Governing AI Across Clinical Care Pathways in Viet Nam: Managing Multiple AI Touchpoints, Handoffs, Conflicting Outputs and Pathway-Level Patient-Safety Risk
IMPLEMENTATION INSIGHTS
Governing AI Across Clinical Care Pathways in Viet Nam: Managing Multiple AI Touchpoints, Handoffs, Conflicting Outputs and Pathway-Level Patient-Safety Risk
IMPLEMENTATION INSIGHT
Last reviewed: 7 October 2026
A healthcare organization can govern several AI systems well and still create an unsafe AI-enabled care pathway.
Consider a patient arriving at an emergency department. AI may influence triage and prioritization. Another model may estimate deterioration risk. Imaging may be prioritized by one system and interpreted with support from another. Clinical decision support may influence the next step. An ambient AI tool may generate documentation. Later, another application may summarize the record for discharge, while an automated system helps prepare patient communication.
Each AI system may have passed its own procurement review, local validation, cybersecurity assessment and internal authorization. Yet the patient does not experience those systems separately. The patient experiences one care pathway.
Once several AI systems influence different stages of that pathway, a different governance question emerges: Are these systems safe not only individually, but also when their outputs, users, workflows, dependencies and failures interact across the patient's journey?
That is the focus of pathway-level AI governance.
Pathway governance is different from portfolio governance
Healthcare AI portfolio governance looks across the organization. It asks which systems exist, where governance attention should be prioritized, which controls should be standardized, where important dependencies are concentrated and whether the overall AI portfolio is producing sufficient value.
Clinical care pathway governance looks at AI from another direction: it follows the patient. The relevant question is not only “Which AI systems do we have?” but also “Which AI systems can influence this episode or sequence of care, how does information move between them, and what happens when their effects interact?”
An organization might operate dozens of AI systems while only a small number affect a particular stroke, cancer or emergency-care pathway. Conversely, one enterprise platform may contain several AI functions that influence the same patient at different stages.
Pathway governance therefore does not replace portfolio governance or system-specific governance. It adds a care-delivery lens.
The terms AI-enabled care pathway, pathway-level AI governance and pathway-level patient-safety risk are used in this article as practical organizational concepts. They are not defined statutory categories under Vietnamese law and are not formal Joint Commission, CHAI or WHO frameworks.
Why this matters in the current Vietnamese context
Viet Nam's Law No. 134/2025/QH15 on Artificial Intelligence has been effective since 1 March 2026. Article 4 establishes human-centred principles, including that AI should serve people rather than replace human authority and responsibility. It also requires human control and capacity to intervene to be maintained and emphasizes safety, accountability, data security and the ability to inspect and supervise AI development and operation [1].
Article 6 is particularly relevant to healthcare. It provides that AI used in essential sectors directly affecting life, health, lawful rights and interests, or public order and safety should be subject to stricter risk management appropriate to the characteristics of the sector. For healthcare specifically, the Law refers to patient safety, reliability under real-world conditions and protection of health data in accordance with law [1].
Decree No. 142/2026/NĐ-CP, effective since 1 May 2026, further details the AI Law, including rules concerning risk classification, notification, transparency, incident management, conformity assessment and management of high-risk AI systems [2]. Decision No. 33/2026/QĐ-TTg, effective since 15 August 2026, establishes Viet Nam's current List of High-Risk Artificial Intelligence Systems [3].
Decision No. 33 also reinforces an important governance principle: using AI does not alter, transfer or exclude the authority and responsibility of competent organizations and individuals, and human supervision, control and capacity to intervene must be maintained during operation [3].
For healthcare, the current list under Decision No. 33 is relatively specific. Its healthcare entries focus on defined AI-assisted surgical and robotic-surgery systems that meet specified conditions [3].
This distinction matters because many other AI systems can still materially influence clinical care without necessarily appearing in those healthcare entries. A triage-support tool, documentation assistant, imaging application, predictive model or patient-communication system may affect information flow, clinician attention, workflow or downstream decisions. The fact that an AI use is not captured by a particular entry in the high-risk list does not remove the broader need to manage patient-safety risk under healthcare law, the AI Law and other applicable requirements.
