Governing AI-Assisted Medication Prescribing in Viet Nam: Dose Accuracy, Pharmacist Review and Medication Safety
IMPLEMENTATION INSIGHTS
PRACTICAL FRAMEWORKS & TOOLS
Last reviewed: 20 October 2026
An older patient with diabetes and chronic kidney disease visits an outpatient clinic. An AI-assisted prescribing system reviews the available medical record and recommends a medication at a standard dose. The recommendation appears reasonable, and the medicine is commonly used for the patient's condition.
There is one problem. The system has relied on a kidney function result obtained several weeks earlier. Since that test, the patient has experienced an acute illness that may have affected renal function. The physician must now decide whether the available information is reliable enough to support the recommended dose.
A similar problem can arise when AI calculates an antibiotic dose for a child using an outdated weight measurement or recommends a medication without recognizing another treatment prescribed by a different healthcare facility.
In each situation, the AI recommendation may appear clinically convincing. The calculation itself may even be mathematically correct. What remains uncertain is whether the recommendation is appropriate for the patient receiving treatment.
Medication prescribing involves much more than selecting a drug for a diagnosis. Safe treatment requires consideration of the indication, dose, route, frequency, duration, relevant contraindications, potential interactions and the patient's clinical condition. Some medicines also require laboratory monitoring or dose adjustment as that condition changes.
For healthcare organizations in Viet Nam, introducing AI into prescribing raises practical questions. How can clinicians verify that the system used the right patient information? Who is responsible for confirming a proposed dose? When should a clinical pharmacist become involved? Which warnings need to be resolved before a prescription is finalized, and how should significant medication safety concerns be recorded?
The challenge is to make prescribing support more useful without allowing incomplete information, inaccurate calculations or poorly designed workflows to create new risks for patients.
AI-assisted prescribing is not one type of technology
Healthcare organizations already use different digital tools to support medication decisions.
Conventional electronic prescribing systems may check for drug interactions, allergies, duplicate therapies or doses exceeding predefined limits. Many of these functions rely on established clinical rules and medication databases rather than artificial intelligence.
Other applications use machine learning to identify potentially inappropriate prescriptions, predict medication-related complications or detect patterns in clinical information. Generative AI can summarize medication histories, interpret clinical text or suggest treatment options in response to a healthcare professional's question.
These technologies do not necessarily have the same capabilities, evidence base or regulatory status.
A system that identifies a documented drug allergy presents different risks from one that recommends an individualized anticoagulant dose. There is also an important difference between displaying a recommendation for clinical consideration and automatically transferring medication details into an electronic order.
The closer an AI function comes to influencing a treatment decision, the more important it becomes to understand which information the system uses, how its recommendations are generated and what happens when an output is incomplete or incorrect.
This is particularly relevant to generative AI. A language model may provide a detailed medication recommendation in confident, professional language even when it has misunderstood a clinical detail, relied on an inappropriate reference or failed to recognize that essential information is missing.
A convincing explanation does not necessarily establish a clinically reliable recommendation.
Before implementation, healthcare organizations need to define what the system is intended to do and which clinical decisions it is expected to support. Evidence from conventional medication alerts should not automatically be treated as evidence that generative AI can safely recommend individualized treatment.
What the evidence tells us about medication decision support
Medication safety remains a major international patient-safety priority.
The World Health Organization's Medication Without Harm initiative identifies high-risk situations, polypharmacy and transitions of care as important areas for reducing avoidable medication-related harm. Its policy brief, published on the WHO website in March 2024, emphasizes that medication safety depends on coordinated action involving patients, healthcare professionals, medicines and the systems through which treatment is delivered [1].
Digital decision support can contribute to these efforts, although its benefits vary across applications and clinical settings.
In 2025, Syrowatka and colleagues published an expanded review in BMJ Digital Health & AI, examining 34 articles on medication-related clinical decision support systems [2].
The review found moderate-certainty evidence that computerized order-entry systems incorporating medication-related decision support were associated with fewer medication errors. Evidence for preventing adverse drug events was less certain, while more targeted or improved decision support showed encouraging results.
The researchers also identified unintended consequences, including alert fatigue, inappropriate overrides, excessive dependence on computerized recommendations and errors introduced by the systems themselves.
These findings suggest that medication decision support can improve aspects of prescribing safety. They do not establish that every system will reduce patient harm or that newer AI technologies necessarily perform better than established prescribing tools.
Research from Viet Nam provides a useful example of how digital alerts and clinical pharmacy activities can complement each other.
In April 2026, Nguyen Thi Thu Phuong and colleagues published a study in Hospital Pharmacy involving 51,437 inpatient and outpatient prescriptions for domperidone at a Vietnamese hospital [3].
