Generative AI in Healthcare Organizations in Viet Nam: How to Govern LLMs, Clinical Copilots and Staff Use
IMPLEMENTATION INSIGHTS
Generative AI in Healthcare Organizations in Viet Nam: How to Govern LLMs, Clinical Copilots and Staff Use
IMPLEMENTATION INSIGHT
Last reviewed: 2 October 2026
Generative AI is entering healthcare organizations through more than one door. Some systems arrive formally through procurement: a clinical copilot, an AI documentation assistant, a patient chatbot or a generative AI function embedded in an existing health information system. Others arrive much more quietly. A doctor uses a public large language model to improve a referral letter. A quality officer asks AI to summarize an incident report. A nurse uses it to draft patient education material. An administrator uploads a spreadsheet and asks for an analysis.
The governance challenge therefore begins well before an organization purchases a product marketed as “clinical AI.” The practical question is no longer simply whether healthcare staff should be allowed to use generative AI. It is which tools may be used, for which purposes, with what data, under what level of human oversight, and with what evidence that the use is safe and appropriate.
That distinction matters because generative AI is unusually flexible. The same underlying model may help rewrite an email, summarize a medical record, generate discharge instructions, answer patient questions or suggest a differential diagnosis. The technology may look similar across these tasks, but the consequences of an error are very different. For healthcare organizations in Viet Nam, responsible governance therefore needs to focus on the actual use case, not simply on whether a system is described as an LLM, chatbot or copilot.
Generative AI is broader than ChatGPT
Generative AI refers to AI techniques capable of generating new content such as text, images, audio, video or combinations of these. Large language models, or LLMs, primarily work with language, while large multimodal models can accept and generate multiple forms of information.
WHO identifies five broad areas in which large multimodal models may be used in health: diagnosis and clinical care, patient-guided use, administrative and clerical work, health-professional education, and scientific research and drug development. It also identifies important risks, including false, inaccurate, biased or incomplete outputs, automation bias, privacy concerns and cybersecurity threats.[1]
In practice, healthcare organizations may encounter generative AI as general-purpose public tools, enterprise AI assistants, clinical copilots, ambient documentation systems, AI-assisted search, coding support, automated summarization, translation, patient-facing chatbots or generative functions embedded in other clinical software. These applications should not all be governed in the same way.
Using AI to improve the grammar of an internal administrative document is not equivalent to asking it to summarize a complex patient history. Generating ideas for staff education is not equivalent to generating a medication recommendation. Drafting a patient message that will be reviewed by a clinician is not equivalent to allowing an AI chatbot to respond autonomously to a patient describing chest pain. A useful governance model therefore starts with two questions: What is the AI being asked to do, and what happens if its output is wrong?
Generative AI is not a separate legal risk category in Viet Nam
Viet Nam’s Law on Artificial Intelligence No. 134/2025/QH15, effective from 1 March 2026, classifies AI systems as high-, medium- or low-risk. Decree No. 142/2026/NĐ-CP, effective from 1 May 2026, provides the operational framework for applying that classification.[2,3]
The legislation does not create a separate statutory risk category for “generative AI,” “LLMs” or “clinical copilots.” A healthcare organization should therefore not assume that an LLM is automatically high-risk, medium-risk or low-risk merely because it generates content. The relevant classification depends on the particular AI system and how it is deployed.
Under Decree No. 142, an AI system outside the Prime Minister’s high-risk list may be classified as medium-risk where the statutory criteria are met, including where users may be misled, influenced or manipulated because they do not recognize that they are interacting with AI or AI-generated content. The Decree also identifies circumstances that are not classified as medium-risk, including certain office-support systems where the AI function is clear from context and certain systems that do not directly interact with or provide services or content to the public.[3]
Decision No. 33/2026/QĐ-TTg establishes the current list of high-risk AI systems. The healthcare entries currently focus on specified forms of AI-supported surgery and surgical robotics.[4] A generative-AI application should therefore be classified under the applicable legal framework based on the specific system and deployment, rather than by assuming that “clinical GenAI” automatically belongs to a particular statutory category.
