Which Healthcare AI Systems Are High-Risk in Viet Nam? What Decision No. 33/2026/QD-TTg Means for Healthcare Organizations
Decision No. 33/2026/QD-TTg identifies high-risk AI systems in Viet Nam. What does the healthcare list cover, what does it not cover, and what should healthcare organizations do now?
POLICY & REGULATORY INSIGHTS
Which Healthcare AI Systems Are High-Risk in Viet Nam? What Decision No. 33/2026/QD-TTg Means for Healthcare Organizations
Artificial intelligence is moving rapidly into healthcare. Hospitals and clinics are beginning to encounter AI in medical imaging, endoscopy, clinical decision support, documentation, patient communication, remote monitoring, workflow management and increasingly sophisticated medical devices.
Since Viet Nam’s Law on Artificial Intelligence took effect in March 2026, an important question has emerged for healthcare organizations: is an AI system high-risk simply because it is used in healthcare?
The answer is no.
On June 30, 2026, the Prime Minister issued Decision No. 33/2026/QD-TTg, promulgating Viet Nam’s List of High-Risk Artificial Intelligence Systems. The Decision took effect on August 15, 2026. Healthcare is one of the sectors covered, but the current list does not classify every healthcare AI application as high-risk. Instead, it identifies specific systems and the circumstances in which they fall within the high-risk category.
This distinction matters. An AI system used for appointment scheduling, an AI-assisted endoscopy system, a medical-imaging algorithm, a generative AI documentation tool and an autonomous surgical robot should not automatically be treated as if they have the same legal risk classification. At the same time, being outside the current high-risk list does not mean that a healthcare AI system is risk-free, lightly regulated or exempt from clinical governance.
Understanding both sides of that distinction is essential for responsible AI implementation in Vietnamese healthcare.
How Viet Nam’s AI risk framework works
The Law on Artificial Intelligence No. 134/2025/QH15, effective March 1, 2026, classifies AI systems into three risk levels: high risk, medium risk and low risk. Under Article 9, a high-risk AI system is one capable of causing significant harm to life, health, the lawful rights and interests of organizations or individuals, national interests, public interests or national security.
The Law also gives healthcare particular importance. Article 6 requires stricter risk management for the use of AI in essential fields that directly affect life, health, lawful rights and interests, or public order and safety. In healthcare, the Law specifically refers to patient safety, reliability under real-world conditions and protection of health data.
That broader requirement applies to the use of AI in healthcare as a field, but it should not be confused with formal classification as a high-risk AI system.
Decree No. 142/2026/ND-CP, effective May 1, 2026, provides the more detailed framework for risk classification. Under the Decree, an AI system is legally classified as high-risk when it is included in the List of High-Risk AI Systems promulgated by the Prime Minister.
The criteria used in developing and updating that list include the potential impact on life, health, property and rights; the degree of automation; the system’s role in final decision-making; the ability of humans to supervise and intervene; the sector in which the system is used; and the scale of potential impact.
The Decree also identifies circumstances in which a system meeting one or more general risk criteria should ordinarily not be proposed for inclusion in the high-risk list. These include systems whose functions are limited to collecting, processing, standardizing, classifying, translating or improving data without directly creating decisions affecting lawful rights and interests; systems with meaningful human oversight allowing an authorized person to independently review, intervene in, reject or change a decision before it takes effect; systems used only for internal administrative activities without directly affecting external parties; and systems that provide analysis, forecasts, assessments or recommendations that remain advisory rather than serving as the sole basis for a final decision.
There is a limited exception. Even where one of these circumstances applies, a system may still be considered for inclusion in the high-risk list where measurable technical or practical evidence demonstrates a serious vulnerability or risk capable of directly affecting national defense, security, public order or safety.
This is why the statement that “healthcare AI is high-risk” is too broad under the current Vietnamese legal framework.
What does Decision No. 33/2026/QD-TTg actually list in healthcare?
Under the current version of Decision No. 33/2026/QD-TTg, the healthcare section of the official high-risk list contains two entries. Both concern AI systems supporting surgery and surgical robots, but the two entries are legally distinct and are assigned different conformity-assessment routes under Article 13 of the AI Law.
The first entry covers AI systems supporting surgery or surgical robots where AI is integrated into a robot that participates directly in an intervention or guides surgical manipulation, or is integrated into a fully autonomous or intelligent robot that performs actions in place of human behavior.
