Not Legally High-Risk Does Not Mean Low Clinical Risk: How Healthcare Organizations in Viet Nam Should Govern AI Beyond Regulatory Classification
IMPLEMENTATION INSIGHTS
Not Legally High-Risk Does Not Mean Low Clinical Risk: How Healthcare Organizations in Viet Nam Should Govern AI Beyond Regulatory Classification
POLICY & REGULATORY INSIGHT
Last reviewed: 2 October 2026
A healthcare organization may be considering an AI system that helps identify abnormalities on medical images, highlights possible lesions during endoscopy, predicts deterioration, prioritizes patients for review or supports another clinical decision. It checks Viet Nam’s current list of high-risk AI systems and does not find that particular type of system or deployment there. It can then be tempting to conclude that if the AI is not legally high-risk, its clinical risk must also be low. That would confuse two different questions.
Viet Nam now has a statutory framework for classifying AI systems by risk under the Law on Artificial Intelligence No. 134/2025/QH15, effective since 1 March 2026, and Decree No. 142/2026/NĐ-CP, effective since 1 May 2026. The Law establishes high-, medium- and low-risk categories, while the Decree provides the operational rules for determining which category applies.[1,2]
For healthcare organizations, however, legal classification should not be the end of the risk conversation. An AI system may fall outside the current statutory high-risk list and still influence diagnosis, triage, treatment, surveillance or another part of care where failure, inappropriate reliance or poor implementation could contribute to patient harm. The practical challenge is therefore to keep regulatory classification and clinical risk assessment connected, but not confused.
Legal risk classification and clinical risk answer different questions
Article 9 of the AI Law classifies AI systems as high, medium or low risk. High-risk systems are systems capable of causing significant harm to life, health, lawful rights and interests, national interests, public interests or national security. Medium-risk systems concern the possibility of misleading, influencing or manipulating users because they do not recognize that they are interacting with AI or AI-generated content. Low-risk systems are those that do not fall within those two categories.[1]
Decree No. 142 makes the classification process more specific. A system is classified as high-risk when it belongs to the high-risk list issued by the Prime Minister. A system outside that list may be classified as medium-risk if it meets the relevant criteria; otherwise, it is classified as low-risk.[2]
These statutory categories matter because different legal obligations follow from them. But clinical risk, as the term is used in this article, asks something different: what could happen to patients if an AI system is wrong, unavailable, misunderstood, poorly integrated or relied upon inappropriately? Clinical risk in this sense is an organizational patient-safety concept. It is not another risk category created by Viet Nam’s AI legislation.
Regulatory classification asks: Which statutory AI risk category applies, and what legal obligations follow? Clinical risk assessment asks: Given the way this system will actually be used in care, what could go wrong, how serious could the consequences be, and what controls are needed? Healthcare organizations need answers to both.
Viet Nam’s current high-risk list is specific in healthcare
Decision No. 33/2026/QĐ-TTg, issued on 30 June 2026 and effective from 15 August 2026, establishes the current list of high-risk AI systems.[3] Within the healthcare section, the list currently contains two entries, both involving AI-supported surgery or surgical robots under specified conditions.
The first covers AI integrated into robots that directly participate in an intervention or guide surgical actions, including fully automated or intelligent robots performing actions in place of humans. The second concerns AI integrated into robotic equipment that directly carries out therapeutic actions on a patient under a medical order without requiring confirmation from healthcare personnel for each change in parameters.[3]
This distinction matters. The current list should be read as the legal list that it is, not as a catalogue of every AI application capable of having clinically important consequences.
Clinical AI already spans a much wider range of functions, including radiology and other imaging support, pathology, endoscopy, ophthalmology, risk prediction, deterioration alerts, triage, medication support, clinical documentation and decision support. Whether a particular system is legally high-risk must be determined from the applicable legal framework, its intended purpose and the way it is deployed. A system should not be labelled legally high-risk simply because its use is clinically important. The reverse is equally important: being outside the current statutory high-risk list does not by itself establish that an AI system is clinically safe or that only minimal organizational oversight is appropriate.
