Building an AI-Ready Healthcare Workforce in Viet Nam: Competency, Training and Safe Clinical AI Use
IMPLEMENTATION INSIGHTS
Building an AI-Ready Healthcare Workforce in Viet Nam: Competency, Training and Safe Clinical AI Use
IMPLEMENTATION INSIGHTS
Last reviewed: 1 October 2026
Healthcare organizations can spend considerable time evaluating clinical AI, reviewing evidence, negotiating contracts and designing governance structures. Yet even a technically strong system can become unsafe if the people expected to use, supervise or support it do not understand what it is designed to do, where its limitations lie, when its output should be questioned and what to do when something goes wrong. Workforce readiness is therefore a patient-safety issue, not simply a training issue.
An AI-ready healthcare workforce does not mean that every doctor, nurse, technician or administrator needs to understand model architecture, write code or become a data scientist. It means that people should have the competence appropriate to the decisions they make and the risks they control. A clinician using AI to support diagnosis needs different knowledge from an executive approving an investment. A quality manager investigating an AI-related incident needs different capabilities from an IT engineer maintaining an interface, while a procurement team evaluating a vendor needs different competence from a nurse responding to an AI-generated clinical alert.
The practical question is therefore not simply “Have we trained our staff on AI?” It is “Do the people who select, approve, configure, use, monitor and govern this AI have the competence needed to perform their roles safely?”
That distinction is increasingly important in Viet Nam. The Law on Artificial Intelligence No. 134/2025/QH15 has been effective since 1 March 2026 and establishes human-centered principles for AI. The Law states that AI should serve people rather than replace human authority and responsibility and requires human control and the ability to intervene in AI decisions and actions. Healthcare is also identified as a sector requiring stricter risk management appropriate to its sensitivity, including attention to patient safety, reliability under real-world conditions and protection of health data.[1]
For AI systems classified as high-risk, the obligations become more specific. The deployer must operate and supervise the system within its intended purpose and risk classification, protect data and preserve the ability for human intervention. Users must follow operating procedures, technical instructions and safety measures and notify the deployer promptly when incidents occur.[1] These requirements do not create a universal clinical-AI training curriculum, but they reinforce a basic principle: human oversight only works when the human is capable of exercising it.
AI literacy is necessary, but it is not enough
Many healthcare AI programs begin with general awareness training: what artificial intelligence is, examples of machine learning and generative AI, potential benefits and common ethical concerns. That can be useful, particularly for organizations at an early stage of adoption, but awareness alone does not establish safe clinical competence.
A clinician may understand AI broadly and still not know the intended use of a particular system. Someone may recognize the concept of algorithmic bias but not know how bias could appear in the population they treat. A staff member may complete a course on generative AI but still enter sensitive patient information into an unapproved public tool. A doctor may know that AI can be wrong yet become over-reliant on a confident-looking output during a busy shift.
Competence therefore needs to go beyond knowing what AI is. It should include the ability to use a particular system appropriately, recognize unsafe or uncertain situations, respond correctly and understand the limits of one’s own authority.
WHO’s 2026 landscape analysis of digital-health competency frameworks supports this broader approach. Across existing frameworks, recurring competency clusters include patient care, data, informatics, communication, technical proficiency, digital professionalism and administration. WHO also emphasizes the need to adapt competency standards to professional, educational and organizational contexts rather than treating digital competence as one generic set of skills.[5]
For clinical AI, the same principle is useful: competency should follow role, use case and risk.
Start by identifying who actually interacts with the AI lifecycle
Healthcare organizations should not limit workforce planning to the clinicians who see the final AI output. Clinical AI usually passes through a much wider organizational system. Senior leaders may approve strategy, investment and risk tolerance. Clinical leaders define intended use and determine whether outputs make clinical sense. Quality and patient-safety teams may establish monitoring, investigate incidents and manage corrective actions. IT and informatics teams integrate systems, configure access and maintain data flows. Privacy and data teams manage information governance, while procurement, legal and compliance teams evaluate vendors, contractual responsibilities and change provisions. Department managers oversee day-to-day adoption, and frontline clinicians determine how AI outputs affect care.
