Patient-Facing Healthcare AI in Viet Nam: Governing Symptom Checkers, Chatbots, Red Flags and Safe Human Escalation
PRACTICAL FRAMEWORKS & TOOLS
Patient-Facing Healthcare AI in Viet Nam: Governing Symptom Checkers, Chatbots, Red Flags and Safe Human Escalation
PRACTICAL FRAMEWORKS & TOOLS
Last reviewed: 8 October 2026
A patient may visit a hospital's website or mobile application before deciding whether to seek medical care. Instead of speaking with a receptionist or healthcare professional, the patient may first interact with an AI-powered chatbot.
The conversation may begin with a routine question about appointment availability. The patient then mentions chest discomfort, difficulty breathing or another symptom that could require urgent assessment. The AI responds politely, provides general information and suggests booking an appointment. The response may sound helpful, but if the patient needs urgent care, a routine appointment recommendation could create a dangerous delay.
This illustrates an important difference between using AI for administrative convenience and allowing AI to influence how patients understand symptoms, assess urgency and decide when to seek professional care.
Patient-facing AI can make healthcare information more accessible, answer routine questions and help people navigate services. However, once an AI system begins responding to personal symptoms or recommending a level of care, its outputs may influence decisions with direct consequences for patient safety.
The central governance question is therefore not simply whether a chatbot provides accurate information or a satisfactory user experience. It is whether the healthcare organization can keep AI interactions within appropriate clinical boundaries, recognize situations requiring urgent attention and help patients reach suitable professional care when needed.
For healthcare organizations in Viet Nam, this requires an approach that combines clinical safety, responsible AI governance, clear communication and realistic operational arrangements.
Patient-facing AI is not one type of healthcare application
The term patient-facing AI covers a wide range of technologies. Some systems answer questions about hospital services, opening hours, costs or appointment procedures. Others provide general health education, explain medical terminology, collect information before consultations or help patients understand instructions previously provided by healthcare professionals. More clinically influential applications may ask about symptoms, suggest possible causes, recommend when to seek care or direct patients toward particular services.
These functions should not be governed as though they carry the same risks. An AI assistant providing clinic opening hours generally presents a different patient-safety concern from a symptom checker advising whether someone should seek emergency care. A chatbot explaining how to prepare for a prescribed examination is also different from one suggesting changes to medication.
The important distinction is the clinical influence of the AI interaction, not whether the technology is described as a chatbot, virtual assistant or symptom checker.
Even an administrative chatbot may encounter clinical information unexpectedly. A patient asking to reschedule an appointment may mention severe pain or sudden weakness. Although the system was not designed to assess symptoms, the conversation has entered a potentially important clinical context.
Healthcare organizations therefore need to decide not only which functions the AI is authorized to perform, but also how it should respond when patients raise concerns outside those functions. Describing a chatbot as intended only for administrative support does not prevent people from using it to describe symptoms or seek health advice. The organization needs to anticipate these situations rather than assume they will not occur.
An apparently reassuring answer may still be unsafe
Patient-facing AI can produce inaccurate, incomplete or inappropriate medical information. However, some of the most concerning responses may sound reasonable and reassuring.
A patient may describe symptoms that require urgent assessment but receive advice to monitor the situation at home or arrange a routine consultation. The response may contain accurate general information while recommending an inappropriate level of care.
This is particularly important for symptom checkers and AI-supported triage.
Undertriage occurs when the recommended level or urgency of care is lower than clinically appropriate. Overtriage occurs when the recommendation is more urgent than necessary. Both matter, although their potential consequences differ. Undertriage may delay essential assessment or treatment, while overtriage may lead to unnecessary emergency visits, investigations and additional pressure on healthcare services.
Healthcare organizations should therefore avoid evaluating patient-facing AI solely through overall response accuracy, conversational fluency or patient satisfaction. A system may answer most routine questions correctly while performing poorly in less frequent situations where the consequences of an incorrect recommendation are greatest.