Why regulatory classification and clinical significance may diverge
Viet Nam’s regulatory framework helps explain why these concepts should remain separate. Article 8 of Decree No. 142 sets out the criteria and principles used to develop the high-risk list. These include the potential magnitude of harm to life, health, property, human rights and public interests; the degree of automation; the system’s role in supporting a final decision; the ability of humans to supervise and intervene; the field of use; and the scale of impact.[2]
The same Article provides that certain systems are not proposed for inclusion in the high-risk list even where they otherwise meet relevant criteria. These include systems with technical and operational arrangements that ensure substantive human oversight, allowing an authorized person to independently review, intervene in, reject or change a decision before it takes effect. It also includes systems that only perform analytical, predictive, assessment or recommendation functions for reference, where the output is not used as the sole basis for the final decision.[2]
These provisions show why autonomy, meaningful human oversight and the role of AI in final decision-making matter when the high-risk list is developed. They should not, however, be treated as a separate classification test that an individual healthcare organization can use to override the statutory framework.
From a patient-safety perspective, even an advisory system can still matter. Consider an AI tool that highlights possible lesions during gastrointestinal endoscopy. The endoscopist remains responsible for the examination and may disregard the AI output. That human oversight is important. But the organization should still understand what happens if the system misses a lesion, creates excessive false alerts, performs differently in the local patient population or gradually changes the way clinicians search for abnormalities.
Similarly, an imaging AI may technically provide only a recommendation while the clinician retains final authority. If that recommendation becomes deeply embedded in workflow and clinicians routinely rely on it, its real-world influence may nevertheless be substantial. Legal classification describes a system within a regulatory framework; it does not replace examination of how the technology actually functions within clinical care.
A low-risk legal classification is not a clinical safety determination
The words low risk can easily carry more meaning than the legislation gives them. Where an AI system is legally classified as low-risk, that classification should be recorded accurately. A healthcare organization should not relabel it as legally high-risk simply because the clinical use deserves greater scrutiny.
At the same time, a low-risk classification under AI legislation does not by itself establish that a system is clinically effective, suitable for a particular patient population, appropriate for a particular workflow, adequately validated locally or safe to use without additional organizational controls.
Healthcare organizations already make similar distinctions with other technologies. A product may lawfully be available for use while the healthcare organization still needs local controls around selection, competency, workflow, monitoring and escalation. AI should be approached with the same discipline. The intensity of internal governance should reflect what the system actually does in care and what could happen if it fails.
Start with the clinical function, not the technology label
A useful assessment begins with a precise description of the use case. “AI for radiology” is too broad. “AI used to flag possible intracranial hemorrhage and influence worklist prioritization in the emergency setting” is much more informative. “AI-assisted endoscopy” is also too broad. “Real-time AI displaying visual alerts to support lesion detection during colonoscopy while the endoscopist retains responsibility for examination and diagnosis” describes the clinical function far more clearly.
The organization should understand who sees the output, what action may follow, whether the AI changes priority or treatment, how time-sensitive the decision is, whether another clinician independently reviews the case, and what happens when the system is unavailable or wrong.
The same underlying technology can create very different risks in different deployments. A prediction displayed in a dashboard for optional review is not equivalent to a prediction that automatically triggers an urgent clinical pathway. An image-analysis system used as a second reader by an experienced specialist is not equivalent to the same technology becoming the principal screening mechanism in a setting with limited specialist review. Clinical AI governance therefore needs to assess the deployed use case, not only the model.
The consequence of error should shape governance intensity
One of the most useful questions is also one of the simplest: If this AI is wrong, what happens next?
Not every error has the same consequence. An error in an administrative scheduling tool may cause inconvenience. A false-negative clinical alert could contribute to delayed diagnosis or treatment. An inaccurate prediction may influence escalation of care. A poorly calibrated triage model may repeatedly deprioritize patients who need earlier attention. An AI-generated clinical summary may introduce inaccurate information into the medical record that later clinicians assume to be correct.