Pathway-level patient-safety risk therefore cannot be inferred from statutory AI classification alone.
A further distinction is also important. Where an AI-enabled product falls within Viet Nam's medical-device regulatory framework, medical-device classification and management remain a separate legal analysis. Circular No. 24/2026/TT-BYT, effective from 1 July 2026, addresses risk determination and management for medical-device products and amends parts of the existing medical-device regulatory framework [4]. AI risk classification under the AI Law and medical-device risk classification should therefore not be treated as the same question.
Existing healthcare law already supports a systems approach to safety
The Law on Medical Examination and Treatment provides an important healthcare foundation.
Article 71 states that prevention of medical incidents should be based on identifying and reporting incidents, analysing causes, developing recommendations and preventive measures, avoiding recurrence and improving healthcare quality and patient safety. Responsibility for medical-incident prevention rests with the head of the healthcare establishment and people working within it [5].
The current consolidated text is Consolidated Document No. 26/VBHN-VPQH dated 26 February 2026.
Circular No. 19/2013/TT-BYT also remains recorded as in force. It establishes hospital quality management as a continuous organizational responsibility and provides that quality-management decisions should be grounded in law, scientific evidence and the practical needs of quality improvement [6].
Neither instrument establishes a specific framework for AI-enabled clinical pathways. They do, however, support a broader principle that becomes increasingly important as AI enters care: patient safety needs to be managed as part of the healthcare delivery system, not only as a property of individual technologies.
International frameworks are also moving beyond isolated AI tools
Joint Commission's Responsible Use of AI in Healthcare (RUAIH) Certification, launched on 1 June 2026, is a voluntary organizational certification for healthcare organizations and health systems in the United States. It focuses on whether an organization has the governance, safeguards, monitoring processes and education needed to use AI responsibly; it does not certify individual AI products or tools. Its standards are organized around five areas: governance; effective data management; risk and bias reduction; monitoring, evaluation and validation of safety performance, effectiveness and responsible use; and transparency, education and training [7].
CHAI's 2026 Governance Playbooks similarly approach responsible AI as an organizational capability. Their baseline controls are intended to be adapted to the context of individual healthcare organizations and cover eight areas: AI policy; organizational structures; organizational resources; responsible AI lifecycle management and use; risk and impact assessments; responsible data management and use; third-party management; and education, training and feedback [8].
WHO/Europe's June 2026 report Bridging Theory and Practice – Implementation Insights on Artificial Intelligence in Health Care focuses specifically on translating responsible-AI principles into real healthcare implementation [9]. A separate WHO/Europe report published in September 2026 summarizes discussions within its Knowledge Community on responsible AI in health. Participants identified fragmented and biased data, governance gaps, unclear accountability and AI-literacy limitations as important barriers, while highlighting strong governance, transparency, equitable access and codesign with patients and clinicians as important enablers [10].
These international sources support a broader health-system perspective on responsible AI. They do not prescribe the pathway-governance framework proposed in this article.
Individually well-governed AI does not guarantee a safe pathway
A common assumption is that if every AI system has been appropriately reviewed, the resulting workflow should also be safe. That assumption can fail.
Imagine an emergency-care pathway using four AI applications. The first supports acuity assessment. The second predicts deterioration. A third prioritizes imaging. A fourth provides diagnostic support. Each application may perform within its own validated specifications.
The combined pathway can still create new risks. A lower triage priority may delay information needed by a deterioration model. A reassuring imaging output may cause another warning to receive less attention. Two systems may describe a patient as “high risk” or “low risk” while referring to completely different outcomes or time horizons.
The safety problem is no longer contained within one algorithm. It arises from the interaction between systems, people, information and workflow.
That is pathway-level risk.
Start by mapping where AI actually touches the patient journey
The first practical step is to map the clinical pathway from the patient's perspective. The organization should identify material AI touchpoints from entry into care through assessment, diagnosis, treatment, monitoring, transition and follow-up.
Some will be obvious: radiology AI, deterioration prediction, clinical decision support or robotic surgery. Others may be less visible. A scheduling algorithm can affect when a patient is seen. An ambient documentation tool can shape what enters the medical record. A summarization system can influence what the next clinician receives. An administrative application can affect downstream utilization processes. A patient-facing assistant can shape the information the patient receives after discharge.