A structured study published in Nature Medicine in February 2026 evaluated ChatGPT Health's triage recommendations using 60 clinician-authored clinical vignettes across 21 medical domains. Each vignette was tested under 16 conditions, producing 960 responses [11].
The researchers reported undertriage in 52% of presentations classified as emergencies according to the study's reference standard. They also found that statements from family or friends minimizing symptoms could influence recommendations in certain borderline situations, sometimes shifting the suggested level of care toward lower urgency.
These findings are important, but they require careful interpretation. The study used simulated clinical cases rather than observing real-world patient outcomes. It evaluated a particular system at a defined point in January 2026, and its findings cannot automatically be generalized to other AI products, later model versions or routine healthcare services in Viet Nam. The reported percentages describe performance within the study's structured testing conditions, not the probability that any particular patient will receive unsafe advice.
Nevertheless, the study highlights a relevant patient-safety concern: a conversational AI system may perform well in ordinary situations while still requiring careful evaluation of its behavior when clinical consequences are potentially serious.
International experience highlights the importance of clinical governance
A different implementation experience was reported in NEJM Catalyst in June 2026.
Boogaard and colleagues described the development, pilot implementation and clinical governance of an AI digital assistant designed to provide general health guidance while directing higher-risk situations to human clinicians [12].
The approach included risk-based classification, safeguards that restricted AI-generated advice in concerning situations, clinical involvement and ongoing review of interactions.
The authors reported encouraging findings within their evaluated implementation. However, this was a specific organizational case study and should not be treated as evidence that every patient-facing AI assistant can achieve the same level of safety. Its operational results also need to be interpreted within the service model, population and evaluation arrangements described by the authors.
The significance of this example lies in how the service was organized. The chatbot was not treated simply as a conversational interface ready for deployment after technical testing. Its functions, boundaries, escalation arrangements and monitoring processes were deliberately designed.
WHO's guidance on large multimodal models in healthcare similarly identifies risks associated with inaccurate information, bias, inappropriate reliance, privacy and the use of AI in decisions affecting health [10]. Its broader 2021 guidance emphasizes human autonomy, safety, transparency and accountability in healthcare AI [13].
Together, these sources support an important principle: the safety of patient-facing AI depends not only on the underlying model, but also on what the organization allows it to do, how patients actually use it and what happens when professional care becomes necessary.
What Viet Nam's current legal framework means
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. Article 6 requires stricter risk management appropriate to essential sectors directly affecting life, health and other protected interests. For healthcare, the Law specifically refers to patient safety, reliability under real-world conditions and protection of health data [1].
Article 11 is particularly relevant to patient-facing applications. Under Clause 1, AI providers must ensure that systems interacting directly with people are designed and operated so that users can recognize when they are interacting with AI, unless otherwise provided by law. The Article also establishes additional transparency responsibilities for providers and deployers concerning specified AI-generated or modified content [1].
Decree No. 142/2026/NĐ-CP, effective since 1 May 2026, further details the AI Law. Article 16 addresses transparency and accountability, including the provision of information about an AI system's intended purpose, scope, conditions of use and limitations, in accordance with applicable requirements [2].
These provisions reinforce the importance of making patient-facing AI interactions understandable. However, the legal responsibilities of providers and deployers are not necessarily identical.
An important distinction arises under Article 3 of the AI Law. A provider is defined in relation to placing an AI system on the market or putting it into use under its own name, brand or trademark, regardless of whether the system was developed internally or by a third party. A deployer uses an AI system under its control in professional, commercial or service activities [1].
Consequently, a healthcare organization offering an AI assistant under its own brand should not automatically assume that purchasing the underlying technology makes it only a deployer. Its legal role should be assessed according to how the system is provided, branded, controlled and operated. Depending on the actual arrangement, both provider and deployer responsibilities may need to be considered.
