Shadow AI in Healthcare Organizations in Viet Nam: How to Identify, Govern and Reduce Unapproved AI Use
IMPLEMENTATION INSIGHTS
Shadow AI in Healthcare Organizations in Viet Nam: How to Identify, Govern and Reduce Unapproved AI Use
IMPLEMENTATION INSIGHT
Last reviewed: 3 October 2026
AI does not always enter a healthcare organization through procurement, an innovation project or a formal clinical pilot. Sometimes it enters through a browser tab.
A clinician may use a public generative AI tool to summarize a document, improve the wording of a referral letter, translate patient information or think through a clinical question. An administrator may use an AI assistant to draft a report. A quality team may upload a policy for comparison. A staff member may install an AI transcription tool or browser extension. An existing software vendor may add an AI-enabled feature through a routine product update.
Some of these uses may appear harmless, and some may genuinely save time. The governance problem begins when the organization does not know which AI is being used, what information is being entered, how the outputs are influencing work, whether the technology has been reviewed for that purpose, or who is responsible when something goes wrong.
This is increasingly described as shadow AI: AI tools or AI-enabled functions being used for organizational work outside the organization’s approved governance, technology or oversight pathway. It is a practical governance term, not a legal category defined in Viet Nam’s Law on Artificial Intelligence. For healthcare organizations, the objective should therefore not simply be to ban unfamiliar AI or assume that every unapproved use presents the same risk. The more useful question is: How can an organization make hidden AI use visible, distinguish low-consequence experimentation from clinically significant risk, protect sensitive information and create safe pathways for legitimate AI use?
Shadow AI is a governance problem before it becomes a technology problem
The attraction of generative AI is easy to understand. Many tools are free or inexpensive, immediately available and simple enough to use without formal training. A staff member facing an urgent task can open an account and begin using AI within minutes, while organizational procurement, cybersecurity review, contracting, privacy assessment and clinical validation may take considerably longer. That difference in speed creates the conditions in which shadow AI develops.
Healthcare workers may not be deliberately trying to bypass governance. They may be trying to finish documentation, translate information, prepare educational materials, find evidence or manage workload pressure. In August 2026, the American Medical Association described the rise of shadow AI, unsanctioned AI tools used by healthcare professionals, as part of a wider challenge for health systems trying to build appropriate governance, data infrastructure, cybersecurity safeguards, accountability and monitoring around rapidly expanding AI use.
A U.S. survey commissioned by Wolters Kluwer Health among 518 healthcare professionals and administrators illustrates the issue. Forty percent of respondents said they had encountered unauthorized AI tools in the workplace without using them, while 17% reported using an unapproved tool themselves. Almost half of those using unapproved tools cited faster workflow, and around one in three pointed to either the absence of an approved tool or inadequate functionality in available alternatives. These findings should not be treated as prevalence estimates for Viet Nam; the survey was conducted in the United States and was not designed to characterize Vietnamese healthcare. They do, however, show how easily AI adoption can move ahead of formal governance when staff have a real operational need.
What should count as shadow AI?
A useful organizational definition needs to be broader than “employees using ChatGPT.” Shadow AI may include a public chatbot accessed through a personal account; an unreviewed writing, translation, transcription or meeting assistant; a mobile application used to analyze work information; an AI browser extension that can access content displayed in other systems; a staff-created chatbot or workflow built on an external platform; or an autonomous or semi-autonomous AI agent connected to organizational files, email, databases or other tools without approval.
The same visibility problem can arise inside software that the organization has already approved. A vendor may introduce a summarization, recommendation, transcription or generative feature after the original procurement review. The underlying software may be authorized, while the new AI function has never been assessed for its data flows, intended use, clinical implications or controls. This may not always be described as shadow AI in the narrowest sense, but from an organizational assurance perspective the core problem is similar.