The second entry also concerns AI systems supporting surgery or surgical robots, but it is more specifically directed at machinery or robotic equipment using AI for autonomous control. It applies where AI is integrated into a robot that directly carries out a therapeutic action on a patient’s body according to a medical order and performs that action without requiring confirmation from a healthcare professional for each change in parameters.
The practical point is important. Decision 33 does not simply say that every product marketed as a “surgical robot” is automatically treated in exactly the same way. The actual functionality of the system, the extent of autonomous control, the relationship between the AI and the physical intervention, and the role of healthcare-professional confirmation all matter.
A product name, marketing description or the fact that a device contains AI is therefore not enough by itself to determine the applicable classification.
The two healthcare entries do not have the same conformity-assessment route
Article 13 of the AI Law requires high-risk AI systems to undergo conformity assessment before being put into use and again where significant changes affect the original assessment.
The Law provides different conformity-assessment routes. For high-risk systems falling within the category that requires conformity certification before use, assessment must be performed by a registered or recognized conformity-assessment organization. For other high-risk systems, the provider may conduct the assessment itself or engage a registered or recognized conformity-assessment organization.
Decision 33 assigns the first healthcare entry to the latter route. The provider may therefore conduct the conformity assessment itself or use a qualified external organization. The second, more autonomous robotic-treatment category is assigned to the stricter route requiring conformity certification before the system is put into use, with the assessment performed by a registered or recognized conformity-assessment organization.
This distinction matters for procurement. A healthcare organization considering an AI-enabled surgical system should not ask only whether the product has been “assessed.” It should understand which legal route applies, who performed the assessment, which version and intended use of the system were covered, what evidence formed the basis of the assessment, and whether subsequent changes have affected the original determination.
Are medical imaging AI, AI-assisted endoscopy and clinical decision support automatically high-risk?
Under the current version of Decision 33, these systems are not automatically classified as high-risk simply because they are used in healthcare.
The healthcare section does not currently list all AI used for radiology interpretation, pathology, AI-assisted endoscopy, diagnostic decision support, risk prediction, remote patient monitoring, clinical documentation, hospital administration or generative AI. Its current scope is narrower and focuses on the two surgical and robotic configurations described above.
That does not mean other forms of healthcare AI are unregulated or clinically low-risk. Article 6 of the AI Law continues to require stricter risk management for AI used in healthcare, while other legal and regulatory frameworks may apply depending on the technology, intended purpose, data involved and operating environment.
A more useful assessment therefore starts with practical questions. What does the system actually do? What is its intended use? Which clinical decisions or physical actions does it influence? How autonomous is it? What meaningful human oversight remains? Which patients or users may be affected? What data does it process? What other regulatory frameworks may apply?
This is more reliable than starting with the label “AI” alone.
“Medium risk” does not mean moderate clinical risk
Another important point is that the terms high risk, medium risk and low risk under Viet Nam’s AI Law are legal classifications. They should not be treated as a complete measure of clinical or patient-safety risk.
Under Decree 142, an AI system that is not on the high-risk list may be classified as medium risk where it can create confusion, influence or manipulation because users do not recognize that they are interacting with an AI system or with AI-generated content, subject to the more detailed conditions and exclusions provided by the Decree. Systems that fall into neither the high-risk nor medium-risk categories are legally classified as low risk.
This has an important implication for healthcare. An AI application may potentially fall into a lower legal risk category under the AI-law taxonomy while still creating clinically meaningful risks that require careful evidence review, medical-device analysis, cybersecurity controls, health-data protection, human-factors assessment or patient-safety governance.
Legal classification should therefore not replace a healthcare organization’s own assessment of clinical, operational, information-security and patient-safety risk.
Put another way: not legally high-risk does not necessarily mean low clinical risk.
The high-risk list can change
Decision 33 should be understood as the current high-risk list, not as a permanent statement that other healthcare AI applications can never become high-risk.
Decree 142 establishes a process through which ministries, provincial authorities and other relevant bodies can review AI deployment and propose changes to the list. Ministries and ministerial-level agencies review systems within their areas of responsibility, while the Ministry of Science and Technology coordinates, appraises and submits proposed amendments to the Prime Minister.