Decision No. 33/2026/QĐ-TTg, effective since 15 August 2026, establishes Viet Nam's current list of high-risk AI systems. Its healthcare-specific entries concern defined AI-assisted surgical and robotic-surgery systems meeting particular conditions [3].
A healthcare chatbot or symptom checker should therefore not automatically be described as legally classified high-risk merely because it processes health information. Its classification must be assessed against the actual system and applicable legal criteria, including the relevant risk classification and notification provisions.
Equally, an AI system that is not included in the statutory high-risk list may still present significant clinical risk if it gives inappropriate symptom guidance or delays access to care.
Another distinction concerns healthcare regulation. The Law on Medical Examination and Treatment No. 15/2023/QH15, as amended, and Decree No. 96/2023/NĐ-CP establish requirements for medical examination and treatment, including relevant forms of remote care [4, 5].
Not every healthcare chatbot is automatically a remote medical consultation service. A system answering administrative questions or providing general health education may have a different legal status from a service performing functions that fall within regulated medical examination or treatment.
However, descriptions such as “health information only” or “not medical advice” do not, by themselves, determine the legal status of a service if its actual functions involve individualized clinical assessment or other regulated activities.
Where AI-enabled software falls within Viet Nam's medical-device regulatory framework, Decree No. 98/2021/NĐ-CP, as amended, and applicable Ministry of Health requirements must also be considered, including Circular No. 24/2026/TT-BYT where relevant [8, 9].
Personal-data protection is another important consideration. Law No. 91/2025/QH15 on Personal Data Protection and Decree No. 356/2025/NĐ-CP have been effective since 1 January 2026 [6, 7].
The applicable obligations depend on the system's functions, the information being processed and the legal roles of the organizations involved.
Taken together, these instruments provide a foundation for responsible patient-facing AI implementation. They do not establish a universal Ministry of Health safety-review form or one standardized clinical escalation process for every healthcare chatbot.
Define what the AI may do before designing the conversation
Healthcare organizations should first establish what the patient-facing AI is intended and authorized to do.
This matters because a conversational interface can make different functions appear similar to patients. The same chatbot may answer questions about appointments, explain general health information and respond to descriptions of symptoms. From the patient's perspective, these interactions may feel like one continuous conversation. From a governance perspective, they may involve substantially different risks.
An organization may authorize AI to explain clinic services, describe preparation for a scheduled examination or provide general information drawn from approved educational materials. It may decide that individualized interpretation of symptoms, treatment recommendations, medication changes or decisions about the urgency of medical care require additional evaluation or professional involvement.
These boundaries should reflect the system's intended purpose, available evidence, applicable legal requirements and the organization's actual clinical and operational capabilities.
This is particularly important when a chatbot uses a general-purpose language model capable of generating responses across many topics. Technical capability does not establish that every possible response or use case is clinically appropriate or organizationally authorized.
An AI assistant introduced for administrative support should not gradually become an informal diagnostic service simply because patients begin asking clinical questions.
The organization should also decide how the system responds to requests outside its approved scope. Some questions may be redirected to routine professional services. Others may contain information suggesting an immediate safety concern that requires urgent-care guidance.
Even where the chatbot has no authorized symptom-assessment function, patients should have a clear way to obtain urgent-care information without depending entirely on the system's ability to recognize medical keywords.
The objective is not to make every conversation restrictive. It is to ensure that the chatbot does not move into clinically consequential activities without appropriate evidence, authorization and safeguards.
Recognizing red flags requires more than matching medical keywords
One of the most important safety considerations is whether the system can recognize when a patient's message may indicate a need for urgent professional assessment.
Potentially serious symptoms may include severe chest pain, significant breathing difficulty, sudden neurological changes, uncontrolled bleeding or other presentations where delay could cause serious harm.
The relevant clinical signs, combinations and response expectations should be determined using appropriate clinical guidance and professional expertise. A short list of keywords cannot safely represent the full range of urgent medical presentations.