The ten-year interrupted time-series study examined an integrated clinical decision support system combined with pharmacist-led interventions targeting contraindicated drug interactions, excessive daily doses and inappropriate treatment duration.
The researchers reported substantial reductions in targeted prescribing errors following successive interventions. Importantly, computerized decision support alone did not produce immediate improvements across every measure. Subsequent pharmacist monitoring, education and audit feedback were associated with further improvements in prescribing practices.
The study provides a relevant Vietnamese example of how medication alerts can be integrated with professional pharmacy activities over an extended period.
However, it involved one medication and one institution. Its observational design limits causal interpretation, and the findings cannot establish that similar improvements would occur with other medicines, healthcare facilities or AI applications.
It should also not be interpreted as evidence that generative AI independently improved prescribing safety.
A 2026 scoping review by Al-Taie and colleagues in Future Healthcare Journal identified 36 publications involving AI applications in clinical pharmacy [4]. These included drug-interaction detection, adverse drug effect monitoring, medication-error identification and potentially inappropriate prescribing.
Although the review identified promising applications, much of the available evidence came from retrospective evaluations and early-stage systems. Stronger research is still needed to establish sustained clinical benefits during routine care.
More recent research illustrates why professional review alone may not eliminate the risks of incorrect AI recommendations.
In September 2026, Surdilla and colleagues published a randomized simulation study in the American Journal of Health-System Pharmacy involving 77 pharmacists, pharmacy residents and pharmacy students [5].
Participants evaluated medication-order verification scenarios with or without assistance from a large language model. AI assistance was associated with higher overall accuracy, but the benefit depended strongly on whether the model's recommendation was correct.
When the AI recommendation was correct, participant accuracy reached 80.2%. When the recommendation was incorrect, accuracy fell to 26.1%. Participants followed 73.9% of incorrect AI recommendations in the simulated scenarios.
These findings highlight the risk of automation bias, in which users place excessive trust in computerized recommendations despite the possibility of error.
The study involved a particular language model, a limited number of scenarios and a simulated verification environment. Its findings should not be interpreted as real-world medication-error rates or assumed to apply to all pharmacists and AI systems.
Nevertheless, it raises an important implementation concern. Asking a healthcare professional to review AI output does not automatically guarantee that incorrect recommendations will be detected. Meaningful review requires relevant expertise, trustworthy information and a workflow that supports independent clinical judgment.
Together, the available evidence supports a cautious but constructive approach. Digital prescribing support may improve medication safety, particularly when integrated with appropriate professional practice. Its actual benefits and limitations still need to be evaluated in the healthcare settings where it will be used.
Reliable prescribing begins with reliable patient information
An AI system cannot produce a reliable patient-specific medication recommendation using information it has not received or cannot interpret correctly.
Before accepting a suggested medicine or dose, the responsible healthcare professional needs to consider whether the system has used sufficiently accurate, complete and clinically relevant information.
Depending on the treatment, this may include the patient's age, body weight, diagnosis, severity of illness, kidney and liver function, allergy history, current medications and previous treatment response.
Other information may become important in particular circumstances. Pregnancy or breastfeeding status can influence the suitability of certain medicines. Frailty, nutritional status, electrolyte abnormalities and relevant pharmacogenomic findings may also affect selected treatment decisions.
Not every prescription requires all these data. The information needed depends on the medicine, clinical indication, patient characteristics and potential consequences of an incorrect recommendation.
Consider a child whose antibiotic dose is calculated according to body weight. If the system uses a measurement recorded during an earlier consultation, the calculation may be accurate while the resulting dose is no longer suitable.
The problem is not necessarily a failure of arithmetic. It is a failure to recognize that the information used may no longer reflect the patient's condition.
Kidney function presents a similar challenge.
For medicines requiring renal dose adjustment, the latest laboratory result in the electronic medical record is not automatically recent enough to guide treatment.
A result obtained several weeks earlier may remain useful for a clinically stable patient. The same result may be inappropriate for someone who has subsequently experienced dehydration, acute kidney injury or another condition affecting renal function.
There is no single acceptable age for laboratory results across all medicines and clinical situations.
The relevance of a result depends on the medicine, the patient's clinical stability, the purpose of testing and the potential consequences of prescribing with outdated information.
The method used to assess kidney function also matters. Estimated glomerular filtration rate and creatinine clearance are not interchangeable for every medication. Product information and relevant professional references may specify a particular approach to dose adjustment, while creatinine-based estimates may be unreliable when kidney function is changing rapidly.
Liver function requires similarly careful interpretation when relevant. Individual laboratory values do not necessarily provide a complete assessment of hepatic impairment or establish whether a particular medicine is suitable.