There is also a broader healthcare-specific requirement that should not be overlooked. Article 6 of the AI Law provides that AI applications in essential fields that directly affect life, health, lawful rights and interests or public order are subject to stricter risk management appropriate to the sector. For healthcare, the Law specifically identifies patient safety, reliability under real-world conditions and protection of health data.[2]
Legal classification matters, but it should therefore be accompanied by a separate organizational assessment of clinical, operational, privacy and cybersecurity risk.
A general-purpose tool can become part of a clinical workflow through use
One of the most difficult governance issues with generative AI is that the product name tells an organization very little about the risk. A general-purpose LLM may not have been designed or marketed as a medical system. But if clinicians begin using it to interpret symptoms, summarize records, generate treatment suggestions or draft clinical documentation, the tool has entered the clinical workflow.
That does not automatically change the product’s legal status or make it a medical device. It does, however, change the governance question for the healthcare organization. The important principle is: organizations should govern the use case, not only the product.
A healthcare organization cannot reasonably evaluate an enterprise AI platform once and assume that every future use is covered. The same platform may be acceptable for administrative drafting, require additional controls for clinical documentation and be inappropriate for autonomous diagnostic decision-making. “Approved AI” should therefore not mean “approved for everything.” Approval should define permitted uses, permitted data, required review and prohibited or restricted uses.
Fluent output can still be wrong
Generative AI does not function like a conventional clinical database that simply retrieves a stored result. It generates outputs from patterns learned from data and from the context provided in the interaction. As a result, an answer can be fluent, detailed and professionally written while still being wrong.
The problem is not limited to dramatic hallucinations. More subtle failures may be harder to detect: omission of a critical contraindication, inaccurate summarization of a long record, misplaced certainty, fabricated references, loss of temporal context, incorrect interpretation of abbreviations or a recommendation that sounds plausible but does not fit the patient. WHO specifically identifies false, inaccurate, biased or incomplete outputs as important risks of large multimodal models used in health.[1]
Healthcare organizations should therefore avoid assuming that competent staff will simply “notice mistakes.” The likelihood of detecting an error depends heavily on the task. A physician asking AI to improve the wording of a paragraph can readily compare the result with the original. A clinician receiving an AI-generated summary of 150 pages of records may have far less ability to know what was omitted.
The more an AI output substitutes for direct review of underlying information, the stronger the verification and assurance process should be.
Clinical copilots should support professional judgment, not quietly replace it
The term clinical copilot is increasingly used for AI systems that support clinicians with documentation, information retrieval, summarization, recommendations or workflow assistance. The term is useful only if the professional genuinely remains in control.
Article 4 of Viet Nam’s AI Law establishes a human-centered principle: AI is to serve people rather than replace human authority and responsibility, while human control and the ability to intervene must be maintained over decisions and actions of AI systems.[2]
For a clinical copilot, meaningful oversight therefore means more than placing a clinician’s name at the end of the process. The clinician needs sufficient information, competence, time and authority to examine the output and disagree with it.
This is particularly important for AI-generated clinical documentation. A draft consultation note may save time, but inaccuracies can become part of the medical record if staff begin approving polished-looking drafts too quickly. Once incorrect information enters the record, it may influence later clinicians, downstream processes and potentially future AI systems.
Healthcare organizations should define which AI-generated clinical content must be reviewed before it enters the medical record, who is accountable for that review, and what level of verification is expected. The underlying principle is simple: AI may assist the professional task; responsibility for the professional decision remains human.
Staff use is already a governance issue, even without an enterprise AI project
Some of the most important generative-AI risks do not come from systems formally purchased by a hospital or clinic. They arise when staff independently use consumer AI tools that the organization has not evaluated or approved. This is often described as shadow AI.
A healthcare employee may paste a clinical note into a public chatbot to improve the wording, upload a document for summarization, enter patient information while asking for a differential diagnosis or use an AI tool to translate patient instructions. The motivation may be entirely reasonable: staff are often trying to save time, communicate better or reduce administrative burden.