Patients do not necessarily describe symptoms using formal medical terminology. Someone experiencing chest discomfort may describe tightness, pressure, heaviness or a sensation that is difficult to explain. Another patient may minimize symptoms because they do not want to visit a hospital or because someone in the family believes the problem is minor.
These differences matter because AI can miss the clinical significance of a message even when it recognizes individual words.
A responsible approach should therefore consider the meaning of the patient's description, relevant context and uncertainty rather than relying solely on exact phrase matching.
When a message suggests a potentially serious condition, the system should respond conservatively within its authorized role. If the information indicates a possible emergency, prolonged questioning by an AI assistant should not create unnecessary delay before the patient receives appropriate urgent-care guidance.
Depending on the clinical situation, the system may need to direct the patient to seek immediate professional assessment rather than continue a routine chatbot conversation.
In Viet Nam, 115 is the designated medical emergency number under the national telecommunications numbering framework [14]. However, organizations should also consider local emergency-service availability and provide practical instructions appropriate to the patient's circumstances.
The presence of an emergency number in a chatbot response should not be treated as proof that timely assistance will necessarily be available everywhere. Nor should an AI assistant imply that an ambulance has been dispatched unless that action has actually occurred through an authorized service.
A red-flag response should not imply that the AI has established a diagnosis. Its purpose is to recognize that professional assessment may be urgently necessary.
The governing principle is that uncertainty about a potentially serious condition should not be converted into false reassurance simply because the AI lacks enough information to identify a specific diagnosis.
Human escalation must be real, available and understandable
Many healthcare chatbots include functions labeled “Talk to a doctor,” “Contact our team” or “Request human support.”
These options may be useful, but their presence does not establish that a safe escalation process exists. A patient may click a button believing a healthcare professional has received the concern, while the request is actually placed in a general customer-service queue. A chatbot may promise a quick response even though the service is not staffed overnight.
For someone requiring urgent medical attention, this difference may be clinically significant.
Healthcare organizations should distinguish between providing emergency-care instructions, transferring an interaction to available personnel and submitting a request for later contact. These are different services and should not be presented as interchangeable.
If a chatbot offers live human support, the organization should know who receives the request, what professional qualifications are required, when the service operates and what happens when a transfer fails.
Where the concern requires clinical assessment, customer-service staff may facilitate contact but should not be assumed to have the competence or authority to provide clinical triage.
If qualified clinical support is unavailable, the system should not imply that a healthcare professional is actively monitoring the patient's symptoms. It should provide clear guidance for reaching suitable care through another appropriate route.
For urgent concerns, a delayed callback request should not be presented as an adequate substitute for immediate emergency assessment.
A receiving professional may also need relevant details from the conversation, subject to appropriate confidentiality and data-protection safeguards. The information transferred should preserve important symptoms and context without presenting unverified AI conclusions as established clinical facts.
The organization should also determine how to identify an unsuccessful transfer. An electronic message marked as sent does not necessarily mean that a professional has accepted responsibility or that the patient has received the promised support.
For clinical escalation, the process should distinguish successful connection, an unanswered request, a failed transfer and a request that remains pending. Where timely clinical attention is necessary, the system should guide the patient toward an alternative appropriate care route rather than leaving the interaction unresolved.
A successful transfer requires more than sending a message. The organization needs reasonable assurance that the patient understands the next step and that the promised support has actually become available when the service represents it as available.
Human escalation is meaningful only when it connects the patient to an appropriate and operationally available response.
Vietnamese language and patient communication need local evaluation
Patient-facing AI implemented in Viet Nam should be evaluated under realistic Vietnamese communication conditions.
Patients may describe symptoms in formal Vietnamese, regional expressions, abbreviated messages or text without diacritical marks. Some may combine Vietnamese and English medical terms, while others may use voice messages that introduce transcription errors.
A patient may write only a few words without explaining how severe or persistent a symptom is. Another may ask a family member to communicate on their behalf.
These patterns can affect whether AI recognizes an important concern.
