Background <p>Emergency department triage is commonly conceptualised as a standardised classification of patient urgency based on vital signs and presenting symptoms. However, research shows that triage is deeply shaped by the clinical context such as organizational structure, clinical uncertainty and subjective decision-making. This contextual complexity presents a challenge for the implementation of artificial intelligence decision-support in emergency care. While Artificial Intelligence-supported triage has demonstrated promising accuracy in controlled settings, these evaluations capture only a limited part of triage work and rarely account for the underlying mechanisms through which triage is sustained in everyday clinical practice. To address this gap, this study aimed to examine how triage unfolds in real-world emergency care and to understand the underlying generative mechanisms shaping triage practice.</p> Methods <p>An ethnographic study was conducted across three Danish emergency departments with different organisational triage configurations. Data consisted of participant observations of triage practice and semi-structured interviews with nurses and physicians. Data were analysed using inductive qualitative content analysis. Findings were interpreted, informed by a critical realistic lens, to understand underlying generative mechanisms shaping triage practice.</p> Results <p>Triage was not a standardized classification event but a dynamic, negotiated practice across the acute admission pathway. Two core generative mechanisms were identified: 1) Coherence work sustained continuity, safety, and flow through transitional coordination, gap compensation, situated judgement, and pragmatic prioritisation. 2) Interpretive reasoning enabled healthcare professionals to navigate clinical complexity through situated interpretive filtering and experience-driven judgement. Both mechanisms were activated by managing competing demands between safety, continuity, and flow within fragmented, resource-constrained organisational setups. Both mechanisms drew on tacit and embodied knowledge not routinely documented or available to artificial intelligence systems.</p> Conclusions <p>These findings demonstrate that safe and effective triage depends not only on accurate classification but on largely invisible adaptive work embedded in everyday practice under conditions of clinical uncertainty and organisational fragmentation. For artificial intelligence-supported triage to achieve clinical impact, system design must recognise and support the coherence- and interpretive reasoning work through which continuity, safety, and flow are maintained in real-world emergency care.</p>