What matters is not the brand name of the tool. It is whether the organization understands the AI function, intended and actual use, data involved, users, outputs, downstream actions and controls. Shadow AI should also not be confused with illegal AI. A tool does not become unlawful merely because a hospital or clinic has not approved it, and organizational approval does not by itself establish legal compliance or clinical safety. Those are different questions.
What Viet Nam’s legal framework does, and does not, say about shadow AI
Viet Nam’s Law on Artificial Intelligence No. 134/2025/QH15 has been effective since 1 March 2026, with Decree No. 142/2026/NĐ-CP effective from 1 May 2026. The framework establishes AI risk classification and responsibilities for actors including providers, deployers and users, but it does not create a legal category called “shadow AI” or a separate offence of unauthorized organizational AI use.
The distinction is important because organizational approval status and statutory AI risk classification are not the same thing. Article 6 of the AI Law requires stricter risk management for AI applications in essential sectors that directly affect life, health, legitimate rights and interests, or social order and safety. For healthcare specifically, it refers to patient safety, reliability under real-world conditions of use and protection of health data. For high-risk systems, the Law and its implementing framework impose more specific obligations on relevant providers and deployers relating to areas such as risk management, documentation, human oversight, monitoring, transparency, incident handling and accountability.
Decision No. 33/2026/QĐ-TTg, effective from 15 August 2026, contains the current list of high-risk AI systems. In healthcare, the list currently identifies two entries involving AI-assisted surgery and surgical robots under specified conditions. It does not classify every AI tool used in a hospital, clinic or other healthcare organization as legally high-risk merely because it is used in healthcare. A general-purpose AI tool used for an administrative task therefore should not automatically be described as a high-risk AI system under Vietnamese law; the applicable classification depends on the statutory framework and the actual system and use case.
Healthcare organizations may nevertheless choose to apply stricter internal controls. An AI system can create meaningful patient-safety, privacy, cybersecurity or operational concerns without appearing on the statutory high-risk list. Internal organizational risk tiers should therefore be clearly distinguished from the legal risk classifications established under Viet Nam’s AI framework.
The National AI Ethics Framework, issued under Circular No. 05/2026/TT-BKHCN and effective from 10 March 2026, provides an additional governance reference. The Circular applies within its stated scope to state management and public-service AI, while Vietnamese and foreign organizations and individuals engaged in AI research, development, provision, deployment or use in Viet Nam outside that scope are encouraged to apply it. Its guidance includes protecting personal data and confidential information, not entering sensitive information into AI without authorization or appropriate controls, verifying AI outputs, particularly in areas such as health, and maintaining human judgment for important decisions. It should therefore be treated as a useful governance framework where relevant, not as a universal statutory prohibition on every form of unapproved AI use in every healthcare organization.
Personal and health data create an immediate governance boundary
For many healthcare organizations, the most immediate shadow-AI risk is not the model itself. It is the information being sent to it.
Viet Nam’s Law on Personal Data Protection No. 91/2025/QH15 and Decree No. 356/2025/NĐ-CP have both been effective since 1 January 2026. The Law requires personal data to be collected and processed within a specific and clear scope and purpose in compliance with the law and calls for appropriate institutional, technical and human measures to protect personal data. Decree No. 356 identifies health status as sensitive personal data and requires organizations processing sensitive personal data to establish access restrictions, processing procedures and security measures.
Healthcare-specific law adds another layer. Under the current consolidated Law on Medical Examination and Treatment, patients have the right to confidentiality of information in their medical records and other private information provided during care, subject to the exceptions established by law. Article 69 also requires medical records to be stored and kept confidential, while persons permitted to use information from those records must maintain confidentiality and use the information only for the permitted purpose.
This creates a practical boundary that healthcare organizations need to explain clearly to staff. Copying a patient history, laboratory result, medical image, discharge summary, consultation note, recording or screenshot into an external AI service is not simply “asking AI a question.” It may involve processing sensitive personal data, disclosure to another party, storage outside organizational systems or, depending on the architecture of the service, cross-border processing or transfer.