Pathway mapping should therefore look beyond products marketed specifically as “clinical AI.” The practical question is whether an AI function can materially affect information, timing, prioritization, action or communication within the care pathway.
The organization should also understand the role of each AI. One system may simply present information, while another summarizes, predicts, prioritizes, recommends, triggers a workflow or initiates an action. A useful operational sequence is:
informs → summarizes → predicts → prioritizes → recommends → triggers → acts
These are not legal classifications. They are practical descriptions that help teams understand where and how AI influences care. A system that summarizes previous notes creates a different pathway risk from one that automatically changes patient priority or initiates a downstream process.
Handoffs, provenance and error propagation deserve special attention
Healthcare has always been vulnerable at handoffs. AI introduces additional forms.
There may be an AI-to-human handoff, where a model produces information that a clinician must interpret. There may be a human-to-AI handoff, where information entered by staff becomes input to another system. Increasingly, there can also be an AI-to-AI handoff, even when nobody intentionally designed one.
For example, an ambient AI scribe may generate a clinical note. That note enters the electronic record. A later risk-prediction or summarization system then consumes the note as part of its input. The downstream AI may have no awareness that some of its source material originated from another AI system.
An upstream error can therefore be carried forward as if it were established information. This is different from independently using two separate AI systems.
Once AI-generated information enters a trusted medical-record or workflow environment, its origin may also become less visible. A generated summary may later be treated like directly observed clinical information. An inferred diagnosis may be copied forward. A risk classification may influence subsequent prioritization without the next clinician understanding how that information was produced.
Pathway governance should therefore identify which downstream decisions depend materially on upstream AI-generated or AI-influenced information. This does not mean every data element needs an elaborate technical provenance label. But where an AI-mediated handoff can materially affect subsequent care, enough traceability should exist to understand where important information originated and whether appropriate human review occurred.
Single-system validation usually asks whether one AI output is sufficiently reliable for its intended use. Pathway review needs to ask another question: What happens downstream when that output is wrong?
An incorrect triage classification may change waiting time. An AI-generated documentation error may travel to another department. A false-negative interpretation may influence another system or clinician that assumes the information is reliable. An inaccurate discharge summary may eventually affect what the patient is told.
The relevant issue is therefore not only the probability that AI A produces an error, but what that error can do to B, C and D. Some failures remain local and are easy to detect. Others propagate because downstream processes increasingly trust the upstream information.
Conflicting outputs need an agreed clinical response
As healthcare organizations deploy more AI, clinicians are increasingly likely to encounter systems producing apparently conflicting signals.
A deterioration model may classify a patient as high risk while another application shows no immediate concern. An imaging system may raise suspicion while a generative summary sounds reassuring. Two predictive systems may use different outcomes, time horizons or thresholds and therefore appear to disagree even when both are technically functioning as intended.
The organization should not wait until clinicians encounter these situations in real care before deciding how they should be handled. Pathway governance should define a response to materially conflicting AI outputs.
This does not necessarily mean designating one AI as the authoritative system. The more useful question is: Which clinical evidence, professional role or workflow governs the decision in this context, what additional review is needed, and when should the discrepancy be escalated?
AI disagreement should not become an unstructured problem left to individual clinicians to resolve without organizational guidance.
Pathway design should also recognize that apparently conflicting outputs may not actually be comparable. One system may predict deterioration within six hours, while another predicts thirty-day mortality. One may rely on imaging, while another uses laboratory and physiological data.
If both are displayed together, the clinical context needs to make those differences understandable. Otherwise, individually valid outputs can become misleading when presented as if they answer the same question.
Meaningful human oversight has to work across the whole pathway
Many healthcare AI governance programs correctly require human review. But having a human reviewer for each individual system does not automatically create meaningful human oversight across the pathway.
A clinician may be expected to verify AI-supported triage information, review an imaging alert, confirm an AI-generated note, assess a treatment recommendation and check a generated discharge summary during the same episode of care. Each system can truthfully claim that a human remains involved. The combined verification burden may nevertheless become unrealistic.
When human review accumulates across several technologies, people can begin confirming outputs routinely rather than actively challenging them. Pathway governance should therefore assess cumulative human-review burden, not only whether each individual tool contains a nominal human-review step.
