Foundation and multimodal models now match expert-level performance on many diagnostic, prognostic, and biomarker tasks in pathology. Yet only a handful of AI systems are used in routine clinical practice, and the products that have reached patients differ substantially from research prototypes. We define this mismatch as the adoption paradox of computational pathology.
We first survey the technical landscape from task-specific deep learning to large unimodal foundation models, multimodal systems, and early agentic architectures. We then examine what has actually entered the clinic, identifying four product archetypes (digital pathology platforms, population scale cytology screening, assistive detection in surgical pathology, and quantitative immunohistochemistry scoring). Using a three stage maturity model algorithmic capability (Stage 1), system integration (Stage 2), and institutional adoption (Stage 3), we analyze the structural barriers that gate each transition.
Three interconnected barriers explain most of the gap: (1) Data and infrastructure fragility [(pre-analytical variability, scanner-induced domain shift, format fragmentation, annotation scarcity, manual quality control (QC)]; (2) Workflow misalignment (cognitive rhythm of pathologists, automation bias, scenario-dependent latency); (3) Institutional trust deficits (shallow interpretability, incomplete prospective validation, unclear reimbursement, unsettled liability, and regulatory gaps for generative/adaptive systems).
We outline system-level pathways for each stage, including infrastructure first, AI, workflow-embedded intelligence, and adaptive governance. Our central claim is that the next phase of progress will depend less on architectural novelty than on the slower institutional work that turns capability into clinical benefit. The framework provides an actionable lens for regulators, developers, and healthcare organizations to diagnose why a given AI system remains a prototype and what is needed to move it into routine use.