The governance problem is that the organization may not know what information is being transmitted, where it is processed, whether it is retained, whether inputs may be used to improve models, which subprocessors are involved, who can access the information or whether processing occurs outside Viet Nam.
A blanket instruction to “never use AI” is unlikely to be an effective long-term strategy. But having no organizational rule is equally problematic. Healthcare organizations need clear staff-use rules that distinguish approved and unapproved tools, identifiable and non-identifiable information, administrative and clinical uses, and low-consequence versus patient-impacting tasks.
Patient information should not casually enter public AI tools
This is one of the most immediate governance issues for healthcare organizations in Viet Nam. The Law on Personal Data Protection No. 91/2025/QH15 has been effective since 1 January 2026, together with Decree No. 356/2025/NĐ-CP. The implementing framework treats health-related information as sensitive personal data and applies additional protection requirements to its processing.[5,6]
Article 26 of the Personal Data Protection Law specifically addresses health information. As a general rule, the data subject’s consent is required during the collection and processing of personal data relating to health information, except where one of the statutory exceptions in Article 19(1) applies.[5]
The practical implication for generative AI is straightforward: staff should not assume that because a public AI platform is easy to access, patient information may be freely entered into it. Before approving the use of identifiable patient information, the organization should understand the processing chain: where prompts and uploaded files go, where processing occurs, what is retained, whether data may be used for model improvement, which third parties are involved, what administrative controls exist and how information can be deleted.
Cross-border processing requires particular attention. Article 20 of the Personal Data Protection Law treats several activities as cross-border personal-data transfers, including the use by an organization in or outside Viet Nam of a platform located outside Viet Nam to process personal data collected in Viet Nam. Except for the statutory exemptions, organizations carrying out activities covered by Article 20 must prepare a cross-border personal-data transfer impact-assessment dossier and submit it in accordance with the statutory requirements.[5]
The practical rule should therefore not be “never use cloud AI.” Enterprise cloud environments may be capable of being governed appropriately. A more useful rule is: do not place patient or other sensitive personal data into an AI system until the organization has approved the system, the purpose, the data flow and the applicable legal and security controls.
De-identification helps, but it is not a universal solution
Removing a patient’s name does not necessarily make clinical information anonymous. A detailed narrative may still contain dates, rare diagnoses, locations, occupations, family relationships, images or combinations of information that allow a person to be identified.
Healthcare organizations should therefore avoid telling staff simply to “remove the patient name before using AI.” Where de-identification is part of an approved workflow, the organization should define an appropriate method and assess the remaining risk of re-identification. In some circumstances, synthetic data or an organization-controlled environment may be preferable.
The same principle applies to confidential organizational information. Incident reports, internal investigations, contracts, cybersecurity information, staff records and unpublished business information may not all be patient data, but that does not mean they are appropriate for entry into a public generative-AI service.
Patient-facing AI requires additional safeguards
Generative AI becomes a different governance problem when patients interact with it directly.
Article 11 of the AI Law requires providers of AI systems that interact directly with humans to design and operate those systems so that users can recognize when they are interacting with AI, unless otherwise provided by law. The same Article places separate transparency obligations on deployers when certain AI-generated or AI-modified content is made public.[2]
For healthcare organizations, legal transparency should be treated as a starting point rather than the entire patient-safety control. A patient-facing chatbot needs a clearly defined scope. What questions may it answer? Can it discuss symptoms? Can it interpret laboratory results? Can it recommend medication? What happens when a patient describes an emergency? When must the interaction be escalated to a clinician?
Those boundaries should be decided before deployment rather than being left to the model to determine dynamically. Patients should also be able to distinguish general health information from individualized clinical advice and should have a clear route to human support when the AI cannot safely manage the interaction. The greater the potential clinical consequence, the less appropriate it is to rely on a generic disclaimer as the primary safeguard.
Transparency does not mean every AI-assisted sentence needs a label
Healthcare organizations should also avoid over-interpreting Viet Nam’s AI transparency requirements.