Meaningful human oversight also requires more than placing a clinician somewhere in the workflow. The person needs enough information, competence, time and authority to recognize a problem and intervene appropriately. The practical question is not simply “Is there a human in the loop?” It is whether that person can realistically understand the relevant context, challenge the AI, override or interrupt an AI-supported action when appropriate, escalate uncertainty and continue care safely.
This is closely aligned with the human-centred principles of Viet Nam's AI Law.
Automation itself can also accumulate gradually. One system automates a small step, then another automates prioritization, another documentation, another summarization and another patient communication. No single implementation may appear to transfer a large amount of control away from clinicians, yet the combined pathway can become increasingly dependent on automated outputs.
Organizations should therefore periodically examine how much of a clinically important pathway has become AI-mediated, where professional judgment remains essential and whether staff can still operate safely when AI is wrong, degraded or unavailable.
Alert burden, shared dependencies and fallback should be reviewed together
An AI vendor may demonstrate that its product generates an acceptable number of alerts. Another vendor may demonstrate the same. The combined experience for the nurse or clinician can still become unmanageable.
A pathway-level review should therefore consider the cumulative number of alerts, warnings, flags, prompts and required confirmations reaching the same professionals during the same workflow. Alert fatigue can become more than a product-usability problem. It can become a pathway architecture problem.
Shared technical dependencies create another pathway-level issue. Several AI applications may carry different product names while relying on the same cloud service, foundation model, API, identity service, data pipeline, network connection or enterprise platform.
If the common dependency fails, several AI-supported functions may become unavailable simultaneously. The relevant pathway question is not simply whether each product has a downtime plan. It is: What happens to patient care if several AI-supported steps fail at the same time?
This also changes how fallback should be designed. A radiology system may revert to manual reading. An ambient scribe may revert to manual documentation. A triage system may revert to conventional nurse assessment. Individually, each contingency may be reasonable. But if all three are triggered simultaneously, the organization may not have enough staff or time to operate every manual fallback at once.
Pathway governance therefore needs coordinated fallback planning. Teams should understand which AI-supported functions can fail together, what manual workload will return, which activities need priority and how staff will know that the pathway is operating in a degraded mode.
Cross-department pathways need cross-functional visibility and accountability
Clinical pathways rarely stay within one department. A patient with cancer may move through outpatient consultation, imaging, pathology, multidisciplinary review, treatment planning, pharmacy, inpatient care and follow-up. Different departments may have very different levels of AI maturity.
One department may rigorously validate its AI system. Another may be using an embedded AI feature with relatively little organizational awareness. A third may rely on a vendor-hosted application. The patient journey crosses all of them.
Pathway governance should therefore provide enough cross-department visibility that one part of the organization is not unknowingly relying on assumptions made elsewhere.
This does not mean every clinical pathway needs a new AI committee. It means important cross-functional AI dependencies need somewhere to be seen and governed.
Every material AI use should already have an accountable system or use-case owner. Pathway governance adds another question: Who has enough visibility to understand the safety of the overall AI-enabled care pathway?
The answer will vary. It may be an existing clinical pathway lead, service-line leadership, quality and patient-safety function, multidisciplinary governance structure or another appropriately authorized role.
This article does not propose a new statutory position called an “AI pathway owner.” The objective is simply to avoid a situation where radiology owns one system, emergency medicine another, IT owns the integration and a vendor manages another component, while nobody can see how the combination behaves from the patient's perspective.
System ownership and pathway accountability are related, but they are not the same thing.
Incident investigation and near-miss learning should follow the pathway of causation
When an AI-related incident occurs, organizations may naturally focus on the system closest to the final harm. That can miss the actual causal chain.
Suppose an inappropriate discharge recommendation is generated. The immediate investigation might focus on the discharge AI. But that system may simply have summarized an earlier documentation error originating from an AI-generated note, which was not corrected because another alert had already diverted clinician attention.
The pathway crosses several systems, people and handoffs. Incident investigation should therefore examine how information and decisions moved through the pathway, rather than only whether the final AI system operated according to specification.