Healthcare organizations should therefore look beyond headline accuracy and consider the severity, likelihood and detectability of potential harm. A relatively uncommon failure may still justify strong controls when the possible consequence is serious and the error is difficult to detect before it affects a patient. Conversely, some imperfect systems may be manageable where errors are readily identifiable, clinicians independently review the relevant clinical information, resulting decisions are reversible and reliable fallback processes exist.
The important question is not whether an AI system can ever be wrong. It is whether foreseeable failures have been understood and whether the surrounding clinical process makes it reasonably likely that they will be detected or contained before they become patient harm.
Human oversight must work in practice
Human oversight is an important part of Viet Nam’s AI framework. Decision No. 33 requires human supervision, control and the ability to intervene in the operation of AI systems, while Decree No. 142 recognizes substantive human oversight as relevant to the principles for developing the high-risk list.[2,3]
But placing a clinician nominally “in the loop” does not automatically make an AI-supported process safe. Meaningful oversight depends on whether that clinician has enough information, time, authority and competence to question the output. It also depends on whether the workflow preserves independent judgment rather than quietly encouraging automatic acceptance.
Healthcare organizations should therefore ask whether users can recognize situations in which the AI may be unreliable, access the relevant clinical information independently, reject or override the output, understand important limitations and escalate suspected problems.
Real-world behavior matters. A tool formally described as advisory may become highly influential if users almost always accept its recommendations. That behavior does not by itself change the system’s legal classification, but it may change the clinical risk that the organization needs to manage. Human oversight should therefore be assessed as an operating control, not merely described in a policy.
Clinical AI risk belongs to the whole system of care
AI-related patient harm does not arise only from algorithms. Risk may also come from poor data quality, inappropriate patient selection, integration errors, interface design, alert fatigue, insufficient training, automation bias, cybersecurity failures, changing workflows, unclear accountability or inadequate fallback arrangements.
A model can perform well in a study and still perform differently after deployment because the local patient population, equipment, disease prevalence, workflow or data pipeline differs. A technically accurate alert can become ineffective if clinicians receive too many alerts to respond meaningfully. A useful recommendation can become unsafe if staff misunderstand what it represents.
WHO’s regulatory considerations for AI in health emphasize intended use, lifecycle risk management, validation, data quality, transparency and post-deployment monitoring rather than treating model performance as the only safety issue.[4] More recently, WHO/Europe’s September 2026 work on responsible AI in health identified governance gaps, unclear accountability, fragmented or biased datasets and deficits in AI literacy among important barriers to responsible implementation.[5]
The practical implication is that clinical AI governance needs to manage the interaction between technology, people, workflow and organization.
Use proportionate controls beyond the statutory minimum
Internal clinical risk assessment should not reproduce the legal high-risk regime for every AI application. That would be unnecessary and could blur the distinction between regulatory compliance and organizational governance.
Instead, healthcare organizations can apply controls proportionately. An AI tool with little connection to patient care may require basic ownership, privacy, cybersecurity and operational controls. A system that informs clinical judgment may justify stronger evidence review, local validation, clearly defined human oversight, user competency requirements and post-go-live monitoring. A system capable of materially influencing diagnosis, triage, treatment or another time-sensitive clinical action may warrant stronger controls still, even when the particular system does not fall within the current statutory high-risk list.
The objective is not to change the system’s legal classification. It is to match organizational controls to the risks created by its actual use. Depending on the use case, those controls may include clearer approval authority, tighter use restrictions, local validation, subgroup performance review, monitoring of false-positive and false-negative patterns, incident reporting, change control, periodic review and reliable fallback arrangements. This is healthcare governance, not an alternative regulatory classification scheme.
Keep legal classification and internal clinical risk separate in the AI inventory
A practical mistake would be to create a single field in an AI inventory labelled simply “Risk: High / Medium / Low.” Months later, it may no longer be clear whether “high” means high-risk under the AI Law, high clinical risk, high cybersecurity risk or simply a judgment made by the implementation team.