Article 11 requires direct human-AI interaction to be recognizable. It also requires a clear notification when AI-generated or AI-modified text, audio, images or video are made public in circumstances where they could create confusion about the authenticity of an event, person or object. Separate labeling requirements apply to certain audio, image and video content that simulate a real person or recreate real events.[2]
Decree No. 142 further provides technical-marking rules for AI-generated or AI-modified content. Machine-readable technical marking is generally not mandatory for text output unless another law provides otherwise.[3] That does not mean publicly provided AI-generated text can never require notification under Article 11; technical marking and disclosure are different obligations.
At the same time, this framework does not mean that every internal email whose grammar was improved by AI needs an “AI-generated” label. Healthcare organizations should distinguish statutory transparency requirements from additional organizational disclosure practices and define those practices deliberately.
A clinical copilot may also raise medical-device questions
AI regulation is not the only legal layer that may apply.
Under Viet Nam’s current medical-device framework, software can fall within the definition of a medical device where the relevant legal conditions and intended medical purposes are met.[7] This does not mean that every clinical copilot, LLM or generative-AI application is a medical device. Whether a particular software product falls within the framework depends on its characteristics, intended purpose and applicable regulatory criteria.
Circular No. 24/2026/TT-BYT, effective from 1 July 2026, forms part of the current framework governing medical-device risk classification and management measures.[8] A general writing assistant and a software system intended to generate patient-specific diagnostic or therapeutic information are therefore not necessarily treated in the same way.
Healthcare organizations should avoid assuming that AI classification under the AI Law is the entire regulatory analysis. Depending on the use case, medical-device regulation, personal-data protection, cybersecurity, professional-practice requirements, contractual obligations and other specialized legislation may all be relevant. Viet Nam’s Law on Cybersecurity No. 116/2025/QH15 has also been effective since 1 July 2026, reinforcing the need to address cybersecurity alongside AI and data governance.[9]
Vendor review needs to go beyond a demonstration
Generative AI can be particularly convincing in a product demonstration because it produces fluent answers immediately. A successful demonstration, however, is not evidence of reliable clinical performance.
Before approving a generative-AI system, healthcare organizations should ask much more specific questions. What model or models power the product? Can the vendor replace them without notification? Where are prompts, files and outputs processed? Are they retained? Can they be used for model training or improvement? Which subprocessors are involved? What safeguards address unauthorized access, prompt injection or data leakage? What evidence supports the intended clinical use? How has performance been tested across relevant patient groups, languages and workflows? What happens when the underlying model changes?
For Vietnamese healthcare organizations, language performance deserves particular scrutiny. Strong performance in English should not be assumed to translate directly to Vietnamese clinical terminology, local abbreviations, mixed Vietnamese-English medical records or local communication patterns.
Vendor evaluation should therefore examine not only the claimed capability of the model, but also the specific configuration being purchased and the environment in which it will actually be used.
Model updates create a hidden change-control problem
Traditional software may remain relatively stable between defined releases. Generative-AI systems can change more dynamically. The underlying model may be replaced, system prompts modified, retrieval sources changed, safety filters updated or new tools connected. A cloud provider may alter model behavior while the visible user interface remains almost unchanged.
A system can therefore look the same to staff while behaving differently.
Healthcare organizations should include generative AI within formal change-control processes. Contracts and internal governance arrangements should define which changes require notification, review, testing or revalidation.
A statement that a vendor “continuously improves the model” is not sufficient assurance for a clinical environment. Improvement in aggregate does not establish unchanged or improved performance for a particular healthcare organization, language, population or workflow.
Monitoring should look for more than technical uptime
Once a generative-AI system is live, monitoring should not stop at whether the service is available. Healthcare organizations should consider signals such as clinically significant inaccuracies, omitted information, inappropriate certainty, staff correction rates, signs of over-reliance, unexpected changes after model updates, patient complaints, near misses, privacy or cybersecurity incidents and use outside the approved scope.