This systems-oriented approach is consistent with Article 71 of Viet Nam's Law on Medical Examination and Treatment, which bases medical-incident prevention on identifying incidents, analysing causes, learning and implementing measures to avoid recurrence and improve patient safety [5].
Near misses can be equally informative. A clinician who catches a conflicting AI recommendation before harm occurs has identified a useful safety signal. An AI-generated note corrected before another department relies on it may reveal a handoff vulnerability. An alert ignored because several systems were demanding attention simultaneously may reveal cumulative workload risk.
Pathway governance should therefore make use of near misses, frontline concerns and weak safety signals, not only serious adverse events. The aim is to identify interaction problems before they become repeated harm.
Pathway monitoring should measure care, not only models
A collection of well-performing AI models does not necessarily produce better healthcare. Pathway monitoring should therefore include outcomes related to the care process and the patient, not only technical model metrics.
Depending on the pathway, useful measures may include time to diagnosis, time to treatment, delayed escalation, unexpected transfers, duplicate testing, length of stay, readmission, documentation corrections, clinician override patterns, alert burden, workflow delays, staff experience, patient experience and relevant clinical outcomes.
Equity should remain visible as well. A pathway can appear effective overall while creating longer delays, higher override rates or poorer outcomes for a particular patient group.
The objective is not to produce hundreds of indicators. It is to select enough measures to answer a practical question: Is the AI-enabled pathway becoming safer, more effective, more equitable or more efficient in the real healthcare environment?
This distinction matters because model success and pathway success are not the same thing. A diagnostic AI may improve sensitivity while increasing downstream testing enough to create new bottlenecks. A triage model may improve prioritization while shifting workload to a service that lacks capacity to respond. An ambient documentation tool may improve clinician experience while increasing downstream review work.
Those outcomes do not automatically mean the AI should be rejected. They mean the healthcare organization needs to evaluate effects beyond the immediate output of each technology.
The ultimate governance question is not simply whether the models perform well. It is: Did the AI-enabled pathway improve care?
Pathways need reconsideration when either technology or workflow changes
An AI system does not need to fail for pathway risk to change.
A vendor may change a threshold. A foundation model may be replaced. A user interface may influence clinician behaviour differently. A department may begin using an existing AI output for a new purpose. A downstream system may start consuming information that previously had been used only for display.
System-level change control should determine whether an individual AI requires revalidation or another formal review. Pathway governance adds another question: Does this change alter a downstream dependency, handoff, workload, clinical interpretation or decision?
A change that appears minor within one application may still be important to the pathway.
Technology changes are also not the only reason to reassess. Staffing models change. Clinical guidelines evolve. New services open. Documentation practices change. Care may be reorganized across facilities. The AI itself can remain unchanged while its environment changes enough to alter risk.
Pathway governance should therefore consider sociotechnical change, not only model or software versions.
Multi-site deployment should not assume the same pathway exists everywhere
Hospital groups often refer to one stroke pathway, oncology pathway or emergency pathway across multiple facilities. In practice, workflows can differ substantially.
One hospital may have an on-site radiologist overnight while another relies on remote reporting. One facility may have enough staff to respond immediately to deterioration alerts while another does not. One site may have seamless electronic integration while another requires manual transfer of information.
The same combination of AI systems can therefore create different pathway-level risks at different sites.
Central governance can establish common minimum requirements, but pathway review should remain sensitive to local workflow, workforce and infrastructure. This is also consistent with the broader organizational direction reflected in Joint Commission RUAIH: common governance can improve consistency across an organization or health system, but safe implementation still depends on how AI is actually used in care [7].
Patients and caregivers can reveal pathway problems that organizations miss
Patients experience healthcare across departmental boundaries. They may notice repeated questions, inconsistent explanations, conflicting instructions, unnecessary delays or confusing changes in care that individual teams do not see.
Patient and caregiver feedback can therefore contribute to pathway-level learning. Patients do not need to evaluate algorithms technically. The organization can ask about the care experience: whether explanations were consistent, whether instructions conflicted, whether technology caused confusion, whether transitions felt coordinated and whether patients understood how to ask questions or raise concerns.
Those observations can complement technical, clinical and operational monitoring.