A system that performs well on carefully written medical scenarios in English may not perform equally well when patients use everyday Vietnamese. Healthcare organizations should therefore examine how the system responds to realistic local expressions, incomplete information and ambiguous descriptions.
Testing should also consider how AI responds when a patient minimizes a concerning symptom. Someone may say that chest discomfort is probably not serious because a family member has reassured them or because they cannot afford to miss work.
Such statements should not automatically lead the system to recommend a less urgent level of care when the clinical information suggests otherwise.
Practical barriers may also influence patient behavior. Cost, transportation, distance, work commitments and family responsibilities may affect whether people seek care. A responsible system should acknowledge these barriers without allowing them to distort clinically important safety guidance.
Where voice interaction is supported, evaluation should also consider background noise, accent variation, speech-recognition errors and poor connectivity.
Local evaluation should reflect how patients actually communicate, not only how clinicians or developers expect them to communicate.
Some patients may need additional safeguards
Certain patient groups may have greater difficulty providing complete information or understanding the limitations of an automated service.
Children, older adults, people with cognitive or communication difficulties and patients experiencing significant emotional distress may require additional attention.
A parent asking about a child's symptoms may provide incomplete information. An older person may have difficulty describing the onset or severity of a condition. A patient experiencing psychological distress may communicate risk indirectly rather than through an explicit request for emergency help.
Patient-facing AI should not assume that every user can provide complete, accurate and clinically structured information. Healthcare organizations should consider whether their applications are suitable for these users and whether additional restrictions, human support or alternative communication channels are necessary.
Mental health and self-harm-related concerns require particular care. A chatbot providing general emotional support should not be assumed capable of safely evaluating or managing an acute mental health crisis without appropriate evidence, safeguards and professional support.
In potentially urgent situations, the service should avoid unsupported reassurance and provide suitable routes to timely human assistance or emergency care.
Interactions involving caregivers also deserve attention. A family member may ask questions on behalf of a patient, creating uncertainty about identity, consent, clinical information and who should receive subsequent communication.
These considerations do not mean vulnerable patients should be excluded from the potential benefits of digital services. They mean that accessibility and safety should be designed together.
Transparency must match what the chatbot actually does
Patients should understand when they are interacting with AI rather than a healthcare professional.
This matters for compliance with applicable transparency requirements, but it also affects how people interpret the information they receive.
A chatbot using a hospital's logo, a professional-looking interface and confident language may appear highly authoritative. Some patients may assume that its responses have been reviewed by a doctor or represent individualized clinical assessments.
Clear communication should therefore explain the nature of the service, its intended purpose and important limitations.
Where the AI provides general health information rather than individualized clinical care, that distinction should be understandable. Where human review is offered, the system should make clear whether the review has actually occurred.
A generic disclaimer should not be treated as a substitute for safe system behavior.
For example, a chatbot that states it does not provide medical advice may still create patient-safety risks if it subsequently gives personalized reassurance about potentially serious symptoms.
Transparency should therefore be consistent with the AI system's actual functions and behavior. The service should also avoid creating the impression that a clinician is available, monitoring conversations or assuming responsibility when this is not the case.
The objective is not to overwhelm patients with technical explanations. It is to help them understand who or what is providing the information, what the information can reasonably support and how to obtain professional care when necessary.
Patient information and conversation data require appropriate governance
Patient-facing AI may receive sensitive information even when the organization originally intended the chatbot for general administrative support.
Patients may voluntarily disclose symptoms, medications, previous diagnoses, pregnancy-related concerns or other personal information through website chat, mobile applications or messaging services.
Healthcare organizations should therefore understand what information is collected, where it is processed, who can access it and how it is retained or transferred.
The applicable requirements under Viet Nam's Law on Personal Data Protection and Decree No. 356/2025/NĐ-CP must be assessed according to the actual processing activities and the roles of the organizations involved [6, 7].
Where consent is legally required, suitable arrangements must be considered. Any other applicable legal basis or exception should be examined under the relevant provisions rather than assumed.