For documentation systems, the organization may need to determine whether AI reduces administrative burden without increasing documentation errors. For patient communication, it may need to review whether information remains understandable and clinically appropriate. For clinical copilots, real-world performance should be monitored rather than assumed from pre-deployment evaluation.
WHO/Europe’s September 2026 work on responsible AI in health identified fragmented and biased datasets, governance gaps, unclear accountability and AI-literacy deficits among key barriers to responsible deployment. Strong governance frameworks, codesign with patients and clinicians, transparency and equitable access were identified as important enablers.[10]
Training staff is part of the control system
AI literacy should mean more than teaching employees how to write better prompts. Healthcare staff need to understand what generative AI can and cannot do, what information they may enter, which tools are approved, which uses are restricted, how outputs should be verified, how common failure modes present themselves and how suspected AI-related incidents should be reported.
Different groups need different competencies. Administrative teams using AI for low-consequence drafting do not need the same training as clinicians using a copilot in patient care. IT teams need to understand integration and cybersecurity. Quality and patient-safety teams need to understand monitoring and incidents. Privacy and legal teams need to understand data flows and contractual responsibilities. Senior leadership needs enough AI literacy to make governance decisions without relying entirely on vendors or technical specialists.
Safe generative-AI use is therefore an organizational capability, not merely an individual digital skill.
A practical Generative AI Governance Gate
Before approving a generative-AI use case, a healthcare organization can ask:
Purpose: What problem are we trying to solve, and is generative AI an appropriate tool for it?
Tool: Is the system organization-approved, and what models, infrastructure and vendors are involved?
Data: What information will users enter, and have personal-data, health-data, confidentiality and cross-border processing requirements been addressed?
Clinical consequence: Could an incorrect, incomplete or misleading output affect diagnosis, treatment, triage, documentation or another patient-care decision?
Human oversight: Who reviews the output, what must be verified, and can that reviewer realistically detect an error?
Evidence: What evidence supports this use in the intended population, language, workflow and setting?
Transparency: Do patients, staff or other users need to know that they are interacting with AI or receiving materially AI-generated content?
Integration: Where will the output enter the workflow, and could it automatically become part of the medical record or trigger downstream action?
Vendor controls: What arrangements apply to retention, model training, subprocessors, cybersecurity, model updates, incident notification and service continuity?
Monitoring: What errors, incidents, user behaviors and performance indicators will be reviewed after deployment?
Change control: What model, prompt, retrieval, integration or workflow changes require reassessment?
Accountability: Who owns the use case, who can restrict or suspend it, and who is responsible for responding when something goes wrong?
This Generative AI Governance Gate is a practical organizational framework. It is not a statutory AI classification method and does not replace legal, medical-device, personal-data, cybersecurity, clinical or professional requirements applicable to a particular system.
Organizations need an approved-use model, not only an AI policy
A long AI policy that staff rarely read will not solve the everyday problem of generative-AI use. Healthcare organizations need operational rules that people can actually apply.
An effective approved-use model should make four things immediately clear: which tools are approved, which data may be used, which tasks require human verification, and which uses require additional approval or are prohibited.
For example, an organization may allow an approved enterprise AI environment for drafting non-clinical communications while prohibiting identifiable patient information in public consumer tools. It may permit AI-assisted clinical documentation only within an approved system and with mandatory clinician review. It may require formal evaluation before any generative-AI system is used to produce patient-specific diagnostic or treatment recommendations.
The exact boundaries will differ between organizations. The important point is that staff should not have to invent their own governance every time they open an AI tool.
Do not govern every generative-AI use as if the risk were the same
Overly restrictive governance can create its own problems. If every AI-assisted task is treated as if it were a high-risk clinical activity, staff may simply move outside the formal process and use consumer tools privately. Governance then becomes less visible rather than safer.
A more effective approach is proportionate. Low-consequence administrative uses can move through a simpler pathway. Uses involving sensitive information need stronger privacy and security controls. Uses that influence patient care require evidence, human oversight, validation, monitoring and incident management proportionate to the potential harm.