Healthcare organizations should therefore consider whether their prescribing systems make the source, timing and clinical relevance of important patient information available for verification.
Where essential information is missing or uncertain, the system should communicate that limitation rather than generate an apparently definitive recommendation.
This does not mean requiring new laboratory tests before every prescription. It means recognizing when existing information is insufficient for a particular decision and determining whether further verification, assessment or an alternative treatment approach is appropriate.
An incomplete medication history can make a reasonable recommendation unsafe
Medication histories can be difficult to establish when patients receive care from several healthcare providers.
A patient attending a hospital may also take medicines prescribed by another specialist, products purchased from a community pharmacy or traditional medicines and dietary supplements that are not recorded in the hospital's information system.
In Viet Nam, healthcare facilities may use different information systems, and complete medication histories are not always available across organizations.
An AI application connected to one hospital or clinic may therefore assess a proposed prescription without knowing about another medicine the patient is taking.
The suggested treatment may be appropriate for the documented diagnosis but unsafe because of an unrecognized interaction, duplicate therapy or contraindication.
Medication reconciliation becomes particularly important when patients move between healthcare settings or when treatment plans change.
A patient discharged from hospital may continue taking a medicine that was intended to be stopped. Another patient may receive similar treatments from different specialists without realizing that two products contain the same active ingredient.
Allergy information can also be incomplete or misleading.
An electronic record may contain an allergy label without describing the previous reaction. A medication side effect may have been recorded as an allergy, or an important reaction may never have been documented.
These differences matter because an automated system may either generate an unnecessary warning or fail to identify a clinically significant risk.
Healthcare professionals may need to clarify the previous reaction, the implicated medicine and the relevance of the proposed treatment before deciding whether a prescription is appropriate.
The absence of a recorded medication or allergy should not automatically be interpreted as confirmation that no such issue exists.
AI can assist with medication reconciliation and discrepancy detection, but it cannot reliably resolve information that is missing or clinically ambiguous without an appropriate source of verification.
Dose accuracy requires more than correct arithmetic
Medication dose calculation is an attractive application of digital decision support because calculations can be standardized, performed quickly and compared with established reference values.
However, mathematical accuracy does not necessarily establish clinical appropriateness.
The correct dose may depend on the indication, route of administration, formulation, patient characteristics, organ function and stage of treatment.
A medicine may require one dose during initial treatment and another during maintenance therapy. Different indications may also require different dosing regimens.
Errors can arise when a system confuses the amount administered on each occasion with the total daily dose.
In pediatric prescribing, the distinction between milligrams per kilogram per dose and milligrams per kilogram per day is particularly important.
Confusion between milligrams and micrograms, or between different concentrations of oral liquid medicines, can also result in serious errors.
Formulation creates another practical challenge.
An AI system may calculate a theoretically appropriate dose that cannot be administered directly using the available tablet strength, liquid concentration or injectable preparation.
Converting that amount into a practical administration instruction can introduce further risks, particularly when a medicine requires dilution or a complex preparation process.
A clinically acceptable dose must also be consistent with the intended route, dosing interval, treatment duration and relevant maximum limits.
For high-alert medicines, medicines with narrow therapeutic ranges or treatments requiring complex individualized monitoring, additional safeguards may be justified.
Healthcare organizations should therefore evaluate the complete dosing process rather than focus only on the final numerical result.
This includes the reliability of patient measurements, the calculation method, the medication reference used, the available formulation and the proposed administration instructions.
Where appropriately validated calculators or dependable rule-based checks are available, they should not be replaced unnecessarily with calculations generated through unrestricted interaction with a language model.
Generative AI may help explain a dosing recommendation or identify information requiring attention. Safety-critical calculations should rely on appropriately validated methods and relevant professional references.
Medication alerts should reflect the potential for harm
An electronic prescribing system may generate warnings about allergies, contraindications, interactions, dose limits or laboratory abnormalities.
These warnings do not all carry the same clinical significance.
A potentially serious allergy or major dose discrepancy may require immediate attention. A less consequential interaction may be manageable through ordinary clinical assessment, particularly when the prescriber already understands the relevant circumstances.
Treating every warning as equally urgent can create excessive interruptions without necessarily improving medication safety.
Repeated exposure to low-value notifications may make clinicians less responsive to important alerts. It can also increase workload, delay treatment and encourage routine dismissal of warnings.
Healthcare organizations should therefore determine which alerts require active resolution before a medication order proceeds and which can be presented as advisory information.
The response should reflect the likelihood and potential severity of harm, the reliability of the information supporting the warning and the clinical circumstances.