If the chatbot relies on a third-party AI platform, the healthcare organization should understand whether patient information is transmitted to external services, retained or used for purposes beyond the immediate interaction.
Processing or transferring personal data outside Viet Nam may create additional legal obligations.
The organization should also determine what information needs to be retained for patient safety, continuity of care and accountability. Not every administrative conversation necessarily belongs in the medical record. However, information that becomes materially relevant to clinical assessment or treatment may require appropriate documentation under the applicable clinical-record requirements.
These decisions should balance patient safety, confidentiality, data minimization and the need for authorized professionals to access relevant information.
A chatbot should not collect extensive medical histories simply because its interface makes doing so easy.
Testing must reflect complete patient interactions
Before routine deployment, healthcare organizations should evaluate patient-facing AI using scenarios that reflect its intended functions and foreseeable uses outside its authorized scope.
Testing should go beyond asking whether the system provides medically accurate answers to isolated questions.
A useful scenario may begin with an administrative request and develop into a potential emergency. Another may involve a patient describing symptoms vaguely, changing details during the conversation or providing information that contradicts an earlier statement.
Other scenarios may involve caregivers, communication barriers, unavailable clinical staff, failed transfers or a patient continuing to request reassurance after receiving urgent-care guidance.
The purpose is to examine how the system behaves throughout the interaction, not merely whether one response contains an appropriate warning.
For clinically consequential functions, scenarios should be assessed by appropriately qualified professionals using relevant clinical guidance. The organization should examine false reassurance, missed red flags, excessive escalation, misleading explanations and failures to connect patients with suitable care.
A high overall accuracy score should not compensate for repeated failures in high-consequence situations.
Testing should also examine cases where the AI recognizes a possible emergency but the operational process fails. A warning has limited practical value if it directs patients to an unavailable service or incorrectly states that a clinician has been notified.
Healthcare organizations should not assume that an AI system will recognize every serious condition. Where a patient-facing application cannot safely evaluate an uncertain situation, the response should remain within its authorized purpose and make suitable professional care options clear.
The extent of evaluation should reflect the system's clinical influence and potential consequences. An administrative chatbot does not necessarily require the same level of assessment as an AI symptom checker intended to influence care-seeking decisions.
Where appropriate, organizations may begin with a limited scope, controlled testing and a supervised pilot before expanding availability.
The guiding question is whether the complete patient interaction remains acceptably safe under realistic conditions, including when the AI is uncertain, the patient provides incomplete information or human assistance is unavailable.
Monitoring should focus on patient safety, not only satisfaction
After deployment, healthcare organizations should monitor patient-facing AI according to its actual risks.
Common chatbot performance measures include response speed, number of conversations, user satisfaction and the proportion of inquiries resolved without human assistance. These indicators may be useful for service management, but they do not establish clinical safety.
A chatbot could reduce the number of patients contacting staff while failing to recognize some people who genuinely need professional attention.
For applications with clinical implications, monitoring should examine whether potentially serious symptoms were recognized, whether appropriate guidance was provided and whether escalation occurred when necessary.
Organizations may also consider reviewing cases where patients received low-urgency recommendations, particularly when subsequent information suggests that more urgent assessment may have been appropriate.
Escalation performance deserves attention. Useful measures may include successful transfers to human support, failed transfers, promised callbacks that were not completed and patients who ended conversations before receiving necessary guidance.
However, these measures should not be interpreted mechanically.
A high escalation rate is not automatically evidence of safety if the system directs large numbers of low-risk patients to already constrained services. A low escalation rate is not necessarily evidence of efficiency if clinically important concerns are being missed.
Organizations should therefore balance the risks of undertriage, unnecessary escalation and failure to complete promised support.
Review should also consider differences in performance across patient groups, Vietnamese language patterns, communication channels and service hours.
Patient complaints, frontline concerns and safety incidents may reveal weaknesses that are not apparent in aggregated performance dashboards. Reviewing relevant samples of actual conversations, within appropriate privacy and access controls, may also help identify misleading reassurance, unexpected uses and gaps in the escalation process.