The purpose is not to make generative AI difficult to use. It is to make appropriate use easier than unsafe use.
What this means for healthcare organizations in Viet Nam
Generative AI is likely to become part of everyday healthcare work whether organizations introduce it deliberately or allow it to emerge informally. The governance opportunity is therefore to move early from individual experimentation to organizational control.
Healthcare organizations should know which generative-AI systems are being used, what staff use them for, what information enters them and where outputs go. They should distinguish low-consequence drafting from clinical decision support, public consumer tools from controlled enterprise environments, and productivity assistance from applications capable of materially affecting a patient.
Legal compliance is also layered. The AI Law and Decree No. 142 provide the AI-specific framework. Personal-data law governs how personal information, particularly health information, is processed. Medical-device regulation may apply to certain software based on its intended purpose. Cybersecurity obligations continue to apply. Professional accountability does not disappear because a recommendation or document was generated with AI.
There is also a transitional issue. Article 35 of the AI Law provides that providers and deployers of AI systems that had already been put into operation before 1 March 2026 have 18 months from the Law’s effective date to meet the Law’s compliance obligations for systems in healthcare, education and finance.[2]
That provision should not be read as a general 18-month exemption from other applicable laws, nor does it create a transition period for systems first put into operation after 1 March 2026. During the transitional period, an existing system may also be required to suspend or cease operation if the competent AI authority determines that it presents a risk of serious harm.
The most useful organizational question is therefore not simply “Can our staff use generative AI?” It is “For which purposes can they use it safely, with which data, under which controls, and who remains accountable?”
Generative AI can support healthcare professionals, reduce administrative burden and improve access to information. But healthcare organizations should not allow ease of access to become ease of uncontrolled deployment. The systems that are easiest to use may be the ones for which clear organizational boundaries are most urgently needed.
Responsible use begins when the organization knows where those boundaries are—and can demonstrate that they work in practice.
Professional education notice
This article is intended for professional education and knowledge exchange. It does not constitute individualized legal, regulatory, clinical, medical-device, cybersecurity, personal-data or conformity-assessment advice.
Terms such as clinical copilot, shadow AI and Generative AI Governance Gate are used here as practical governance concepts and are not statutory AI risk categories under Viet Nam’s Law on Artificial Intelligence. The statutory classification of a particular system should be determined under applicable Vietnamese law based on the specific system and deployment. A generative-AI application, LLM or clinical copilot should not be assumed to belong to a particular statutory risk category solely because of its technology label.
Similarly, use of a general-purpose AI system in a clinical workflow does not, by itself, establish that the product is legally a medical device. Medical-device status depends on the applicable definition, intended purpose and relevant regulatory requirements. Healthcare organizations should separately assess the AI, personal-data, medical-device, cybersecurity, professional-practice and other requirements applicable to each use case.
References
[1] World Health Organization. Ethics and Governance of Artificial Intelligence for Health: Guidance on Large Multi-Modal Models. 2024.
[2] National Assembly of Viet Nam. Law No. 134/2025/QH15 on Artificial Intelligence. Adopted 10 December 2025; effective 1 March 2026.
[3] 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.
[4] 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.
[5] National Assembly of Viet Nam. Law No. 91/2025/QH15 on Personal Data Protection. Adopted 26 June 2025; effective 1 January 2026.
[6] Government of Viet Nam. Decree No. 356/2025/NĐ-CP Detailing a Number of Articles and Measures for Implementation of the Law on Personal Data Protection. Issued 31 December 2025; effective 1 January 2026.
[7] Ministry of Health of Viet Nam. Consolidated Document No. 08/VBHN-BYT on Medical Device Management. 6 March 2026.
[8] 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.
[9] National Assembly of Viet Nam. Law No. 116/2025/QH15 on Cybersecurity. Adopted 10 December 2025; effective 1 July 2026.
[10] WHO Regional Office for Europe. Report of the Knowledge Community on Responsible Artificial Intelligence in Health. 1 September 2026.