The 2026 WHO/Europe Knowledge Community report similarly records participants identifying codesign with patients and clinicians, transparency and strong governance as important enablers of responsible AI implementation [10]. This should be understood as a synthesis of the Knowledge Community's discussions, not as a formal WHO requirement for the pathway model proposed here.
Transparency should also remain proportionate and meaningful. If many AI systems contribute to one pathway, providing a separate technical explanation for every algorithm may overwhelm rather than inform the patient.
Where AI materially affects care, organizations should consider what information is relevant for patient understanding, what professional oversight remains and how questions or concerns can be raised. The objective should be meaningful transparency, not disclosure volume.
A practical AI-Enabled Care Pathway Governance Review
Healthcare organizations in Viet Nam can structure pathway-level governance around the following cycle:
map the clinical pathway → identify AI touchpoints → define each AI's role → map information and decision handoffs → identify dependencies and shared infrastructure → define human decision authority → anticipate conflicting outputs → test error propagation and cumulative workload → design coordinated escalation and fallback → define pathway-level outcomes → monitor real-world performance → review changes and safety signals → learn and redesign
Map the clinical pathway means following the actual patient journey rather than the organizational chart. Identify AI touchpoints means locating material AI influence across clinical, operational, documentation and communication processes. Define each AI's role means understanding whether a system informs, summarizes, predicts, prioritizes, recommends, triggers or acts.
Map information and decision handoffs means identifying where AI-generated or AI-influenced information moves to another person, department or system. Identify dependencies and shared infrastructure means understanding where multiple parts of the pathway rely on the same data pipeline, vendor, model, cloud service, API or operational resource.
Define human decision authority means making clear which professionals or organizational roles remain responsible for interpreting information, resolving uncertainty and escalating concerns. Anticipate conflicting outputs means determining how important disagreement among AI systems, or between AI and other clinical evidence, should be handled.
Test error propagation and cumulative workload means considering what happens when an upstream system is wrong and whether combined alerts, reviews and verification demands remain operationally realistic. Design coordinated escalation and fallback means ensuring that the pathway can continue safely when one or several AI-supported functions become unreliable or unavailable.
Define pathway-level outcomes means selecting measures that reflect safety, quality, timeliness, equity, workload and real-world benefit rather than relying only on model performance. Monitor real-world performance means looking for emerging patterns after deployment, including near misses, workflow effects and cross-system interactions.
Review changes and safety signals means reconsidering the pathway when technology, workflow, dependencies or evidence materially change. Learn and redesign means using those findings to improve not only individual AI systems but the care pathway itself.
This AI-Enabled Care Pathway Governance Review is proposed here as a practical organizational approach. It is not a statutory Vietnamese framework, a Ministry of Health requirement, a Joint Commission standard, a CHAI framework or a WHO guideline.
Pathway governance should remain proportionate
Pathway governance should not create another large committee for every clinical service, nor should every low-consequence AI feature require a comprehensive pathway review. The intensity of governance should reflect the potential effect on care.
More structured pathway review may be appropriate where several AI systems influence clinically important decisions; where AI-generated information becomes input for another system; where several tools depend on the same professionals or infrastructure; where automation is accumulating; where failure can propagate downstream; or where the pathway serves particularly vulnerable or clinically high-consequence populations.
A simple administrative process may require little additional pathway governance. A complex emergency, oncology, surgical or intensive-care pathway involving multiple interacting AI systems may justify much more.
What healthcare organizations should avoid
One failure mode is system-by-system tunnel vision, where each AI passes individual review but nobody examines their combined effect on care. Another is nominal human oversight, where every product technically includes human review but the combined verification workload makes meaningful intervention unrealistic.
Organizations should also watch for invisible AI-to-AI handoffs, where generated information becomes downstream input without sufficient awareness of its origin; unresolved output conflict, where clinicians receive competing AI signals without a clear review pathway; and uncoordinated fallback, where each technology has a contingency plan but simultaneous fallback processes cannot realistically be operated.
Other warning signs include gradual automation accumulation, where successive implementations change the balance between human and automated decision-making without a deliberate pathway-level review; metric fragmentation, where each AI meets its individual KPI while patient outcomes remain unchanged or deteriorate; and cross-department accountability gaps, where every system has an owner but nobody can see the combined patient journey.