For a serious, well-supported concern, the appropriate response may involve correcting inaccurate information, reconsidering the medicine, adjusting the dose or obtaining additional professional advice.
In some circumstances, a clinically justified decision may allow treatment to proceed despite a warning, with appropriate safeguards and documentation.
The process also needs to account for urgency.
An alert that appropriately interrupts routine outpatient prescribing may require a different handling pathway during an emergency, when delaying necessary treatment could cause greater harm.
The timing of the warning matters as much as its content.
An alert displayed before an order is signed gives the prescriber an opportunity to reassess the treatment. A problem discovered after dispensing or administration may require patient contact, clinical reassessment, treatment modification or additional monitoring.
The objective should not be to maximize the number of alerts generated or acknowledged. It should be to ensure that clinically important medication risks receive appropriate attention while the healthcare team can still prevent or reduce harm.
Prescriber responsibility and the role of clinical pharmacists
AI-assisted prescribing must operate within the professional responsibilities established by Vietnamese healthcare law.
The Law on Medical Examination and Treatment No. 15/2023/QH15, as amended and reflected in Consolidated Document No. 26/VBHN-VPQH dated February 26, 2026, provides the general legal framework for professional practice and patient care [6].
For outpatient prescribing of chemical medicines and biological products, Circular No. 26/2025/TT-BYT, effective July 1, 2025, establishes relevant requirements concerning prescribing authority, prescription content and prescribing principles [7].
An AI-generated medication recommendation does not independently establish prescribing authority or replace the professional decision required under applicable law.
The appropriately authorized prescriber must assess whether the proposed treatment is suitable for the patient's clinical circumstances and consistent with relevant professional requirements.
This does not mean clinicians must manually repeat every calculation performed by a properly validated tool. It means the prescribing process must support meaningful assessment of important information, particularly when a recommendation involves uncertainty or significant potential harm.
Clinical pharmacists can provide valuable additional expertise.
Decree No. 131/2020/NĐ-CP regulates the organization and activities of clinical pharmacy in healthcare facilities [8].
The Decree addresses activities including medication-use review, identification of medication-related problems, professional consultation and collaboration with clinical departments.
Article 15 is particularly relevant when clinical pharmacy personnel identify concerns about medication safety, effectiveness or appropriateness.
Under the circumstances covered by that provision, the clinical pharmacy professional provides advice to the prescriber and records the consultation in the prescribed patient medication-use analysis form. The professional is responsible for the advice provided.
Where the advice is not accepted, the decision and responsibility concerning prescribing and medication use remain with the prescriber, as specified in the Decree.
These provisions support defined professional collaboration when medication-related problems arise. They should not be interpreted as establishing a universal requirement for a pharmacist to approve every recommendation generated by AI.
The appropriate level of pharmacist involvement depends on clinical risk, applicable professional obligations, healthcare services and the organization's operating arrangements.
In a hospital with established clinical pharmacy services, pharmacist review may be prioritized for complex renal dose adjustments, high-alert medicines, clinically significant interactions, complicated antimicrobial regimens or patients receiving multiple medications with substantial safety concerns.
Clinical pharmacists may also contribute to evaluating recurring alerts, maintaining medication reference information and identifying prescribing patterns that require quality improvement.
For relatively straightforward, lower-risk prescribing, assessment by an appropriately authorized prescriber and existing medication safety controls may be sufficient.
Smaller healthcare facilities may organize medication review differently, depending on their services, available professional resources and applicable legal requirements.
The value of pharmacist involvement comes from identifying and resolving meaningful medication problems, not from requiring an additional signature for every computerized recommendation.
A practical process for verifying AI-assisted medication recommendations
Healthcare organizations can establish a structured but proportionate process for moving from an AI-generated suggestion to an authorized medication order.
The process begins with confirming the patient, the clinical indication and the medication decision involved.
The responsible professional needs to understand whether the AI system successfully analyzed the relevant information and whether the case falls within its intended use.
The next consideration is the reliability of the underlying patient data.
Important information such as body weight, organ function, allergies, medication history and relevant laboratory results should be verified according to the treatment being considered.
Once the necessary information is available, the proposed medicine can be assessed for clinical suitability.
This includes reviewing the indication, dose, route, frequency, duration, formulation and relevant medication-related risks.
For medicines requiring special dosing considerations, calculations and reference information should be checked using an appropriate method.
If a clinically important concern remains unresolved, the prescriber should determine whether additional information or professional consultation is needed.
Depending on the circumstances, this could involve advice from a clinical pharmacist, specialist consultation, additional laboratory information or another appropriate form of review.