Healthcare organizations should therefore treat regulatory monitoring as part of AI governance. An organization should know which AI systems it is using, whether their functions are changing, whether new modules or models are being added, and whether the regulatory classification applicable to them changes over time.
This matters because an AI system itself may change even when its commercial name stays the same. Software updates, model replacement, additional functionality, increased autonomy, changes in data inputs, new integrations or an expanded intended use may materially alter the system’s risk profile.
Significant changes can trigger reassessment
Decree 142 provides further detail on circumstances in which conformity reassessment may be required for high-risk AI systems.
Relevant changes may include changes to the system’s principal function, intended purpose or scope of application that create new risks or alter the assessed level of risk; changes to system architecture, the AI model or key technical configurations that may affect accuracy, reliability, safety or controllability; material changes to data sources or principal input data; integration with other systems or deployment in a new operating environment; and other changes that materially affect compliance with the high-risk management requirements under Article 14 of the AI Law.
For healthcare organizations, the relevant procurement question is therefore not simply whether the product was compliant when it was purchased. Organizations also need a process for identifying when a software update, model change, hardware change, new integration or expanded intended use requires renewed review or formal reassessment.
That question is particularly important for AI because many systems may evolve more frequently than traditional medical technologies.
A hospital may be a deployer, not simply a customer
Healthcare organizations also need to understand their own legal role.
The AI Law distinguishes among developers, providers, deployers and users. A hospital or clinic operating an AI system under its control as part of its professional or service-delivery activities may fall within the role of deployer, depending on the circumstances.
This matters because Article 14 places responsibilities directly on deployers of high-risk AI systems. These include operating and supervising the system within its intended purpose, scope and classified risk level; maintaining data safety and security and the capacity for human intervention; maintaining applicable standards and technical-regulation compliance during operation; meeting transparency and incident-management obligations; providing information needed for regulatory accountability; and cooperating with providers and authorities during inspection, evaluation, post-market review and incident remediation.
A hospital therefore should not assume that AI compliance belongs entirely to the manufacturer, software company or technology vendor.
In practice, these legal responsibilities connect directly with clinical governance. Hospitals need defined accountability, appropriately trained users, access controls, escalation pathways, incident management, change control, performance monitoring and clear mechanisms for communicating with the provider when problems emerge.
Human oversight remains fundamental
Decision 33 states that the use of an AI system does not change, transfer or eliminate the authority and responsibility of the legally competent organization or individual. It also requires human supervision, control and the ability to intervene during system operation.
In healthcare, meaningful human oversight should involve more than a clinician simply clicking “accept” after an AI recommendation appears on a screen. The responsible person needs sufficient information, competence, authority, time and system support to evaluate the output, recognize situations in which it may be unreliable, reject or override it when appropriate, and stop or escalate use when a safety concern arises.
Healthcare organizations therefore need to determine who may rely on AI output, when independent verification is required, how disagreements between clinicians and AI are handled, whether overrides are technically possible and appropriately documented, and who has authority to restrict or suspend a system when patient safety may be at risk.
These are governance questions, but they are also patient-safety questions.
AI conformity assessment is not the same as clinical evidence
Healthcare organizations should also avoid confusing regulatory conformity with evidence that a technology is clinically appropriate.
A conformity assessment addresses whether an AI system satisfies applicable regulatory requirements. By itself, it does not demonstrate that the technology will improve clinical outcomes, perform appropriately in a hospital’s patient population, integrate successfully into its workflow or create sufficient benefit to justify adoption.
A healthcare organization evaluating AI should therefore ask two different questions. Has the system satisfied the applicable legal and regulatory requirements? And is there sufficient evidence to support safe and effective use in our particular clinical environment?
The second question may require consideration of technical and clinical performance, external validation, applicability to the intended population, usability, human factors, workflow integration, local verification and post-deployment performance.
Regulatory compliance and clinical evidence are complementary. One should not be treated as a substitute for the other.
AI risk classification is different from medical-device classification
This distinction is especially important for healthcare AI.
Viet Nam has a separate regulatory framework for medical devices. Decree No. 98/2021/ND-CP, as amended and supplemented, remains part of that framework, with a consolidated reference text reflected in Consolidated Document No. 08/VBHN-BYT dated March 6, 2026. Circular No. 24/2026/TT-BYT, effective July 1, 2026, further regulates the determination of risk levels and management measures for medical-device products.