The same AI system can therefore require several different competency profiles. The goal is not for every person to understand everything. The goal is to avoid a situation in which an important responsibility exists but nobody has the competence to perform it.
Clinical users need to understand the intended use and its boundaries
For frontline users, one of the most important competencies is also one of the simplest: knowing what the AI is actually intended to do. Users should understand the target population, clinical setting, required inputs, type of output and role of the system in the care pathway. They should know whether the AI is intended for screening, detection, prioritization, prediction, documentation or another purpose, whether its output is advisory and whether any action can occur with a greater degree of automation.
Understanding what the AI is not designed to do is equally important. An AI system evaluated for adult chest radiographs should not quietly become a decision-support tool for children without appropriate evidence and authorization. A colonoscopy detection system should not be assumed to classify pathology it was never designed to assess. A generative documentation tool should not gradually become an autonomous source of treatment advice simply because users discover that it can produce recommendations.
Safe use therefore begins with clear boundaries. Users should be able to answer four basic questions: Who is this system for? What decision does it support? What should I do with the result? When should I not rely on it?
Users need enough performance literacy to interpret outputs responsibly
Clinicians do not need to become statisticians, but they need enough understanding of performance to avoid giving an AI output more authority than the evidence supports. Depending on the system, this may involve understanding the practical meaning of sensitivity, specificity, false positives and false negatives. For a risk-prediction model, users may need to understand that a probability is not a diagnosis and that calibration can differ across populations. For imaging AI, users may need to know whether the system detects, localizes, prioritizes or classifies findings and whether performance may change under poor image quality or particular acquisition conditions.
A single “accuracy” number should not become the clinician’s mental model of a technology. A confident-looking interface does not necessarily mean a highly certain result. A predicted probability of 85% is not certainty. An AI-generated recommendation can sound plausible and still be incorrect, while the absence of an alert does not automatically mean that disease is absent.
Performance literacy is therefore not primarily about memorizing statistics. It is about interpreting outputs with the right degree of confidence and skepticism.
Human oversight has to be taught as an active skill
Healthcare organizations often say that “a clinician remains in the loop.” By itself, that statement provides little assurance. Meaningful oversight requires the person to know when independent review is expected, when an AI output should be challenged, when escalation is necessary and whether the user has genuine authority to reject or override a recommendation.
Viet Nam’s AI Law places human control and the ability to intervene within its governance framework.[1] Decree No. 142/2026/NĐ-CP goes further in its risk-classification rules: a system that otherwise meets high-risk criteria is not proposed for inclusion on the high-risk list when, among other specified circumstances, technical and operational arrangements ensure substantive human oversight such that an authorized person can independently review, intervene in, reject or change the system’s decision before it takes effect. The Decree also contains other exceptions and safeguards that need to be considered in context.[2]
Training should therefore not merely tell clinicians that they “remain responsible.” It should show how that responsibility works in the actual workflow. What should a radiologist do when AI highlights a finding that appears implausible? What should an endoscopist do when a detection system repeatedly flags normal mucosa? When should a nurse escalate an AI-generated deterioration alert? What happens when the AI output conflicts with another clinical test? Who has authority to restrict or temporarily stop use if recurrent unsafe behavior is observed?
Human oversight becomes meaningful only when those questions have practical answers.
Safe use requires calibrated reliance, not automatic trust or rejection
One of the most important risks in clinical AI is not simply that staff may resist the technology. It is that they may trust it too much. Automation bias can arise when users give disproportionate weight to an automated output, particularly when the system has performed well previously, workload is high or the recommendation appears authoritative. Over time, clinicians may begin to review information less independently because they expect the AI to identify important problems.