A better approach is to maintain separate but connected records. The organization should document the statutory AI risk classification, the basis for that classification and the resulting legal requirements. Separately, it should document an internal clinical or organizational risk assessment, including the potential consequences of failure, degree of clinical influence, effectiveness of human oversight, uncertainty in the evidence and controls required for local use.
A system may therefore carry a low-risk classification under the applicable AI framework while receiving more intensive internal clinical governance because of the way it is used within a particular workflow. That is not necessarily a contradiction; the two records answer different questions. Keeping them separate also makes future review easier. If regulation changes, the statutory classification can be reassessed. If the clinical use, workflow or local evidence changes, the internal patient-safety assessment can be updated independently.
Changes in deployment may require clinical review and, in some cases, legal reclassification
Healthcare organizations should not assume that the original classification remains appropriate after meaningful changes to how a system is deployed.
Article 10 of the AI Law provides that a deployer may inherit the provider’s classification, but where the deployer modifies, integrates or changes the function of the system in a way that creates new or higher risks, it must coordinate with the provider to perform reclassification. Decree No. 142 also requires review and reclassification in specified circumstances, including significant changes in function, intended purpose or deployment context that affect the original risk-classification criteria.[1,2]
This can be particularly important in healthcare. An advisory tool may later be integrated into a more automated process. A system may be given a new clinical purpose or deployed in a materially different context. Integration may give its output greater influence over downstream actions.
Other changes may be clinically important without, by themselves, establishing a legal reclassification requirement. Clinicians may begin relying on a system more heavily, the patient population may change or workflow may evolve around the AI. Those developments should still trigger internal clinical review.
When meaningful change occurs, the organization should therefore ask two questions: Has the clinical risk of this use changed? And has anything changed that requires review of the statutory classification or other regulatory obligations? They are related questions, but they are not identical.
The high-risk list can evolve
The current list should not be treated as permanent. Decree No. 142 establishes a process through which provincial authorities, ministries and other competent bodies review AI deployment and propose amendments to the high-risk list, with the Ministry of Science and Technology coordinating the process for submission to the Prime Minister.[2]
That matters in a rapidly evolving field such as healthcare AI. A system outside the current list should not simply be assumed to remain outside it indefinitely. Healthcare organizations should therefore record the legal basis and date used for classification and include regulatory surveillance within their AI inventory and change-control processes.
AI classification does not replace medical-device or other specialized regulation
The AI Law is not the only regulatory framework that may apply. Decree No. 142 recognizes that where an AI system is integrated into a product, good or service governed by specialized legislation, the applicable sector-specific requirements continue to apply alongside AI-specific requirements.[2]
For healthcare, this is particularly important because some AI-enabled software or systems may also fall within the medical-device regulatory framework depending on their characteristics and intended purpose. Viet Nam’s current medical-device framework includes the consolidated medical-device management provisions in Consolidated Document No. 08/VBHN-BYT dated 6 March 2026 and Circular No. 24/2026/TT-BYT, effective from 1 July 2026, concerning risk classification and management measures for medical devices.[6,7]
Other obligations may arise from personal-data protection, cybersecurity, professional practice, contractual arrangements and other applicable legislation. The practical lesson is simple: AI risk classification is one regulatory layer, not the entire regulatory analysis.
A practical Clinical AI Risk Governance Gate
For an AI system used in or around clinical care, particularly one outside the current statutory high-risk list, a healthcare organization can use the following questions as an internal governance gate:
Regulatory status: What is the current statutory AI risk classification and what is the basis for it? Do medical-device or other specialized regulatory requirements also apply?
Clinical function: What exactly does the AI do, where does it enter the care pathway and what decisions or actions can its output influence?
Consequence of error: What could happen if the system produces a false positive, false negative, inaccurate recommendation, delayed result or no result?
Degree of influence: Is the output informational, advisory, prioritizing or linked to a clinical action, and how much influence does it actually have in practice?
Human oversight: Can an appropriately qualified person independently review the case, recognize limitations, reject the output and intervene before harm occurs?
