A multilevel framework of human–AI knowledge work and its epistemic consequences for organizational decision-making

Organizations are adopting generative AI into knowledge work faster than they are adapting the epistemic controls on which knowledge work depends. Existing scholarship has examined the resulting risks predominantly one level at a time (individual reliance and calibration, workflow and oversight design, or institutional legitimacy), and although several traditions acknowledge interactions across actors, routines and institutions, those interactions are rarely theorized as a single cross-level process. Guided by three research questions (how epistemic failures propagate across levels, when such dynamics compound or reverse, and how they could be measured), this paper argues that this insufficient integration becomes consequential under generative AI and develops a framework that extends, rather than replaces, existing multilevel perspectives. We introduce the EPIC (Epistemic Process In Co-production) framework, a multilevel model connecting calibration quality and judgment preservation at the micro level, workflow configuration and oversight architecture at the meso level, and epistemic authority and knowledge legitimacy at the macro level. The framework was developed through a theory-building review of 295 papers retrieved from Scopus and Web of Science (Stage 1, 2018–2026; Stage 2 targeted stream searches, 2004–2026), with per-dimension inter-rater reliability ranging from κ = 0.70 to κ = 0.88. Six constructs and seven testable propositions are advanced, including a provisional four-stage process model of epistemic lock-in in which cognitive legitimacy formation progressively removes the detection capacity that correction requires. Empirical analogues from aviation automation, clinical decision support, satellite navigation, algorithmic hiring and accounting automation establish mechanism plausibility, and we discuss how workforce age and career stage condition these dynamics. We set out four alternative field-level trajectories with leading indicators, map the macro-level constructs onto operative governance instruments, and provide an operationalization roadmap, a worked illustration and a draft diagnostic instrument to support empirical testing.