The opposite can also occur. Frequent false positives, poor integration or repeated low-value alerts can cause users to ignore the system altogether. Neither automatic acceptance nor habitual dismissal represents effective human oversight.
Training should therefore expose users to situations in which AI and human judgment disagree, including cases where the AI is correct and cases where it is wrong. Staff should learn to preserve independent reasoning while using AI where it adds value. The aim is calibrated reliance: trusting the system neither more nor less than its evidence, intended use and observed performance justify.
Different roles need different competencies
A practical workforce model should be organized around responsibility rather than job title alone. Frontline clinical users need competence in intended use, limitations, interpretation, uncertainty, escalation, incident reporting, appropriate data handling and patient communication where relevant. Clinical and departmental leaders need those capabilities plus an understanding of workflow fit, safe-use boundaries, appropriate user groups and decisions about continuation, restriction or suspension.
Quality and patient-safety teams need the ability to assess AI-related risks, define monitoring indicators, investigate incidents, distinguish model failure from workflow, data, integration or human-factor problems, and manage corrective actions. IT and informatics teams need deeper competence in interfaces, access control, data integrity, logging, version management, availability, cybersecurity and technical incident response, while still understanding enough clinical context to recognize why apparently minor technical changes may affect patient safety.
Privacy and information-governance teams need to understand the data the AI requires, processing purposes, access arrangements, vendor involvement, retention, cloud services and cross-border processing where relevant. Procurement, legal and compliance teams need enough AI literacy to examine evidence, intended use, regulatory status, performance commitments, data processing, cybersecurity, software updates, audit rights, incident support and exit arrangements. Senior leadership and the AI governance function need to understand organizational risk, accountability, escalation, resource requirements and when a system should be approved, restricted, monitored more closely or discontinued.
This does not require a separate training course for every group. Many competencies overlap. But one generic “AI awareness” session is unlikely to prepare all of these roles adequately.
Vendor training is useful, but it should not define organizational competence
Vendors often provide product training before go-live, and that can be valuable because they understand the system’s features, technical requirements, intended use and known limitations. But product training is only one part of workforce preparation.
Vendor training can explain the product. The healthcare organization still needs to determine how that information fits its own clinical governance, workflow and patient-safety requirements. Local escalation pathways, incident reporting, downtime procedures, privacy requirements, interaction with other systems, documentation expectations and organizational criteria for overriding, restricting or suspending use may not be covered adequately by standard vendor training.
A practical program therefore combines product-specific training with organization-specific preparation.
Training should use the real workflow, not only presentation slides
Clinical AI is a socio-technical intervention. Staff therefore learn most effectively when training reflects the environment in which they will actually use the system. Case-based learning can include true positives, false positives, false negatives, uncertain outputs and difficult edge cases. Simulation can show what happens when data are missing, an interface fails, a result arrives late or AI advice conflicts with clinical judgment.
Training should also cover the operational details that often determine whether implementation succeeds. Where does the output appear? Who is expected to review it? Is review mandatory? How is an override documented? Who receives an incident report? What happens during downtime? How does the user know which software or model version is operating?
These details may appear less impressive than the AI itself, but they are often where patient safety is won or lost.
Completion of training is not the same as competence
A signed attendance sheet shows that someone attended a session. It does not show that the person can use the technology safely.
Where clinical risk is meaningful, organizations should consider proportionate competency verification before independent use. This may involve short scenario-based assessments, supervised cases, demonstration of correct interpretation, recognition of limitations, appropriate escalation or confirmation that the user knows how to respond to common failure situations. Higher-risk or more complex systems may justify a more structured approach.
The aim is not to create an examination for every AI tool. It is to provide reasonable assurance that people who will rely on a system can perform the tasks expected of them. This is more meaningful than reporting that “100% of staff have completed training.”