Not every AI warning requires a second professional to assess the case. The need for escalation should reflect the clinical importance of the concern and whether the initial prescriber can adequately resolve it within the applicable professional framework.
The authorized prescriber then makes the treatment decision, taking account of the available evidence and any relevant professional advice.
The AI recommendation may be accepted, modified or rejected.
The purpose is not to achieve agreement with AI. It is to arrive at a clinically appropriate prescription supported by sufficiently reliable information.
Where a significant medication safety concern has influenced treatment, the relevant issue and action should be documented according to applicable requirements and organizational procedures.
This process is proposed as a practical organizational approach. It is not a standardized Ministry of Health AI prescribing protocol or an independently validated medication safety assessment instrument.
Medication review should fit the actual clinical workflow
Medication review operates differently across healthcare settings.
In inpatient care, medication orders may pass through prescribing, pharmacy review, preparation, dispensing, administration and subsequent clinical monitoring.
A problem identified during one stage may need to be communicated to professionals responsible for another.
For example, an AI system may identify a possible dose discrepancy during prescribing. A clinical pharmacist may subsequently identify an interaction, while a nurse may recognize a discrepancy involving preparation or administration instructions.
The organization should establish how important concerns are communicated and resolved without relying on assumptions that another professional has already completed the necessary checks.
Outpatient prescribing presents different challenges.
A patient may receive a prescription during a short consultation and obtain the medicine from a community pharmacy that has limited access to the original clinical information.
The dispensing pharmacist may recognize a potential problem but may not know which laboratory results or other patient details were considered when the prescription was issued.
Where necessary, the workflow should support timely clarification between the pharmacy and the prescriber.
Medication safety also continues after treatment begins.
A new laboratory result, newly identified allergy, unexpected adverse effect or change in clinical condition may require reassessment of an earlier prescription.
For selected medicines, follow-up monitoring may be as important as confirming the initial dose.
AI-assisted prescribing should therefore fit within the wider medication-use process rather than operate as an isolated calculation or one-time safety check performed before the order is signed.
Record clinically meaningful medication safety concerns
Documentation is particularly important when AI-assisted prescribing identifies a problem that could materially affect patient safety.
A relevant case may involve an inappropriate dose suggested because of outdated laboratory information, a significant interaction discovered before dispensing or a medication order changed following clinical pharmacy consultation.
The clinical record should contain enough information to explain the concern, the action taken and any necessary follow-up, consistent with applicable professional and documentation requirements.
For selected events, information about the AI system itself may also support investigation.
This may include the system version, the time the recommendation was generated, the relevant information available to the system and whether the output was displayed before the prescription was finalized.
These details can help determine whether a problem arose from incomplete clinical data, an incorrect calculation, outdated medication reference information, a software integration failure or an inappropriate response to a warning.
The corrective action will depend on the underlying cause.
However, not every routine alert needs to become a formal patient-safety incident report.
Healthcare organizations should distinguish ordinary decision support activity from clinically significant discrepancies, near misses and events involving actual patient harm.
A potential error corrected before reaching the patient is not equivalent to an adverse event that caused injury, although both may provide useful information for safety improvement.
Documentation should support clinical continuity, professional accountability and learning without unnecessarily duplicating sensitive patient information across clinical records, technical logs and quality-improvement databases.
Monitoring should measure medication safety, not simply AI acceptance
After deployment, healthcare organizations may be tempted to measure performance primarily by counting how often clinicians accept or reject AI recommendations.
These figures can help explain how the system is being used, but they do not establish whether medication safety has improved.
A high acceptance rate may indicate that the system produces relevant recommendations. It may also reflect excessive trust in computerized advice.
Similarly, frequent rejection of AI recommendations may indicate limited clinical usefulness, but some rejections may represent appropriate professional judgment.
More meaningful evaluation should consider whether clinically important medication problems are recognized and appropriately managed.
Healthcare organizations may monitor confirmed dose discrepancies, significant interaction alerts, medication orders corrected before dispensing, selected near misses and medication-related adverse events.
For complex or high-consequence prescribing decisions, the time required to resolve important warnings may also be informative.
However, reviewing only cases in which AI generated an alert can provide an incomplete picture of safety.
A system may fail to identify a medication problem that is subsequently detected by a pharmacist or another healthcare professional. Other errors may remain unrecognized until the patient receives treatment or experiences an adverse event.
Where practical, quality assurance should therefore include selected prescriptions that did not generate AI warnings, particularly those involving complex dosing, high-alert medicines or patients with substantial medication-related risks.
The evaluation should also consider whether system performance differs across clinical settings or patient populations.
An application evaluated in general adult outpatient care should not automatically be assumed to perform equally well for children, frail older adults, patients with unstable organ function or people receiving complex treatment regimens.