Where a potentially unsafe interaction is identified, the organization should determine whether immediate safeguards are needed, including restricting a function, correcting information, reviewing relevant cases or involving appropriate clinical personnel.
Monitoring should support learning and improvement rather than serve only as evidence that the chatbot is widely used.
A practical Patient-Facing AI Safety and Human Escalation Review
Healthcare organizations in Viet Nam can structure their assessment through a Patient-Facing AI Safety and Human Escalation Review.
This is proposed as a practical organizational approach. It is not a prescribed Vietnamese regulatory form, Ministry of Health requirement, WHO standard or international certification instrument.
The review should begin by defining the AI service and its authorized functions. The organization needs to establish whether it provides administrative information, general health education, symptom-related guidance or other functions that may influence clinical decisions.
It should then identify situations where a conversation could become clinically consequential. These may include possible emergencies, requests for diagnosis or treatment, medication-related questions, vulnerable users or information that the system cannot interpret reliably.
The next consideration is how the AI responds in those situations. The organization should determine which outputs are permitted, when the system should avoid individualized advice, what urgent-care guidance is appropriate and how users can obtain human assistance.
The review should establish whether the promised human support is genuinely available. It should distinguish live clinical assistance, administrative transfers and requests for later contact, including arrangements outside normal operating hours or during service interruptions. The organization should also define how failed or incomplete transfers are recognized and managed.
Local testing should use clinically reviewed scenarios reflecting Vietnamese language, realistic patient behavior, ambiguous symptom descriptions and foreseeable communication failures.
The review should assess whether patients can recognize that they are interacting with AI and understand the limits of the service.
Personal-data processing, third-party technology arrangements and applicable legal requirements should also be examined according to the system's actual functions and the responsibilities of those providing and deploying it.
Finally, the organization should determine who is responsible for ongoing clinical safety oversight, how concerning interactions will be reviewed and when the service should be modified, restricted or temporarily suspended.
These activities should remain proportionate to the system's risks and can be integrated into existing healthcare quality, patient-safety and digital governance arrangements.
A clinic using AI only for administrative information may need relatively simple controls. A healthcare organization deploying AI-supported symptom guidance at scale may require a more structured clinical safety review and ongoing professional oversight.
The objective is not to create another lengthy compliance process. It is to provide reasonable assurance that the AI service can recognize when routine assistance is no longer appropriate and guide the patient toward suitable care.
Start with a service the organization can safely support
For healthcare organizations in Viet Nam, a practical starting point may be to introduce patient-facing AI for clearly bounded administrative or educational functions.
The organization can establish reliable information sources, define authorized responses and provide straightforward ways for patients to contact existing services.
Before expanding into symptom-related guidance, healthcare leaders should consider whether the necessary clinical evidence, professional oversight and operational capacity are available. This includes understanding what happens when patients raise urgent concerns outside normal working hours.
An AI assistant should not promise immediate clinical review if the organization does not provide that service.
Healthcare organizations should also examine whether the chatbot improves access or unintentionally creates new barriers. A patient who prefers to speak with a person, has difficulty using digital services or cannot clearly describe a problem should still have an appropriate route to care.
The commercial benefits of automation, including reduced call volumes and lower administrative workload, may be legitimate objectives. However, they should not become the primary measures of success when the technology influences decisions about patient care.
Clinical leadership, patient-safety professionals, digital health personnel and relevant service teams should work together to ensure that the AI's functions remain consistent with the organization's actual capabilities.
The question is not whether every healthcare organization needs an advanced AI chatbot. It is whether an organization choosing to deploy one can provide a safe, understandable and appropriately governed service within its professional, technical and operational limits.
From convenient conversations to safer patient navigation
Patient-facing AI can improve access to healthcare information and help people navigate services more conveniently. It may also create new opportunities to support patients before, between and after clinical encounters.