Where appropriate, access can also follow competence. Not every employee who can technically open an application necessarily needs authority to use it in clinical decision-making. Organizations may begin with a limited group of trained users, observe early use and expand access gradually. Competency-based access is an organizational governance control; it should not be interpreted as a statement that Vietnamese law requires individual certification for every clinical AI user.
Generative AI creates additional workforce risks
General-purpose generative AI deserves particular attention because its possible uses extend far beyond a narrow clinical function. The same tool may be used to summarize records, draft documentation, translate patient materials, create educational content or answer clinical questions. That flexibility can make boundaries less visible than they are with a purpose-built clinical device or application.
Users should understand that fluent language is not evidence of factual reliability. Clinically important statements should be verified, unsupported conclusions or invented references should be recognized as possible failure modes, and staff should understand organizational rules on what patient information may or may not be entered into approved and unapproved systems.
Healthcare organizations should also address shadow AI, the use of public or unapproved AI tools outside established processes because they are convenient. Simply prohibiting all AI may be ineffective if staff are already using it. A more practical approach is to define approved tools, permitted and prohibited uses, data restrictions and verification expectations. For generative AI, competence includes knowing not only how to use the tool, but when not to use it.
Data protection and cybersecurity belong in workforce competence
AI safety and information security cannot be separated completely. Viet Nam’s Law on Personal Data Protection No. 91/2025/QH15 and Decree No. 356/2025/NĐ-CP have been effective since 1 January 2026.[9,10] Health-status data are classified as sensitive personal data under the current framework, making appropriate handling particularly important.
Staff interacting with AI should therefore understand organizational rules on what information may be entered into a system, who may access outputs, how information may be copied or exported, whether external vendors or cloud services are involved, and what to do if information is disclosed incorrectly. Technical teams require deeper cybersecurity capability, but frontline behavior also matters: shared accounts, weak authentication practices, personal AI tools and informal workarounds can undermine strong technical controls.
Privacy and cybersecurity education is most useful when linked directly to the AI workflow rather than treated only as a generic annual compliance exercise.
Patient communication may also require competence
Some AI systems will raise questions from patients: Was AI involved in my care? Did the computer make the decision? Can the clinician disagree with it? What happens if it is wrong?
Not every AI use case creates the same disclosure or consent requirements, and transparency, clinical communication and informed consent should not be treated as interchangeable legal concepts. But staff who interact with patients should know what their organization expects them to communicate and should be able to explain the role of AI without exaggerating either its authority or its limitations.
Often the explanation can remain simple: what the system does, how it supports the care process, the continuing role of the clinician and what happens if clinical judgment differs from the AI output. Staff should not have to invent those explanations individually at the bedside.
Competence should change when the system changes
AI competency should not be treated as permanent simply because someone completed training before go-live. A material model update may change outputs or performance. A redesigned interface can alter how clinicians interpret alerts. A new patient population, device, site or workflow may introduce risks that were not covered originally. Monitoring may reveal a failure mode staff need to recognize, while an incident may show that an escalation pathway was misunderstood.
Organizations should therefore define triggers for refresher training or competency reassessment. These may include material model or software changes, expansion to a new use case or clinical area, significant workflow changes, recurrent user errors, new safety information, serious incidents or evidence of inappropriate reliance or workarounds. Not every minor update warrants full retraining; the response should be proportionate to what changed and what staff need to do differently.
Monitoring should include human performance, not only model performance
After go-live, organizations often monitor technical availability and AI performance but pay less attention to how people use the system. Human-performance indicators can reveal problems that model metrics alone will not.
Examples include how often clinicians override AI recommendations, whether important alerts are ignored, how frequently users escalate concerns, whether workarounds develop, whether documentation quality changes and whether certain groups of staff consistently struggle to interpret outputs. A sudden reduction in overrides does not necessarily mean the AI improved; clinicians may simply have become more trusting. A high override rate does not automatically mean staff are resistant; the AI may be generating poor recommendations.