Differences in clinical outcomes or alert performance require careful investigation, taking account of patient characteristics, clinical case mix, data quality and the system's intended use.
Monitoring should help healthcare organizations determine whether AI contributes to safer medication decisions and whether new risks emerge after implementation.
Local validation matters before wider clinical deployment
An AI-assisted prescribing application should be evaluated using medication scenarios representative of the healthcare organization where it will be used.
A hospital may need to assess commonly prescribed medicines, including high-alert medications, complex antimicrobial regimens and drugs requiring individualized dose adjustment.
An outpatient or specialty clinic may focus on the medicines, patient groups and treatment decisions most relevant to its services.
Local evaluation should consider the reliability of medication reference information, available formulations, prescribing guidelines and the quality of integration with existing information systems.
This is particularly important in Viet Nam, where an application developed for another healthcare market may rely on different medication products, formulations, clinical references or prescribing conventions.
Testing should include more than complete, straightforward patient records.
It should examine scenarios involving missing laboratory information, outdated weight measurements, incomplete medication histories, clinically significant interactions and dosing calculations requiring special precautions.
The organization should also evaluate how the system behaves when relevant information cannot be retrieved or when it cannot generate a reliable recommendation.
A system that performs well on uncomplicated cases may still fail when important clinical information is unavailable.
For generative AI, evaluation should include the possibility of unsupported medication recommendations, incorrect calculations, fabricated information and misleading expressions of certainty.
Where feasible, organizations may begin with a noninterventional evaluation in which AI recommendations are compared with existing clinical decisions without directly influencing patient care.
This can help identify technical and clinical problems before the system becomes part of routine prescribing.
However, retrospective testing or successful silent-mode evaluation does not establish long-term clinical effectiveness.
Clinician behavior, workload and system performance may change once AI recommendations become visible and begin influencing actual treatment decisions.
Healthcare organizations should therefore plan for continuing evaluation, including how software updates, changes to medication knowledge bases and modifications to electronic prescribing interfaces will be assessed.
Significant or recurring medication safety problems should trigger appropriate investigation and corrective action. Depending on the circumstances, this may involve modifying the workflow, restricting particular functions or suspending use until the concern has been adequately addressed.
Vietnamese legal requirements need to be assessed separately
Several areas of Vietnamese law are relevant to AI-assisted medication prescribing, but they regulate different aspects of healthcare practice and technology.
The Law on Medical Examination and Treatment No. 15/2023/QH15, as amended, and the Law on Pharmacy No. 105/2016/QH13, as amended by Law No. 44/2024/QH15, provide important elements of the professional and pharmaceutical legal framework [6, 10].
Circular No. 26/2025/TT-BYT establishes specific requirements for outpatient prescribing of chemical medicines and biological products [7]. Its provisions should not automatically be extended to all inpatient medication orders or prescribing activities outside its scope.
Where prescribing involves traditional medicines, herbal medicines or their combination with chemical medicines, the applicable specialized requirements should also be considered, including Circular No. 55/2025/TT-BYT, effective March 1, 2026 [9].
Decree No. 131/2020/NĐ-CP establishes the organization and activities of clinical pharmacy services, with responsibilities determined by its relevant provisions and the healthcare setting [8].
These professional and pharmaceutical requirements need to be distinguished from the regulatory classification of the AI technology itself.
Viet Nam's Law on Artificial Intelligence No. 134/2025/QH15 took effect on March 1, 2026 [11]. Decree No. 142/2026/NĐ-CP, effective May 1, 2026, provides further provisions on implementation, including AI risk classification and management [12].
Decision No. 33/2026/QĐ-TTg, effective August 15, 2026, establishes the list of high-risk AI systems [13].
As of this article's review date, the healthcare-specific entries in that list concern defined AI-assisted surgical and robotic surgical systems meeting particular conditions.
Ordinary AI-assisted medication prescribing is not separately identified in those healthcare entries.
This does not mean every AI prescribing application must be classified as low-risk.
The legal classification of a particular system needs to be assessed according to its intended purpose, actual functions, applicable risk criteria and relevant legal provisions.
An application that influences medication selection or dosage may present substantial clinical safety concerns even when it does not fall within a specifically listed high-risk AI category.
Clinical risk assessment and legal AI risk classification are related but distinct considerations.
Medical device regulation may also apply to certain software products.
Decree No. 98/2021/NĐ-CP on medical device management, as amended, provides a relevant regulatory framework [14]. Circular No. 24/2026/TT-BYT, effective July 1, 2026, further addresses risk determination and management measures for products classified as medical devices [15].
