Abstract
Organizations increasingly adopt artificial intelligence (AI) to automate routine and data-intensive forms of knowledge work. Yet, even in settings assumed to be analytically tractable and data-rich, AI systems frequently struggle to generate stable and actionable outcomes once deployed in practice. Drawing on a matched-pair, longitudinal qualitative study of AI implementations in diagnostic radiology and supply chain inventory management, this paper examines why such difficulties persist and how organizations respond to them. We conceptualize AI systems as algorithmic assemblages embedded within existing infrastructures of knowing—historically sedimented sociotechnical arrangements that organize how data are produced, validated, and transformed into actionable knowledge. Through comparative analysis, we show that AI integration repeatedly surfaces epistemic misalignments between algorithmic outputs and organizational ways of knowing. Empirically, we identify three recurring forms of data–knowing gaps: completion gaps, where data assumed to be sufficient lack critical contextual information; disambiguation gaps, where algorithmic outputs remain analytically plausible but ambiguous for action; and adaptation gaps, where environmental volatility undermines the continued relevance of historical data. Rather than resolving these misalignments through technical refinement alone, organizations stabilize AI use by selectively reconfiguring their infrastructures of knowing. In both cases, this process culminates in the institutionalization of automation thresholds—organizationally negotiated boundaries that delimit when algorithmic outputs can be acted upon autonomously and when human expertise must prevail. While these thresholds take different forms across contexts, they serve a common function: constraining algorithmic authority to preserve the workability of organizational knowing under uncertainty. By shifting attention from AI’s epistemic limits to infrastructural adaptation, this study advances research on AI and knowledge work and highlights infrastructures of knowing as critical sites through which AI becomes actionable in organizations.