Clinical artificial intelligence often performs well in demonstrations but fails to become a sustained part of routine care. The challenge is not only technical or regulatory; it also lies in whether healthcare organisations can build the capacity to authorise, operate and maintain AI systems. A recent analysis published in the Journal of Medical Internet Research uses an 18-month implementation of a provincial clinical AI platform in China to outline a six-module governance framework focused on moving clinical AI beyond temporary pilots.
From Pilots to Institutional Capacity
Clinical AI tools can remain confined to short-lived pilots when ownership is fragmented, accountability boundaries are unclear and deployment depends on temporary resources or individual champions. In this context, a technically functional tool may still fail to become part of everyday clinical practice because no durable organisational arrangement supports continued operation. The framework treats governance capacity as something that develops during implementation, as real-world operational tensions expose gaps in responsibility, data coordination and oversight. Governance is therefore not simply a fixed structure to install before deployment; it becomes stronger as institutions respond to practical demands created by clinical use. The sequence links infrastructure to governance and governance to institutionalisation, making sustained use the observable outcome rather than the starting assumption.
The six modules create a governance cycle rather than a single institutional template. A designated institutional carrier gives AI deployment a formal organisational home with decision authority beyond project teams. Infrastructure governance treats data pipelines, computing capacity and interoperability as durable institutional assets with access, security and update controls. Regulatory and ethical governance defines authorised use and accountability before clinical activation. Interdisciplinary coordination links technical development, clinical decision-making and administrative oversight. Translational scaling connects local success to institution-wide capability. Lifecycle oversight keeps updates, operational feedback and changing clinical conditions under accountable review.
Must Read: Governing AI in Health Systems
Governance Emerges Through Clinical Deployment
The implementation took place at a large tertiary academic hospital in Hebei Province, China, where a clinical AI platform was deployed across three pathways. An intelligent preconsultation pathway used a hospital-developed large language model-enabled conversational interface to collect structured symptom histories before clinical encounters. Summarised risk assessments were transmitted automatically to the electronic medical record, providing preliminary information for outpatient triage and physician assessment. An oncology multidisciplinary decision-support pathway embedded AI-assisted treatment recommendations in tumour board workflows, with outputs reviewed alongside imaging findings, pathology reports and patient-specific clinical data. A therapeutic drug monitoring pathway used an in-house machine learning pipeline to estimate duloxetine plasma drug concentrations and support individualised dosing decisions.
Operational indicators showed sustained activity beyond a bounded pilot. The preconsultation pathway recorded more than 24,000 completed patient interactions across five outpatient departments, with recent average daily use of about 110 encounters during the mature post-rollout phase. The pathway remained operational for more than 12 consecutive months, while three AI-enabled clinical pathways stayed in active institutional use during the observation period. AI-assisted documentation tools, including generative medical record functions and natural-language clinical assistants, also supported routine information management. The platform enabled structured capture of research indicators and integration of clinical data management with institutional quality-monitoring processes.
Transferable Functions, Not Fixed Structures
The framework separates governance functions from the organisational forms used to deliver them. Institutional carrier formation may appear as a dedicated AI laboratory, a formally chartered committee or a federated governance arrangement. Infrastructure governance may rely on hospital-owned platforms, shared regional systems or cloud-based infrastructure under institutional data governance agreements. The consistent requirement is continuity of governed data and computational resources, not ownership of large-scale infrastructure. This distinction is important for organisations with different resource levels, as governance capacity can develop through resource pooling, staged implementation or external technical partnerships. The approach also acknowledges economic feasibility by allowing staged deployment focused on high-value pathways rather than comprehensive infrastructure replication.
The framework also supports governance fit rather than uniform procedural expansion. Oversight intensity should align with clinical risk, update frequency and operational complexity. Routine model refinements may follow simplified authorisation pathways, while higher-impact changes require stricter review. Participatory governance remains an area for maturation. Patient and community involvement becomes operationally meaningful when participatory inputs have a clear institutional pathway into lifecycle governance, feedback channels, communication of system use boundaries and review of operational incidents. International standards and regulatory guidance define what accountable AI governance should achieve, while this framework focuses on how organisational ownership, coordination authority and oversight mechanisms become viable during deployment.
Sustainable clinical AI implementation depends on more than moving a capable technology into clinical practice. The framework places institutional capacity at the centre of deployment, with governance emerging through operational tensions, role clarification and repeated coordination across clinical, technical and administrative domains. The platform shows how AI systems can become routine operational capabilities when ownership, infrastructure, authorisation, coordination, scaling and lifecycle oversight develop together. Its emphasis remains on implementation depth, not model performance alone. Without those conditions, technically mature systems risk remaining promising demonstrations rather than sustained components of clinical care.
Source: Journal of Medical Internet Research
Image Credit: iStock
References:
Tian J, Zhao Z, Tang L et al. (2026) From Pilot Trap to Institutional Capacity: A Governance Framework for Sustainable Clinical AI Implementation in Health Systems.
J Med Internet Res;28:e92680.