However, safe use depends on more than accurate language generation.
A chatbot may be helpful in routine conversations and still perform poorly when a patient describes serious symptoms, provides incomplete information or needs urgent professional attention.
Healthcare organizations therefore need to govern what AI is allowed to do, how it recognizes clinically important concerns, when it should avoid individualized advice and how patients can reach appropriate human care.
For healthcare leaders, the essential question is: If a patient describes a potentially serious health problem to our AI assistant, can we demonstrate that the system responds safely and does not create an avoidable delay in obtaining professional care?
That question should be considered before launching the service, revisited during implementation and monitored after deployment.
Responsible patient-facing AI does not mean allowing technology to answer every healthcare question. It means using AI where it can provide value while maintaining clear clinical boundaries, safe escalation arrangements and accountability for the service being offered.
Ultimately, the goal is not simply to make healthcare conversations faster or more convenient. It is to help patients receive appropriate information and reach the right care safely, particularly when their needs extend beyond what AI should handle alone.
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 patient-facing AI safety governance and Patient-Facing AI Safety and Human Escalation Review are used in this article as practical organizational concepts. They are not defined statutory categories or prescribed documentation requirements under Vietnamese law.
The approach proposed here is not a Ministry of Health framework, a statutory chatbot approval process, a WHO certification requirement or a specific international standard governing patient-facing AI.
Applicable legal obligations must be assessed according to the AI system's actual functions, intended use, deployment arrangements and clinical context. These may include the legal roles of providers and deployers, AI risk classification and transparency, medical examination and treatment regulation, medical-device requirements, professional responsibilities, personal-data protection, clinical records and other applicable laws.
WHO guidance and international research are used as reference points for responsible governance and patient-safety improvement. Findings from individual research studies should not be generalized automatically to all AI products or healthcare settings.
The proposed review does not replace clinical evaluation, regulatory authorization where required or professional judgment.
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.
National Assembly of Viet Nam. Law No. 15/2023/QH15 on Medical Examination and Treatment, as amended. Effective 1 January 2024.
Government of Viet Nam. Decree No. 96/2023/NĐ-CP Detailing a Number of Articles of the Law on Medical Examination and Treatment, as amended. Issued 30 December 2023; effective 1 January 2024.
National Assembly of Viet Nam. Law No. 91/2025/QH15 on Personal Data Protection. Adopted 26 June 2025; effective 1 January 2026.
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.
Government of Viet Nam. Decree No. 98/2021/NĐ-CP on Medical Device Management, as amended.
Ministry of Health of Viet Nam. Circular No. 24/2026/TT-BYT on Determining Risk Levels and Management Measures for Medical-Device Products and Amending Certain Provisions of Circular No. 05/2022/TT-BYT. Issued 30 June 2026; effective 1 July 2026.
World Health Organization. Ethics and Governance of Artificial Intelligence for Health: Guidance on Large Multi-Modal Models. Geneva: World Health Organization; 2024. ISBN: 978-92-4-008475-9.
Ramaswamy A, Tyagi A, Hugo H, et al. ChatGPT Health Performance in a Structured Test of Triage Recommendations. Nature Medicine. 2026;32(5):1671–1675. Published online 23 February 2026. DOI: 10.1038/s41591-026-04297-7.
Boogaard CO, Marshall J, Shah A, et al. Blueprint for Safety: Implementing a Clinically Governed AI Digital Assistant for Patient Guidance. NEJM Catalyst Innovations in Care Delivery. 2026;7(7). Published online 17 June 2026. DOI: 10.1056/CAT.25.0377.
World Health Organization. Ethics and Governance of Artificial Intelligence for Health: WHO Guidance. Published 28 June 2021.
Ministry of Information and Communications of Viet Nam. Consolidated Document No. 04/VBHN-BTTTT dated 31 March 2022, Consolidating the Circular on the National Telecommunications Numbering Plan. Identifies 115 as the national medical emergency service number.