The common problem is the same: the organization governs the components but loses sight of the care system they collectively create.
What this means for healthcare organizations in Viet Nam
As healthcare AI moves from isolated applications into combinations of technologies embedded in routine care, organizations will increasingly face questions that cannot be answered one system at a time.
They will need to understand where AI influences the patient journey, what information moves between systems, whether an upstream error can propagate, how conflicting outputs are handled, how much review burden is accumulating around the same clinicians, what happens when shared dependencies fail and whether the pathway remains safe when implemented at another site.
Most importantly, they need to ask whether patients are actually receiving better and safer care.
The practical transition is:
from governing individual AI systems → to governing their interactions → to governing the AI-enabled care pathway experienced by the patient
An organization may reasonably conclude that every individual AI system in a pathway is acceptable. That does not automatically mean the pathway created by combining them is acceptable.
Healthcare quality and patient safety have long shown that harm can emerge not from one component alone, but from interactions among people, processes, technologies, information and organizational conditions.
AI does not change that principle.
It makes it even more important.
Professional education notice
This article is intended for professional education and knowledge exchange. It does not constitute individualized legal, regulatory, clinical, patient-safety, medical-device, cybersecurity, personal-data-protection, accreditation or quality-management advice.
The terms AI-enabled care pathway, pathway-level AI governance, pathway-level patient-safety risk and AI-Enabled Care Pathway Governance Review are used in this article as practical organizational concepts. They are not defined statutory categories under Vietnamese law.
The AI-Enabled Care Pathway Governance Review proposed here is not a Ministry of Health framework, statutory requirement, Joint Commission standard, CHAI framework or WHO guideline.
Pathway-level governance does not replace system-specific obligations. Applicable requirements concerning AI risk classification, high-risk AI, conformity assessment, personal-data protection, medical-device regulation, medical records, professional responsibility, clinical governance and other obligations must still be assessed for each relevant system and use. Where an AI-enabled product is also regulated as a medical device, AI-law classification and medical-device classification should be assessed separately under their respective regulatory frameworks.
Joint Commission RUAIH, CHAI Governance Playbooks and WHO/Europe materials are used as international reference points for organizational governance, implementation and health-system learning. They do not create legal requirements for healthcare organizations in Viet Nam.
References
National Assembly of Viet Nam. Law No. 134/2025/QH15 on Artificial Intelligence. Adopted 10 December 2025; effective 1 March 2026.
Government of Viet Nam. Decree No. 142/2026/NĐ-CP Detailing a Number of Articles and Measures for Implementation of the Law on Artificial Intelligence. Issued 30 April 2026; effective 1 May 2026.
Prime Minister of Viet Nam. Decision No. 33/2026/QĐ-TTg Promulgating the List of High-Risk Artificial Intelligence Systems. Issued 30 June 2026; effective 15 August 2026.
Ministry of Health of Viet Nam. Circular No. 24/2026/TT-BYT on Determination of Risk Levels and Management Measures for Medical-Device Products and Goods and Amendments to Circular No. 05/2022/TT-BYT. Issued 30 June 2026; effective 1 July 2026.
National Assembly of Viet Nam. Law No. 15/2023/QH15 on Medical Examination and Treatment, as amended. Effective 1 January 2024. Current consolidated text: Office of the National Assembly, Consolidated Document No. 26/VBHN-VPQH dated 26 February 2026.
Ministry of Health of Viet Nam. Circular No. 19/2013/TT-BYT Providing Guidance on Management of Quality of Medical Examination and Treatment Services in Hospitals. Issued 12 July 2013; effective 15 September 2013; currently recorded as in force.
Joint Commission. Responsible Use of AI in Healthcare Certification. Launched 1 June 2026.
Coalition for Health AI. AI Governance Playbooks. Released 27 May 2026.
World Health Organization Regional Office for Europe. Bridging Theory and Practice – Implementation Insights on Artificial Intelligence in Health Care: Report of Working Group 1 of the Strategic Partners' Initiative for Data and Digital Health. Published 26 June 2026.
World Health Organization Regional Office for Europe. Report of the Knowledge Community on Responsible Artificial Intelligence in Health. Published 1 September 2026.