Contractual Data Governance for Healthcare AI in Viet Nam: Governing Secondary Use, Model Training, Outputs, Local Performance Data, Re-identification and Audit Rights
PRACTICAL FRAMEWORKS & TOOLS
Contractual Data Governance for Healthcare AI in Viet Nam: Governing Secondary Use, Model Training, Outputs, Local Performance Data, Re-identification and Audit Rights
PRACTICAL FRAMEWORKS & TOOLS
Last reviewed: 7 October 2026
When a healthcare organization buys or deploys AI, contractual discussions often begin with familiar issues: price, implementation, service levels, cybersecurity, support and liability. Healthcare AI creates another set of questions that can be just as important: what happens to the data once the AI relationship begins?
A vendor may need patient information to deliver the service. The healthcare organization may send images, audio, prompts, clinical notes, structured records or other information to the system. AI may generate new information from those data, and outputs may enter the medical record. The vendor may collect logs, error information and clinician feedback. Local validation and post-deployment monitoring may generate new performance data. Some of this information may later be proposed for service improvement, model tuning, benchmarking, research, development of another product or training a model used for other customers.
If the contract says only that the vendor may process information “to provide and improve the services,” the healthcare organization may still not know what it has actually authorized.
For healthcare organizations in Viet Nam, contractual data governance should therefore mean defining which data may be used, by whom, for which purposes, through which parties and locations, for how long, under what restrictions, and with what rights of access, oversight, verification, deletion, portability and exit.
The legal context makes these questions particularly important. Law No. 91/2025/QH15 on Personal Data Protection and Decree No. 356/2025/NĐ-CP have both been effective since 1 January 2026 [1,2]. Article 37 of the Law requires a personal-data controller to specify relevant responsibilities, rights and obligations of the parties in agreements or contracts relating to personal-data processing. A processor may receive personal data only after an agreement or contract with the controller or controller-processor has been established and must process the data in accordance with that arrangement [1].
This makes contractual data terms part of healthcare AI governance rather than merely procurement boilerplate.
One principle should nevertheless remain clear from the beginning:
A contract does not make an otherwise unlawful use of personal data lawful.
Contractual provisions allocate and operationalize responsibilities within an arrangement that must itself comply with Vietnamese law. Contractual permission does not, by itself, satisfy the legal conditions for personal-data processing and cannot override mandatory requirements concerning health information, data-subject rights, personal-data impact assessments, cross-border transfers, cybersecurity, medical records or other applicable obligations.
Contractual data governance is narrower than privacy governance, and deeper than confidentiality
This article is deliberately about contractual data governance, not the entire privacy, cybersecurity or AI-governance framework.
Privacy governance determines whether and how personal data may lawfully be processed. Cybersecurity addresses technical and organizational protections. Procurement assesses the vendor and commercial relationship. Research governance may apply when information is used for research. AI governance determines whether a technology and use case are appropriate for the healthcare organization.
Contractual data governance asks a narrower but deeper question:
Once the organization enters a relationship with another party, what data-related rights, restrictions, responsibilities and evidence need to be built into that relationship?
A general confidentiality clause rarely answers that question.
Confidentiality may restrict disclosure, but it may not tell the organization whether a vendor can use clinical data to train a general model, combine information with data from other customers, generate cross-customer benchmarks, retain derived information after termination, add another processor, move processing to another country, monetize performance information or materially change model-training practices without notice.
Those issues require more precise contractual treatment.
Map the parties, roles and data before negotiating permissions
Contractual governance should begin from what the parties actually do, not from commercial labels.
Calling a technology company “the vendor” does not determine whether it acts as a personal-data processor, controller, controller-processor, third party or potentially performs different roles for different activities.
A company may, for example, process clinical information solely on the healthcare organization's instructions to provide an ambient documentation service. If the same company separately determines that it will use those data to improve a general model for its own broader purposes, the legal and governance analysis may be different.
Before drafting contractual permissions, healthcare organizations should therefore understand what data each party receives, which party determines the relevant purposes and means of processing, whether the vendor has additional independent purposes, and which downstream organizations receive or process the information.