Whether an AI prescribing application falls within medical device regulation depends on its intended medical purpose, actual functions and applicable legal requirements.
Healthcare organizations should not assume that every medication-related AI application is regulated medical device software or that software described as clinical decision support is automatically excluded from that framework.
Medical device classification and AI risk classification are separate regulatory assessments. Compliance with one framework should not automatically be regarded as compliance with the other.
Patient information also raises important legal obligations.
The Law on Personal Data Protection No. 91/2025/QH15 and Decree No. 356/2025/NĐ-CP took effect on January 1, 2026 [16, 17].
AI-assisted prescribing may involve sensitive personal data, including diagnoses, laboratory results, medication histories and other health information.
Healthcare organizations should assess the lawful basis and purposes of processing, access permissions, information security, retention, relevant vendor arrangements and any applicable requirements concerning data sharing or transfers.
Particular care is needed where patient information is processed by external AI services or cloud-based systems.
Internal medication safety reporting should also be distinguished from statutory notification requirements that may apply to particular incidents, regulated technologies or professional activities.
No single AI-specific review or approval procedure should be assumed to satisfy all these different legal obligations.
The applicable requirements depend on the healthcare organization, clinical activity, system functions, regulatory status and circumstances of implementation.
Making AI-assisted prescribing practical for healthcare organizations in Viet Nam
For healthcare organizations considering AI-assisted prescribing, the most useful starting point may be a clearly defined medication safety problem rather than an ambitious attempt to automate the entire prescribing process.
A hospital might begin by evaluating support for renal dose adjustments, selected high-alert medicines or clinically significant drug interactions.
A specialty clinic may choose medicines commonly used in its services, particularly where patient characteristics make dosing or treatment selection more complex.
The initial use case should be specific enough to evaluate whether the system identifies meaningful risks and whether healthcare professionals can respond to its recommendations within their normal workflow.
Prescribers and clinical pharmacy personnel should contribute to selecting appropriate medication references, identifying essential patient information and determining when additional professional assessment is justified.
Information technology and patient-safety teams can help establish how recommendations appear in the prescribing system, what information is available for verification and how clinically significant problems will be investigated.
The workflow should reflect actual clinical capacity.
A large hospital with established clinical pharmacy services may support arrangements that would be difficult for a smaller outpatient facility to maintain.
At the same time, a simpler workflow should not eliminate safeguards necessary for the clinical risks involved.
The appropriate approach is to make medication safety checks proportionate to the potential consequences of an incorrect recommendation and practical enough to function reliably during routine care.
Organizations should also examine whether AI reduces meaningful medication risks without creating excessive alerts, unnecessary treatment delays or additional work that does not contribute to patient safety.
A prescribing system that produces fewer but more relevant warnings may be more useful than one that generates large numbers of notifications clinicians cannot realistically investigate.
Implementation should therefore be evaluated through actual clinical performance rather than the volume of recommendations produced or accepted.
The purpose is to create a medication-use process in which reliable technology and professional expertise reinforce each other.
From an AI-generated dose to a safer medication decision
AI can help healthcare professionals recognize medication-related risks, identify potentially inappropriate prescriptions and make better use of available clinical information.
However, an apparently accurate medication recommendation is only one part of safe prescribing.
The patient's information must be sufficiently reliable and relevant to the decision. Dose calculations need to account for the medicine, indication, formulation and clinical circumstances. Important allergies, interactions and contraindications must be considered when appropriate.
The authorized prescriber remains responsible for making the prescribing decision within the applicable professional framework. Clinical pharmacists can provide additional medication expertise according to clinical risk, professional responsibilities and the organization's operating arrangements.
Meaningful warnings should be addressed at the point where they can influence treatment, and significant medication safety concerns should contribute to continuing improvement.
For healthcare organizations in Viet Nam, responsible AI-assisted prescribing does not mean requiring another professional to approve every computerized recommendation or treating every warning as equally urgent.
It means establishing a process in which medication recommendations can be checked against reliable patient information, important risks receive appropriate attention and professional decisions can be made without unnecessary delay.
The value of AI-assisted prescribing should ultimately be judged by the safety and appropriateness of medication decisions, not by how quickly a system can suggest a medicine or calculate a dose.
Professional Education Notice
This article is intended for professional education and knowledge exchange. It does not constitute individualized medical, pharmaceutical, prescribing, legal, regulatory or patient-safety advice.
The medication recommendation verification process discussed here is a proposed organizational approach. It is not a prescribed Vietnamese Ministry of Health procedure, a mandatory international protocol or an independently validated medication safety assessment framework.
Recommendations concerning patient data verification, dose checking, pharmacist consultation, alert management, documentation and monitoring should not be interpreted as identical statutory obligations for every healthcare organization or AI application.
