The Adoption Gap as architecture, not Behavior: Why clinical AI adoption fails even when the model is good

Clinical artificial intelligence (AI) adoption-failure is dominantly framed in two ways: as a behavioural problem (clinician resistance, training gap, change-management deficit) addressed by frameworks such as NASSS and CFIR, or as a model-quality problem (accuracy, calibration, fairness) addressed by external validation. Both frames are reparative; they engage after the implementation decision has been made. We argue that the actual failure mode lies upstream, at seven architectural pre-deployment decisions that determine whether any clinical AI system can ever be operationally trusted: workflow-embedding position, decision-rights allocation, accountability routing, error-recovery loop topology, escalation hierarchy, counterfactual visibility, and authority layer. When these decisions are left implicit, even high-accuracy models accumulate workarounds and convert into AI-Shadow-Care, a deployment state in which clinicians acknowledge AI alerts to dismiss them, document around outputs, and route around recommendations. We test this taxonomy by re-coding 12 published clinical-AI deployment cases through the architecture lens, including the 2026 multi-centre prospective validation of the Epic Sepsis Model version 2, in which improved discrimination (AUROC 0.82-0.92) coexisted with persistently low positive predictive value (0.13-0.26) and high alert burden. We propose a pre-deployment seven-question architecture audit and outline implications for procurement, regulation, and reimbursement. Adoption is an architectural choice, not a behavioural outcome.