This is particularly important under Article 37 of the Personal Data Protection Law because processors must operate according to the relevant processing agreement or contract and the Law assigns different responsibilities to controllers, processors and controller-processors [1].
The organization should also define the data themselves. Generic terms such as “customer data” or “service data” can conceal materially different categories: source clinical information, medical images, audio recordings, prompts, uploaded documents, structured records, metadata, system logs, AI outputs, embeddings, clinician corrections, user feedback, local validation results, incident information, performance metrics and derived analytics.
Contractual rights do not necessarily need to be identical across all of those categories.
This is particularly important because information created by AI can itself remain personal data. Decree No. 356 provides that information resulting from AI inference must receive personal-data protection where it can identify or help identify a particular individual. The Decree also permits personal data to be used for research and development of self-learning algorithms, AI systems and other automated systems only subject to compliance with personal-data protection requirements [2].
The contract should therefore avoid assuming that anything labeled “derived data” automatically becomes unrestricted vendor data.
Separate service delivery, secondary use and model training
One of the most important contractual distinctions is the difference between using data to provide the contracted service and using those data for another purpose.
A vendor may need information to operate an AI service for the healthcare organization. That does not automatically mean the same information should be available to train a general model for other customers.
The contract should distinguish purposes such as providing and securing the contracted service, generating organization-specific outputs, performing agreed local analytics, improving an organization-specific configuration, fine-tuning an organization-specific model, improving a general commercial model, benchmarking across customers, developing new products and conducting research.
These purposes are not interchangeable.
The phrase “use data to improve the service” can be especially problematic when left undefined. “Improvement” might mean correcting bugs, analyzing error rates, using clinician feedback to improve the organization's configured service, adding local information to a general training corpus or developing an entirely different commercial product. Those are materially different activities and should not be collapsed into one broad contractual permission.
Model training therefore deserves an explicit decision.
If training is not necessary to deliver the contracted service, an organization may choose to prohibit it unless separately authorized. Where some form of model improvement is intended, the contract should define which information may be used, which model is being improved, whether the improvement is organization-specific or shared across customers, whether data may be combined with information from other organizations, which safeguards apply and what happens if the relationship later changes.
The contract should also distinguish retrieval or indexing from model training. A system may create embeddings or retrieve approved organizational content through a RAG architecture without altering the underlying foundation model. Another system may use interactions or corrections to update model parameters. Both use organizational information, but they are not the same technical or contractual activity.
Generic terms such as “AI improvement” should not hide that distinction.
Health information, de-identification and re-identification need careful treatment
Healthcare data are not ordinary enterprise information.
Article 26 of the Personal Data Protection Law generally requires the consent of the personal-data subject when collecting and processing health information, subject to the exceptions specified in Article 19(1) and other applicable legal provisions. Relevant personal-data protection requirements must also continue to be observed [1]. Decree No. 356 lists health status among sensitive personal data [2].
The contractual implication is not simply to state that health data are confidential. Healthcare organizations need to know which health information the vendor receives, what it does with that information, what additional purposes are contemplated, where processing occurs and who receives the information downstream.
De-identification should also not become a contractual shortcut.
International technology agreements frequently use terms such as de-identified, anonymized and pseudonymized. Healthcare organizations in Viet Nam should not assume that terminology derived from another legal framework, including U.S. HIPAA terminology, automatically determines the information's legal status under Vietnamese law.
Under the Personal Data Protection Law, personal data that have been effectively de-identified no longer constitute personal data, while the Law also prohibits re-identification of de-identified personal data except where otherwise permitted by law [1]. Contractual governance should therefore focus on whether the resulting data can still identify or help identify a particular individual, including where they are combined with other information.
Removing a patient's name alone may not achieve that result. Dates, rare conditions, free text, imaging, genomic information or combinations of attributes may still enable identification.
Where de-identified data are provided to another party, the contract should also address unauthorized re-identification. Joint Commission–CHAI guidance specifically recommends defining permitted uses, data minimization, prohibition of re-identification, third-party obligations and audit rights in data-use agreements [5].