The unit of safety is often the human-AI system, not the algorithm alone. IMDRF’s 2025 Good Machine Learning Practice principles explicitly call for assessment of human–AI interactions in the intended use environment, including the performance of the human–AI team rather than the device in isolation.[8]
Training should also connect directly to incident reporting. An AI-related safety issue may begin with an unusual recommendation, repeated missed findings, inconsistent outputs after an update, incorrect patient information or an interface problem. Staff should have a simple route for reporting these observations without first having to prove that the AI caused harm. Investigation may require clinicians, quality and patient safety, IT, data teams, the vendor and leadership to work together, and lessons from those incidents should feed back into workforce preparation.
Viet Nam’s policy direction increasingly supports AI workforce capability
Viet Nam’s wider policy direction now places considerably more emphasis on AI capability. Decision No. 1528/QĐ-TTg, issued and effective on 11 August 2026, approved the National Program for Artificial Intelligence Human Resource Development to 2030, with Orientation to 2035. The Program adopts a three-tier approach covering broad AI literacy, applied AI capability across sectors and advanced specialist capability. It also seeks to make the ability to use and apply AI an essential competence for learners, workers and management personnel.[4]
Importantly, the Program also calls for development of a National AI Competency Framework, with the implementation roadmap targeting June 2027.[4] As of 1 October 2026, the framework reviewed for this article therefore does not yet establish a single national clinical-AI competency standard or a universal certification requirement specifically for healthcare workers using AI. Healthcare organizations still need to translate broader national capability goals into practical, role-based requirements for their own technologies, workflows and risk profiles.
The National Artificial Intelligence Ethics Framework, issued under Circular No. 05/2026/TT-BKHCN, also provides relevant direction. Its mandatory scope covers state agencies and organizations or individuals acting as developers, providers, deployers or users of AI systems for state management or public-service delivery; other Vietnamese and foreign organizations conducting AI activities in Viet Nam are encouraged to apply it.[3] The Framework emphasizes human judgment, verification of AI outputs, protection of sensitive information and development of skills for safe and responsible AI use.
This distinction matters for healthcare organizations. The Ethics Framework can provide a useful governance reference, but its legal application should not be described as identical for every public and private healthcare provider.
Viet Nam is beginning to build health-sector AI capability
Workforce development is also becoming visible within the health sector itself. On 14 January 2026, the National Health Information Center under the Ministry of Health held a health AI kickoff workshop and technical training program within its professional cooperation with HealthAI and officially launched V-RHAIN, described by the Ministry as Viet Nam’s network of pioneering members for responsible health AI solutions. The event included approximately 200 in-person participants and was connected online to thousands of health-sector participants nationwide.[11]
This is an important signal, but participation in a network, workshop or introductory training program should not be confused with competence to use a specific clinical AI system safely. General education can create awareness. Operational readiness requires clear roles, system-specific knowledge, practical supervision, local governance and evidence that people understand what they are expected to do.
Internationally, the challenge is similar. WHO’s 2026 digital-health competency analysis emphasizes contextualization of workforce competencies, while WHO/Europe’s 2026 assessment of AI readiness examines workforce preparedness as an important part of responsible deployment. WHO/Europe has also reported that only a minority of countries in its Region provide AI education before qualification or training once health professionals enter the workforce, and its September 2026 Knowledge Community report identifies AI literacy deficits among barriers to responsible adoption.[5–7]
A practical AI Workforce Readiness Gate
Before routine clinical use, a healthcare organization can apply a simple AI Workforce Readiness Gate:
Role clarity: Have the people responsible for selecting, approving, configuring, using, monitoring and governing the system been identified, with clear responsibilities and escalation paths?
Use-case competence: Do intended users understand the purpose, target population, inputs, outputs, limitations and circumstances in which the system should not be relied upon?
Clinical interpretation and human oversight: Can users interpret outputs appropriately, recognize implausible or uncertain results, exercise independent judgment and override or escalate when necessary?
