mutual constitution of social and technical elements in organizational life. Against this backdrop, this dissertation examines how organizations manage AI through multidisciplinary practices and with what consequences. Rather than defining management as a formal organizational function, the dissertation adopts a practice-based perspective in which management is understood as a situated activity distributed across managers, users, and developers. To address this question, the dissertation draws on less established methodological approaches that are particularly suited to coping with the conceptual ambiguity, organizational complexity, and blurring of boundaries associated with AI. The dissertation consists of three self-contained studies. Chapter 2 provides a meta-narrative literature review of augmentation, a concept frequently presented as a desirable objective of AI in organizations. Based on a multidisciplinary review of 76 papers across five academic fields, the chapter identifies four distinct understandings of augmentation: augmenting the body, cognition, work, and performance. These perspectives differ in what is augmented, how augmentation occurs, and which human-AI relationships are involved. The chapter highlights that AI-based augmentation is far from a unified concept and that an overly superficial use of this term obscures important differences, tensions, and interdependencies between augmentation targets. It calls for more precise theorization of augmentation, and argues that managing AI requires clarity about augmentation ambitions and attention to how augmentation narratives shape organizational practices. Chapter 3 draws on a five-year embedded field study within a public employment service (PES), where the author worked as a manager and conducted analytical autoethnography. The chapter examines how AI-generated fuzziness—stemming from AI’s autonomy, learning capacity, and inscrutability—emerges across development, use, and management practices. Fuzzy reasoning in AI creates new affordances, while also generating constraints experienced as loss of control, of predictability, and of expertise. Organizational actors respond through coping practices that constrain, accept, or leverage this fuzziness. The chapter argues that fuzziness is not a random side effect but an integral sociotechnical property of AI, requiring management practices that challenge the common narrative of AI simplifying work. Chapter 4 focuses on a one-year AI development project within the same PES. This chapter unpacks the role of AI prediction errors in shaping organizational practices, by producing performative effects through the actions, signals, and interactions triggered by errors. As the multidisciplinary project team began anticipating these effects already during the AI development, the PES shifted focus from striving for accuracy to “organizing for errors,” proactively adapting algorithm design and organizational practices to mitigate the future consequences of errors. The chapter highlights what happens when an organization moves from managing AI as a self-standing tool to managing the sociomaterial apparatus as a whole. Together, the three studies show that managing AI is a practice-based, relational, and multidisciplinary activity extending beyond one-off strategies or roadmaps. The dissertation contributes to information systems literature by conceptualizing managing AI as a distinct practice bridging development and use. Second, it advances a relational perspective, demonstrating that managing AI involves coordinating interdependencies across organizational elements. Third, it contributes to theorizing AI management as a multidisciplinary accomplishment, arguing that it unfolds through the interaction of different domains of expertise. Through its autoethnographic lens, the dissertation moves beneath the surface of managerial rhetoric and reveals the situated, negotiated, and ongoing work required to “make AI work” in practice.