Depending on its intended medical purpose and characteristics, an AI-enabled product may therefore also need to be assessed under the medical-device framework. Whether a product is a medical device, and how it is classified as a medical device, are separate questions from whether the AI system is legally classified as high-risk under the AI Law.
An AI-enabled surgical robotic system, for example, may have obligations under both frameworks. Meeting the requirements of one framework should not automatically be assumed to satisfy every requirement of the other.
At the same time, Viet Nam’s AI framework seeks to avoid unnecessary duplication. Decree 142 allows valid results from testing, inspection, certification, conformity assessment, specialized regulation, technical standards, technical regulations or product-quality law to demonstrate corresponding AI requirements where those requirements have already been lawfully and adequately assessed and the relevant results remain valid.
Additional assessment is therefore needed for requirements that have not already been covered rather than simply repeating the same assessment without regulatory purpose.
What about health data?
AI risk classification is only one part of healthcare AI governance.
Many clinical AI systems process personal information and health-related data. Viet Nam’s Law on Personal Data Protection No. 91/2025/QH15 and Decree No. 356/2025/ND-CP took effect on January 1, 2026. Healthcare organizations using AI therefore need to consider not only the system’s AI risk classification but also how personal and health data are collected, accessed, processed, disclosed, transferred, stored, protected and potentially reused.
A system may sit outside Decision 33’s current healthcare high-risk list while still raising significant issues involving personal-data protection, medical-device regulation, cybersecurity, professional responsibility, clinical safety, contractual accountability or intellectual property.
A mature AI governance process should therefore never use the question “Is this system high-risk under Decision 33?” as its only screening test.
Serious incidents require a defined response process
Viet Nam’s AI framework also places significant emphasis on what happens after deployment.
Under Article 19 of Decree 142, a serious AI incident includes an event occurring during operation of an AI system that results in death or serious harm to human health; significant property damage or serious disruption to an organization’s operations; serious infringement of human rights or the lawful rights and interests of organizations or individuals; or serious disruption to public or essential services or effects on national security, public order or safety.
A provider or deployer must submit a preliminary report through the one-stop electronic portal for artificial intelligence within 72 hours from confirmation of urgent serious incidents involving death or serious harm to health, serious disruption of public or essential services or effects on national security, public order or safety, as well as specified uncontrollable serious infringements of rights. Other serious incidents must be preliminarily reported within five working days from confirmation.
Importantly, confirmation does not mean waiting until a complete technical investigation has been finished. The relevant point is when the organization or individual has sufficient initial information to determine that the incident has in fact occurred and is highly likely to originate from a fault of the AI system. Submitting the preliminary report within the required period is not itself treated as an admission of technical fault or legal responsibility.
Providers and deployers are also expected to take measures to limit consequences, cooperate in remediation and preserve relevant data and system logs. An official report on remediation results must subsequently be submitted within 15 days from the preliminary report.
For healthcare organizations, these requirements make it important to design AI incident management before a serious event occurs. The organization should know who receives the first internal notification, who has authority to restrict or suspend use, how an AI-related event connects with existing patient-safety and clinical-risk processes, how the provider will be contacted, what logs and evidence must be preserved, and who determines whether regulatory reporting or parallel reporting under other applicable frameworks is required.
AI safety should therefore be integrated into existing clinical governance, patient-safety, risk-management, information-security and quality-management systems rather than managed as an isolated information-technology issue.
The transitional dates are important
Decision 33 contains specific transitional arrangements that healthcare organizations should not overlook.
For AI systems included in Decision 33 that were already in operation before August 15, 2026, providers and deployers of systems in healthcare, education and finance must complete the applicable AI-law compliance obligations before September 1, 2027. During this transitional period, those systems may continue to operate unless a competent authority determines that a system presents a risk of serious harm and requires suspension or termination in accordance with the law.
A separate rule applies to systems put into operation during the six months following the Decision’s effective date. Providers and deployers of those systems must complete the applicable compliance obligations before March 1, 2027.
The transition timetable therefore depends not only on the sector but also on when the system entered operation. Healthcare organizations should not assume that every system has the same compliance deadline.
These transitional dates are specific to AI systems falling within the high-risk list under Decision 33. They should not be interpreted as a blanket extension of compliance obligations for all healthcare AI. This distinction is particularly important because the transitional provisions address systems already operating before the high-risk list itself took effect, as well as systems entering operation during the six-month period following August 15, 2026.