Evidence and local validity: Is the evidence appropriate for the intended use, population and setting, and has the organization established reasonable confidence that the system performs adequately and fits the local workflow?
Population and equity: Could performance differ across clinically relevant patient groups, disease prevalence, equipment, data sources or care settings?
Detectability and reversibility: How readily can an error be recognized, and can the resulting action be stopped or reversed before significant harm occurs?
Scale and exposure: How frequently will the system be used, how many patients may be affected and could a systematic problem propagate across a department, site or network?
Operational controls: Are competency requirements, use restrictions, escalation routes, downtime arrangements and fallback workflows defined?
Monitoring and incidents: What performance, safety, user-behavior and incident signals will be monitored after go-live, and who will review and act on them?
Change and reclassification: What changes in the system, integration, intended purpose, workflow, population or regulatory environment should trigger renewed clinical assessment or review of the statutory classification?
This Clinical AI Risk Governance Gate is a practical organizational patient-safety framework. It is not a statutory AI classification method and does not replace classification, conformity assessment or any other requirement under Vietnamese law.
What this means for healthcare organizations in Viet Nam
Viet Nam’s AI Law, Decree No. 142 and Decision No. 33 give healthcare organizations a much clearer regulatory starting point than existed previously. The next challenge is to use that framework correctly without expecting a statutory category to answer every patient-safety question.
Legal risk classification determines regulatory treatment. It does not by itself establish whether evidence transfers to a local patient population, whether human oversight works under real clinical pressure, whether an alert fits the workflow, whether clinicians are becoming overly dependent on recommendations or how closely the system should be monitored after go-live. Those are healthcare implementation and patient-safety questions.
A useful organizational principle is therefore:
Classify AI according to the law. Govern its clinical use according to the risks it creates in practice.
For some systems, the two assessments will point toward similar levels of scrutiny. For others, they may not. The purpose is not to create a second regulatory regime inside the healthcare organization. It is to prevent the absence of a statutory high-risk designation from becoming an absence of clinical judgment.
Responsible AI governance starts with compliance, but it should not stop there. For healthcare leaders, the question is not only “Is this AI legally high-risk?” It is also “Given what this system actually does for our patients, in our workflow and in our organization, what governance do we need to use it safely?”
Both questions matter. They should inform each other, but they should not be confused.
Professional education notice
This article is intended for professional education and knowledge exchange and does not constitute individualized legal, regulatory, clinical, medical-device, cybersecurity, personal-data or conformity-assessment advice. Clinical risk and the Clinical AI Risk Governance Gate are organizational patient-safety concepts used in this article; they are not statutory risk categories or assessment methods established by the Law on Artificial Intelligence, Decree No. 142/2026/NĐ-CP or Decision No. 33/2026/QĐ-TTg.
A system that warrants stronger internal clinical governance should not, for that reason alone, be described as legally high-risk. Conversely, a low-risk statutory classification should not be interpreted as a determination of clinical safety or effectiveness. Healthcare organizations should assess the requirements applicable to the specific system, intended purpose and deployment context and re-check the regulatory position when the technology, use case or relevant law changes.
References
[1] National Assembly of Viet Nam. Law No. 134/2025/QH15 on Artificial Intelligence. Adopted 10 December 2025; effective 1 March 2026.
[2] 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.
[3] 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.
[4] World Health Organization. Regulatory Considerations on Artificial Intelligence for Health. 19 October 2023.
[5] WHO Regional Office for Europe. Report of the Knowledge Community on Responsible Artificial Intelligence in Health. 1 September 2026.
[6] Ministry of Health of Viet Nam. Consolidated Document No. 08/VBHN-BYT on Medical-Device Management. 6 March 2026.
[7] Ministry of Health of Viet Nam. Circular No. 24/2026/TT-BYT on Determining Risk Levels and Management Measures for Products and Goods That Are Medical Devices and Amending Certain Provisions of Circular No. 05/2022/TT-BYT. Issued 30 June 2026; effective 1 July 2026.