Article 17 of Decree No. 356 expressly treats several situations as cross-border transfers of personal data, including transferring data collected or stored in Viet Nam to servers outside Viet Nam, storing or processing such data on foreign cloud services, transferring data to a foreign recipient, or moving data collected in Viet Nam to a platform outside Viet Nam for continued processing. This does not mean that every external AI service necessarily involves the same cross-border arrangement; organizations need to understand the actual data flow, provider, infrastructure and contractual structure.
The Personal Data Protection Law provides that data that has been properly de-identified is no longer personal data. But simply deleting a patient’s name does not necessarily achieve de-identification. Age, dates, location, images, a rare condition or a distinctive clinical history may still identify or help identify a specific person when considered alone or in combination.
A practical organizational rule can therefore be stated simply: sensitive patient or organizational information should not be entered into an AI service that has not been approved for that information and that use. This is an organizational safety and governance rule; it does not replace the more detailed legal, privacy, cybersecurity and contractual assessment required for approved systems.
The risk depends on what the AI is doing, not simply whether it is approved
A healthcare organization should not treat using AI to improve the grammar of a generic presentation in the same way as uploading an identifiable medical record and asking for a treatment recommendation. The relevant question is not only whether the tool is approved, but what task it performs, what data it receives and what happens if its output is wrong.
An unapproved tool used with generic, non-confidential information for a low-consequence administrative task may create relatively limited patient-safety exposure, although cybersecurity, intellectual-property, contractual or organizational policy issues may still require attention. Concern increases when the tool handles confidential organizational information or personal data, generates patient-facing material, influences clinical documentation, summarizes medical records, interprets images or test results, or supports diagnosis, triage, medication or treatment decisions. The risk can increase further when AI is connected to other systems: an AI agent that can access files, email, records or applications presents a different control problem from a chatbot that only returns text, because potential harm depends not only on what the model says but also on what it can do.
Clinical use deserves a particularly high threshold. Fluency is not evidence of clinical reliability. General-purpose models can produce plausible but incorrect information, omit important context, reflect bias or respond differently to changes in prompting. Performance demonstrated elsewhere does not establish safe performance for a particular Vietnamese patient population, specialty, organization or workflow. WHO’s guidance on large multi-modal models for health emphasizes issues including governance, safety, transparency, human oversight and evaluation, while NIST’s Generative AI Profile provides a voluntary lifecycle-oriented framework for managing generative-AI risks. These are international governance resources rather than Vietnamese statutory requirements, but they are relevant when healthcare organizations design internal controls.
The closer AI moves toward patient-care decisions, the stronger the case becomes for defined intended use, evidence review, local validation where appropriate, human oversight, monitoring, traceability and clear accountability. Shadow clinical AI is especially concerning because it may bypass exactly those safeguards.
Start by making hidden AI use visible
An organization cannot govern AI it does not know exists.
Discovery should begin with people, not just surveillance. Staff can be asked where they already use AI, what tasks they are trying to solve, which tools they use, whether patient or organizational information is involved and where approved alternatives are missing. Anonymous or low-friction reporting can be useful because staff may underreport informal practices if disclosure is assumed to lead automatically to punishment.
Different functions will see different parts of the problem. Clinical departments may know how AI is entering patient workflows; quality teams may see its use in documentation or policy work; IT and cybersecurity may identify applications and extensions; privacy teams may see data-flow risks; procurement may identify contracted tools; and vendor management may discover AI features added to software originally purchased for another purpose. Technical discovery can support this work through identity-management data, approved SaaS lists, application controls, network-security tools or data-loss-prevention systems, but it should be proportionate and consistent with applicable law and organizational policy.
Vendor-added AI deserves specific attention. Organizations should consider requiring appropriate notice when AI functionality, material model behavior, data processing or relevant subprocessors change. An approved vendor should not automatically mean that every future AI feature from that vendor is approved for every clinical or organizational purpose.