That is a useful international contractual-governance practice, but it should not be presented as a contractual clause mandated in that exact form by Vietnamese law.
Secondary use should be visible rather than assumed
Healthcare AI creates many opportunities for data to acquire another purpose.
Information initially processed to support care may later become useful for quality improvement, model training, product development, commercial analytics, benchmarking, research or algorithm development.
Joint Commission's Responsible Use of Health Data™ (RUHD) Certification focuses on responsible secondary use of de-identified health data. Its standards address governance, de-identification, protection against unauthorized re-identification, limitations on use, algorithm validation and patient transparency [3].
RUHD is a voluntary U.S. certification framework, not Vietnamese law. Its relevance here lies in a broader governance principle: information used appropriately for one purpose should not quietly acquire a new purpose merely because it has scientific or commercial value.
A healthcare AI contract should therefore make secondary use explicit. It may prohibit particular secondary uses, authorize specifically defined uses or require separate approval before additional purposes begin.
Broad phrases such as “internal business purposes,” “analytics,” “service improvement,” “research” or “product development” should be examined carefully. Operational analytics needed to measure service uptime are different from analytics used to build another commercial product. Security monitoring is different from model training. Aggregate operational statistics are different from benchmarking local clinical performance across customers.
Contractual language should reflect those distinctions.
Outputs, local performance data and clinician feedback need defined rights
Healthcare AI generates valuable information after deployment.
A hospital may produce local sensitivity and specificity results, subgroup analyses, override patterns, incident data, error rates, user-feedback information, workflow measures, utilization data and outcome evaluations. These local performance data can become essential to quality management and patient safety.
A healthcare organization should not become dependent on the vendor for access to evidence about how the technology performs in its own setting. Contracts should therefore consider the organization's rights to access, retain, analyze and export relevant local performance information throughout the relationship and, where appropriate, after it ends.
Joint Commission–CHAI guidance specifically recommends considering rights relating to model outputs, local performance data and monetization of data when defining permitted and prohibited uses [5].
The same issue arises with AI outputs. A hospital using an ambient documentation system needs to be able to use generated documentation in clinical care and the medical record. Decision-support systems may generate scores or classifications that need to be retained for governance, while generative systems may produce materials the healthcare organization needs to review, modify or reuse.
The contract should therefore define appropriate rights to use, store, review, reproduce, modify and retain relevant outputs and clarify any rights the vendor claims to reuse them.
Clinician corrections and feedback deserve similar attention. When clinicians edit AI-generated notes, override recommendations, annotate images or report hallucinations, they generate information about system performance. A vendor may reasonably want to use that information for improvement, but the contract should clarify whether it is used to improve only the organization's configured service, to improve a general model used by other customers, or to support another commercial purpose.
Clinician corrections can reflect local clinical practice, professional judgment, safety weaknesses and potentially patient information. They should not automatically become unrestricted training data merely because they were generated while using the service.
Contractual rights should be more precise than a simple claim of “ownership”
Healthcare AI contracts often ask who owns the data, outputs or performance results.
Ownership terminology can be useful, but it may be too blunt for many digital-information relationships. Different legal rules may apply to personal data, confidential information, databases, clinical records, intellectual property, software outputs and derived analytics.
In practice, it is often more useful to define specific rights. The contract can address who may access particular information, which purposes are permitted, whether disclosure to third parties is allowed, whether the information may be retained or combined with other data, whether it may be used to train another model, whether it may be commercialized, whether the healthcare organization can export it, and whether safety findings may be provided to regulators, quality functions or other legitimate oversight mechanisms.
Those practical rights may matter more operationally than one broad statement that a party “owns” the information.
Questions concerning intellectual property or proprietary rights should nevertheless be assessed separately under the relevant legal framework and facts.
Performance evidence should remain usable for patient safety
Vendor confidentiality can be legitimate, particularly where proprietary technology or commercially sensitive information is involved.
Confidentiality should not, however, prevent a healthcare organization from using its own local performance and safety evidence for legitimate internal governance or from making disclosures required by applicable law.