The responsibilities of prescribers, clinical pharmacy professionals, healthcare facilities and technology providers depend on the applicable laws, professional scope of practice, healthcare services, system functionality and circumstances of use.
Evidence from conventional clinical decision support systems should not automatically be generalized to machine-learning or generative AI applications. Findings from individual studies, institutions or patient populations should not be assumed to establish effectiveness across all healthcare settings in Viet Nam.
AI-related regulatory classification, medical device status, personal data protection obligations and incident reporting requirements must be assessed separately under the relevant legal frameworks.
References
World Health Organization. Medication Without Harm: Policy Brief. Published on the WHO website March 7, 2024. ISBN: 978-92-4-006276-4.
Syrowatka A, Motala A, Lawson E, Shekelle PG. Clinical Decision Support Systems' Effect in Reducing Medication Errors and Adverse Drug Events: An Expanded Review from Making Healthcare Safer IV. BMJ Digital Health & AI. 2025;1. DOI: 10.1136/bmjdhai-2025-000083.
Nguyen Thi Thu Phuong, Dao Thuy Duong, Tran Van Anh, et al. Effectiveness of an Integrated Clinical Decision Support System and Clinical Pharmacist-Led Interventions in Domperidone Prescribing: A 10-Year Interrupted Time Series Study. Hospital Pharmacy. Published online April 12, 2026. DOI: 10.1177/00185787261433442.
Al-Taie A, Abdullah AS, Arueyingho O. Integration of Artificial Intelligence Applications in Clinical Pharmacy Services: A Scoping Review. Future Healthcare Journal. 2026;13(3):100547. DOI: 10.1016/j.fhj.2026.100547.
Surdilla D, Michaels B, Michael C, et al. Impact of AI Assistance on Pharmacist and Pharmacy Trainee Accuracy and Automation Bias During Medication Order Verification: A Randomized Simulation Trial. American Journal of Health-System Pharmacy. Published online September 25, 2026. DOI: 10.1093/ajhp/zxag259.
National Assembly of Viet Nam. Law No. 15/2023/QH15 on Medical Examination and Treatment. Adopted January 9, 2023; effective January 1, 2024, as amended. Consolidated Document No. 26/VBHN-VPQH, issued February 26, 2026.
Ministry of Health of Viet Nam. Circular No. 26/2025/TT-BYT on Prescriptions and Prescribing of Chemical Medicines and Biological Products in Outpatient Treatment at Healthcare Facilities. Issued June 30, 2025; effective July 1, 2025.
Government of Viet Nam. Decree No. 131/2020/NĐ-CP on the Organization and Activities of Clinical Pharmacy in Healthcare Facilities. Issued November 2, 2020; effective January 1, 2021.
Ministry of Health of Viet Nam. Circular No. 55/2025/TT-BYT on Prescribing Traditional Medicines, Herbal Medicines and Their Combination with Chemical Medicines in Healthcare Facilities. Issued December 31, 2025; effective March 1, 2026.
National Assembly of Viet Nam. Law on Pharmacy No. 105/2016/QH13, as amended by Law No. 44/2024/QH15 on Amendments and Supplements to the Law on Pharmacy, effective July 1, 2025.
National Assembly of Viet Nam. Law No. 134/2025/QH15 on Artificial Intelligence. Adopted December 10, 2025; effective March 1, 2026.
Government of Viet Nam. Decree No. 142/2026/NĐ-CP Detailing Certain Provisions and Measures for Implementing the Law on Artificial Intelligence. Issued April 30, 2026; effective May 1, 2026.
Prime Minister of Viet Nam. Decision No. 33/2026/QĐ-TTg Promulgating the List of High-Risk Artificial Intelligence Systems. Issued June 30, 2026; effective August 15, 2026.
Government of Viet Nam. Decree No. 98/2021/NĐ-CP on Medical Device Management, as amended, including by Decree No. 07/2023/NĐ-CP and Decree No. 04/2025/NĐ-CP.
Ministry of Health of Viet Nam. Circular No. 24/2026/TT-BYT on Risk Determination and Management Measures for Products Classified as Medical Devices and Amendments to Circular No. 05/2022/TT-BYT. Issued June 30, 2026; effective July 1, 2026.
National Assembly of Viet Nam. Law No. 91/2025/QH15 on Personal Data Protection. Adopted June 26, 2025; effective January 1, 2026.
Government of Viet Nam. Decree No. 356/2025/NĐ-CP Detailing Certain Provisions and Measures for Implementing the Law on Personal Data Protection. Issued December 31, 2025; effective January 1, 2026.