The organization’s AI inventory should gradually bring these sources together. NIST’s voluntary AI Risk Management Framework explicitly calls for mechanisms to inventory AI systems as part of governance, with resources aligned to organizational risk priorities. For a healthcare organization, the practical question is therefore not simply “Which AI products have we purchased?” but “Which AI systems are actually influencing our data, workflows, staff or patients today?”
Shadow AI is the gap between that answer and reality.
Make the safe pathway easier than the workaround
A long AI policy is of little value if frontline staff cannot understand what it means for the task in front of them. A practical policy should allow users to distinguish quickly between activities permitted with approved tools, activities that require prior review and activities that should not occur. It should address patient and confidential information, clinical documentation, patient-facing communication, translation and transcription, images and recordings, clinical decision support, software integrations, personal accounts and autonomous actions.
The policy also needs an accessible path for uncertainty. Staff should know where to ask whether a proposed use is acceptable and how they can obtain a decision. Healthcare organizations can support this by maintaining an approved AI catalogue identifying which tools may be used, by whom, for which purposes and with which types of information. Where justified, enterprise accounts may offer stronger contractual, privacy, security and administrative controls than personal accounts.
Governance should also be proportionate. A low-consequence use case should not necessarily require the same review pathway as a system influencing diagnosis or treatment. A rapid review mechanism for lower-risk requests can reduce incentives to bypass governance, while clinically consequential systems should still receive appropriate evidence review, privacy and cybersecurity assessment, regulatory assessment where relevant, validation and clinical oversight.
Education is equally important. Staff need examples that resemble the choices they actually face: whether AI can be used to translate patient instructions; whether an approved tool can help draft a letter that is then reviewed by a clinician; whether a consultation may be recorded through an AI transcription application; whether an internal policy may be pasted into a public chatbot; or whether a photograph of a clinical image may be uploaded to an AI application. Training should explain which information is sensitive, which tools are approved, which uses require additional review, how AI outputs should be checked and how an accidental disclosure or unsafe output should be reported.
Technical controls such as managed browsers, application restrictions, web filtering, identity management, endpoint management and data-loss-prevention tools can reinforce these arrangements, but they cannot govern AI by themselves. Staff may use personal devices, new services appear continuously, and AI may be embedded in otherwise permitted software. Overly broad blocking can also push legitimate experimentation further out of view if no acceptable alternative exists. The aim should therefore not be to make AI use difficult. Responsible governance should make the safe pathway more visible, credible and usable than the workaround.
When shadow AI is discovered, assess the use before deciding the response
Finding an unapproved AI tool should trigger triage, not an automatic assumption that harm or illegality has occurred. The organization should determine what tool was used, for what purpose, by whom, with what information, under which account, how frequently and whether data was transmitted outside the organization. It should also determine whether the provider retained or further processed the data, whether AI-generated content entered the medical record or other organizational systems, whether an output influenced patient care and whether the same practice is being used elsewhere.
The appropriate response may range from registering and reviewing a low-consequence use to moving the user to an approved alternative, stopping a workflow, securing affected information, contacting the vendor, initiating a privacy or cybersecurity assessment, reviewing affected patient care or opening a formal incident investigation.
If an AI-related event meets the definition of a serious incident under Article 19 of Decree No. 142/2026/NĐ-CP, specific responsibilities apply. Deployers and users must promptly record the incident, take necessary measures to limit its consequences and notify the provider; providers must implement necessary technical measures; and providers and deployers must cooperate in providing information and addressing the incident. Providers or deployers, as applicable, must also submit preliminary reports to the competent authority within the timelines established by Article 19, while providers and deployers must preserve relevant system logs, data and information for verification, assessment and remediation. If the incident is also subject to reporting under cybersecurity, personal-data-protection or other specialized legislation, the corresponding reporting regime applies.