What should healthcare organizations do now?
For most hospitals and clinics, the most useful first step is not to label every AI system “high-risk.” It is to understand what AI is actually being used, what it does and how it is being used.
Healthcare organizations should maintain an inventory of AI systems already in routine use, under pilot evaluation, embedded within medical equipment, included within software or cloud services, or being considered for procurement. For each system, the organization should understand its intended use, provider, deployment model, clinical role, patient population, degree of autonomy, data flows, human oversight, applicable regulatory status and whether its functionality falls within Decision 33.
Where a system appears to fall within the high-risk list, the organization should determine the respective roles of provider and deployer, confirm the applicable classification, understand which conformity-assessment route applies, examine the relevant evidence and documentation, identify the applicable transition timeline and establish governance for human oversight, incident response, monitoring and significant changes to the system.
Where a system does not fall within the current high-risk list, the assessment should continue rather than stop. The organization still needs to determine its classification under the broader AI framework and consider other applicable requirements involving medical devices, personal data, cybersecurity, clinical governance, patient safety and professional accountability.
In practice, this calls for a risk-proportionate approach. A low-autonomy administrative tool should not necessarily undergo the same level of review as a system that influences diagnosis, treatment or direct physical intervention. But every AI system should enter the organization through a sufficiently structured process to establish what it does, what could go wrong, what regulatory obligations may apply, who is accountable and what level of oversight is appropriate.
Most importantly, legal classification should not become the organization’s only measure of risk.
From compliance to responsible implementation
Decision 33 is important because it makes Viet Nam’s risk-based approach to AI more concrete. For healthcare, however, its current scope is significantly more specific than the phrase “high-risk healthcare AI” might initially suggest.
Under the current version of the Decision, the healthcare section focuses on particular AI-supported surgical and robotic systems, with the more autonomous therapeutic configuration assigned to the stricter conformity-certification route. It does not automatically place every clinical AI application into the high-risk category.
At the same time, the AI Law requires healthcare applications more broadly to be subject to stricter risk management because of their potential impact on life and health. The appropriate conclusion is therefore neither that all healthcare AI is legally high-risk nor that AI outside Decision 33 is of little concern.
For healthcare leaders, the more useful questions are what the system actually does, how autonomous it is, which clinical decisions or physical actions it influences, whether meaningful human oversight remains, what evidence supports its intended use, which regulatory frameworks apply, how patient and health data are protected, and how performance and safety will be monitored after deployment.
Decision 33 should therefore be understood as one component of a broader shift toward risk-based AI governance in Viet Nam, rather than as a complete framework for deciding whether a healthcare AI system is safe or appropriate to use.
Responsible healthcare AI requires regulation, evidence, clinical governance, patient safety, data governance and real-world implementation to work together.
Key References
National Assembly of Viet Nam. Law on Artificial Intelligence No. 134/2025/QH15, dated December 10, 2025, effective March 1, 2026.
Government of Viet Nam. Decree No. 142/2026/ND-CP detailing provisions and measures for implementation of the Law on Artificial Intelligence, dated April 30, 2026, effective May 1, 2026.
Prime Minister of Viet Nam. Decision No. 33/2026/QD-TTg promulgating the List of High-Risk Artificial Intelligence Systems, dated June 30, 2026, effective August 15, 2026.
Ministry of Health of Viet Nam. Consolidated Document No. 08/VBHN-BYT on medical device management, dated March 6, 2026.
Ministry of Health of Viet Nam. Circular No. 24/2026/TT-BYT on determining risk levels and management measures for medical-device products, dated June 30, 2026, effective July 1, 2026.
National Assembly of Viet Nam. Law on Personal Data Protection No. 91/2025/QH15, dated June 26, 2025, effective January 1, 2026.
Government of Viet Nam. Decree No. 356/2025/ND-CP detailing provisions and measures for implementation of the Law on Personal Data Protection, dated December 31, 2025, effective January 1, 2026.
This article is intended for educational and informational purposes. It does not constitute legal, regulatory, clinical, medical-device, procurement or technology advice. The classification and regulatory requirements applicable to a particular AI system should be assessed according to its specific intended use, functionality, degree of autonomy, provider and deployer roles, data practices, operating environment and applicable laws and regulations.
Updated: October 7, 2026