Data and security competence: Do relevant staff understand permitted data use, access rules, privacy and cybersecurity requirements and the boundaries around external or generative AI tools?
Workflow competence: Can the system be used safely under real working conditions, including downtime, disagreements, unusual cases and handovers?
Incident and escalation readiness: Do users know what may constitute an AI-related safety issue, how to report it and when use should be restricted or suspended?
Competency assurance and maintenance: Has competence been reasonably verified where warranted by risk, and are there triggers for refresher training when the technology, workflow or risk profile changes?
This Workforce Readiness Gate is a practical organizational governance framework. It is not a statutory checklist under Vietnamese law and should be adapted to the technology, intended use, clinical context and level of risk.
What this means for healthcare organizations in Viet Nam
Hospitals and clinics do not need to wait for a perfect national clinical-AI curriculum before preparing their workforce. They can begin with the AI systems they are actually considering or already using. For each system, identify who makes decisions about it, who uses it, who monitors it and who responds when something goes wrong. Define the minimum competence required for those roles, combine product training with local workflow, patient-safety, privacy and escalation requirements, verify competence where the risk justifies it, and reassess capability when the system changes.
Organizations should resist two extremes. Clinical experience alone does not automatically create AI competence, but clinicians also do not need to become AI engineers before they can use a well-governed system safely. What they need is enough understanding to remain effective professionals when AI enters the decision-making environment.
For healthcare leaders, the practical question is therefore not whether the organization has delivered AI training. It is whether its workforce can use AI without surrendering clinical judgment, patient-safety responsibility or organizational accountability to the technology.
An AI-ready workforce is neither a workforce that automatically trusts AI nor one that automatically distrusts it. It is a workforce that knows when to use it, how to challenge it, when to stop relying on it and what to do when something does not look right.
That is the competence responsible clinical AI requires.
Professional education notice
This article is intended for professional education and knowledge exchange. It does not constitute individualized legal, regulatory, clinical, employment, educational, cybersecurity, personal-data or professional-licensing advice.
The competency domains, role-based approach and AI Workforce Readiness Gate described in this article are practical organizational recommendations. They are not a statutory competency framework, mandatory national training curriculum or professional certification standard under Vietnamese law.
Training and competency requirements should be proportionate to the technology, intended use, clinical risk, organizational environment and applicable legal or regulatory obligations. Where an AI system is subject to high-risk AI requirements, medical-device regulation, professional-practice requirements, personal-data rules or other specialized legislation, those obligations should be assessed separately.
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. 30 April 2026; effective 1 May 2026.
Ministry of Science and Technology of Viet Nam. Circular No. 05/2026/TT-BKHCN Promulgating the National Artificial Intelligence Ethics Framework. 10 March 2026; effective 10 March 2026.
Prime Minister of Viet Nam. Decision No. 1528/QĐ-TTg Approving the National Program for Artificial Intelligence Human Resource Development to 2030, with Orientation to 2035. 11 August 2026; effective 11 August 2026.
World Health Organization. Competency Frameworks and Standards for Digital Health: A Landscape Analysis. 14 June 2026. ISBN 978-92-4-011976-5.
World Health Organization Regional Office for Europe. Artificial Intelligence Is Reshaping Health Systems: State of Readiness Across the European Union. 20 April 2026. WHO/EURO:2026-12707-52481-81471.
World Health Organization Regional Office for Europe. Report of the Knowledge Community on Responsible Artificial Intelligence in Health. 1 September 2026. WHO/EURO:2026-13381-53155-83077.
International Medical Device Regulators Forum. Good Machine Learning Practice for Medical Device Development: Guiding Principles. IMDRF/AIML WG/N88 FINAL:2025. 29 January 2025.
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. 31 December 2025; effective 1 January 2026.
National Health Information Center, Ministry of Health of Viet Nam. Health AI kickoff workshop, technical training and launch of V-RHAIN. 14 January 2026.