The fact that an AI tool was unapproved does not, by itself, mean that a serious incident has occurred. Equally, formal organizational approval does not prevent an AI-related event from meeting the statutory threshold.
A practical Shadow AI Governance Gate
When an unapproved or previously unknown AI use is identified, healthcare organizations can work through the following questions:
Use case: What task is the AI actually performing, and what problem is the user trying to solve?
Tool and ownership: What system, model, application, extension or embedded feature is being used, and under whose account or contract?
Data: Is patient information, sensitive personal data, confidential organizational information, images, audio, documents or other protected information being entered or exposed?
Data destination: Where is the information processed or stored, by whom and for how long, and could cross-border processing or transfer be involved?
Clinical consequence: Can the output influence diagnosis, treatment, triage, medication, clinical documentation, patient communication or another care decision?
Automation: Does the AI only provide information, or can it access systems, change records, send communications, trigger workflows or perform other actions?
Evidence: Has the system been evaluated for the intended purpose, population and setting, and are important limitations understood?
Human oversight: Who reviews the output before it affects a patient, record, decision or external communication?
Traceability: Can the organization reconstruct what information was provided to the AI, what output was produced and how it was used when this is necessary for safety or investigation?
Legal and regulatory status: What classification and other regulatory requirements apply to the particular system and use case? Internal organizational risk should not be confused with statutory AI risk classification.
Security and vendor controls: Have relevant cybersecurity, access, contractual, retention, model-training, subcontractor and vendor risks been reviewed?
Existing exposure: Has the use already created a privacy, cybersecurity, clinical or other incident requiring action?
Decision: Should the use be approved, approved with conditions, moved to an approved alternative, temporarily restricted or stopped?
Learning: Does the discovery reveal a broader workflow problem that the organization should solve instead of repeatedly policing the same workaround?
The Shadow AI Governance Gate is a practical organizational framework proposed for healthcare governance. It is not a statutory AI risk-classification methodology, conformity-assessment procedure or legal requirement established by Viet Nam’s AI legislation.
Connect shadow-AI governance with systems the organization already has
Healthcare organizations do not need a separate bureaucracy for every new AI problem. Shadow-AI governance can connect with structures that already exist: AI inventory and governance, information security, personal-data protection, vendor management, clinical governance, medical-device management where relevant, quality and patient safety, workforce education and incident management.
This integration matters because the same event can cross several boundaries. An unauthorized transcription application might initially appear to be an IT issue but could also involve sensitive personal data, medical-record confidentiality, cross-border processing and inaccurate clinical documentation. A generative-AI workflow may begin as an efficiency experiment but become a patient-safety concern if staff start relying on its outputs in clinical decisions. The objective is therefore not to create an “AI committee” for every tool, but to ensure that existing organizational controls can recognize when AI changes the risk profile of a workflow and route the issue to the right people.
What this means for healthcare organizations in Viet Nam
The need for this capability is becoming more important as Viet Nam actively expands AI adoption. Decision No. 1671/QĐ-TTg, issued on 28 August 2026, approved the National AI Strategy to 2030 with a vision to 2045. The Strategy includes healthcare applications such as diagnosis and treatment support, remote care, healthcare-resource coordination, hospital management, patient-flow optimization and health-record management, while one of its strategic pillars focuses specifically on safe, trustworthy and responsible AI governance.
That direction supports wider AI adoption. It does not reduce the need for healthcare organizations to understand what AI is actually being used inside their operations.
Viet Nam now has an AI Law and implementing Decree, a statutory high-risk AI list, a National AI Ethics Framework, a National AI Strategy, a Personal Data Protection Law and implementing Decree, alongside the Cybersecurity Law No. 116/2025/QH15, which has been effective since 1 July 2026. The external governance environment is becoming substantially clearer.
But legislation cannot tell a hospital that an employee has just opened a personal AI account, that a browser extension can read information displayed on a workstation, that a team has begun using a chatbot to prepare clinical material, or that a vendor introduced a new AI feature last week. Those remain organizational visibility and governance problems.
The appropriate response is neither to assume that every unapproved use is dangerous or unlawful nor to ignore informal AI because the organization never authorized it. A mature approach makes AI use visible, protects sensitive information, distinguishes different levels of consequence, provides workable approved alternatives, evaluates clinically significant uses appropriately and gives staff a realistic way to ask, “Can I use AI for this?”
The most useful question for leadership is therefore not, “Have we banned unapproved AI?” It is, “Do we know where AI is entering our work, and can our people reach a safe, approved pathway before they need a workaround?”
The goal is not to eliminate useful experimentation. It is to prevent hidden experimentation from becoming hidden patient risk.
Professional education notice
This article is intended for professional education and knowledge exchange. It does not constitute individualized legal, regulatory, clinical, cybersecurity, personal-data-protection, employment or medical-device advice.
“Shadow AI” and the “Shadow AI Governance Gate” are used here as practical organizational governance concepts. Shadow AI is not a statutory risk category defined by Viet Nam’s Law on Artificial Intelligence, Decree No. 142/2026/NĐ-CP or Decision No. 33/2026/QĐ-TTg. An AI system does not become legally high-risk solely because it has not been approved by a healthcare organization, and unapproved organizational use should not automatically be characterized as unlawful.
The legal responsibilities applicable to a particular AI use depend on factors including the system, its function, intended and actual use, statutory risk classification, data processed, relevant actors, clinical context and other applicable legislation. Internal healthcare governance may appropriately impose controls beyond minimum statutory requirements when proportionate to patient safety, privacy, cybersecurity or organizational risk.
References to WHO, NIST, AMA and other international materials are included as evidence or governance considerations and should not be interpreted as Vietnamese statutory requirements unless separately made applicable through Vietnamese law, regulation, contract, certification commitment or another binding basis.
References
[1] National Assembly of Viet Nam. Law No. 134/2025/QH15 on Artificial Intelligence. Adopted 10 December 2025; effective 1 March 2026.
[2] 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.
[3] 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.
[4] Ministry of Science and Technology of Viet Nam. Circular No. 05/2026/TT-BKHCN Promulgating the National Artificial Intelligence Ethics Framework. Issued and effective 10 March 2026.
[5] Prime Minister of Viet Nam. Decision No. 1671/QĐ-TTg Approving the National Artificial Intelligence Strategy to 2030, Vision to 2045. Issued and effective 28 August 2026.
[6] National Assembly of Viet Nam. Law No. 91/2025/QH15 on Personal Data Protection. Adopted 26 June 2025; effective 1 January 2026.
[7] 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.
[8] Office of the National Assembly of Viet Nam. Consolidated Document No. 26/VBHN-VPQH on the Law on Medical Examination and Treatment. Consolidated 26 February 2026, incorporating Law No. 15/2023/QH15 and subsequent amendments.
[9] National Assembly of Viet Nam. Law No. 116/2025/QH15 on Cybersecurity. Adopted 10 December 2025; effective 1 July 2026.
[10] World Health Organization. Ethics and Governance of Artificial Intelligence for Health: Guidance on Large Multi-Modal Models. Publication page dated 25 March 2025.
[11] National Institute of Standards and Technology. Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1). Published 26 July 2024; NIST webpage updated 8 April 2026.
[12] National Institute of Standards and Technology. AI Risk Management Framework and AI RMF Playbook. Voluntary resources for organizational AI risk management, including AI-system inventory and third-party governance.
[13] American Medical Association. AI Is Moving Fast. Health Systems Need Guardrails. 20 August 2026.
[14] Wolters Kluwer Health. Shadow AI: A Hidden Risk to Healthcare — Survey Findings. Survey of 518 U.S. healthcare professionals and administrators, conducted in December 2025 and reported in January 2026. Findings should not be generalized as prevalence estimates for Viet Nam.