If an evaluation identifies poor performance, subgroup disparities, unsafe workflow effects or other clinically important concerns, the healthcare organization needs sufficient contractual freedom to act on those findings.
Where appropriate, the parties may separately address external scientific publication, public benchmarking or disclosure of proprietary information.
The aim is not to eliminate confidentiality. It is to prevent confidentiality from becoming a barrier to patient-safety governance.
The same applies to benchmarking and monetization. Cross-customer benchmarking may provide useful quality information, but contracts should specify what information contributes to the benchmark, whether patients or organizations are identifiable, whether local results can be used in marketing or commercial products, and whether benchmark information can be sold or licensed.
If patient data, derived information, AI outputs or local performance evidence can be monetized, that possibility should be visible rather than buried inside generic service terms. Contractual permission, however, does not by itself establish that the underlying processing complies with applicable law.
Subprocessors, onward transfer and processing locations must remain visible
The organization signing the contract with the healthcare provider may not be the only organization processing the data.
An AI service may rely on cloud infrastructure, a foundation-model provider, a speech-transcription service, analytics services, security providers or other subprocessors. Contractual data governance should therefore extend beyond the first vendor.
A risk-proportionate contract may require the vendor to maintain appropriate visibility over material subprocessors, impose relevant protections downstream, notify the healthcare organization of material changes and establish a mechanism for further review where a change materially alters risk or legal obligations.
Onward transfer also deserves attention. A vendor may provide data to another entity, move them to another cloud environment or make information available to an affiliate. The contract should clarify whether those transfers are permitted, for which purposes and subject to what safeguards.
Relevant obligations should flow downstream as appropriate, including purpose restrictions, confidentiality, security, retention expectations, restrictions on unauthorized re-identification and incident cooperation.
The primary contracting party should not be able to make the healthcare organization's protections ineffective merely by outsourcing processing to another provider.
Cross-border processing needs contractual support for regulatory compliance
The Personal Data Protection Law defines cross-border transfer broadly. It includes transferring personal data stored in Viet Nam to storage outside Viet Nam, providing personal data from Viet Nam to a foreign organization or individual, and using an overseas platform to process personal data collected in Viet Nam [1].
Unless an applicable statutory exemption applies, the organization or person carrying out the cross-border transfer must prepare the required cross-border transfer impact-assessment dossier and submit one original dossier to the competent personal-data protection authority within 60 days from the first cross-border transfer [1].
Decree No. 356 is particularly relevant contractually because the applicable dossier requires information and documentation concerning the parties and processing arrangement, processing flows, storage and deletion, safeguards, onward transfer and relevant risks [2].
An AI contract should therefore provide the healthcare organization with enough information and cooperation to meet applicable requirements. The organization should not discover after signing that the vendor cannot clearly describe processing locations, downstream recipients or other information necessary for its regulatory assessment.
Processing locations can also change over time. A vendor may move a service to another region or introduce an overseas subprocessor without materially altering the clinical interface. Where such a change affects legal or risk obligations, the healthcare organization should receive sufficient notice to conduct appropriate review.
That does not mean every infrastructure modification requires separate approval. The contract should identify the types of location or processing changes material enough to trigger governance review.
Contracts should support personal-data impact assessment
The contractual relationship should also support the organization's personal-data processing impact assessment.
Under Article 21 of the Personal Data Protection Law, the controller or controller-processor must prepare and retain the processing-impact dossier and submit one original dossier to the competent personal-data protection authority within 60 days from the first day of processing, subject to statutory exceptions. Processors also have obligations to prepare and retain the relevant impact-assessment dossier for processing performed on behalf of the controller, subject to the applicable statutory framework [1].
A healthcare organization cannot prepare a meaningful assessment if its AI vendor cannot or will not explain important aspects of processing.
Contracts should therefore provide for reasonable cooperation in supplying relevant information concerning data categories, purposes, processing activities, retention, security controls, recipients, processing locations and other matters necessary for governance and regulatory documentation.
Contractual transparency should support regulatory transparency.
Audit and assurance rights should be proportionate and usable
The phrase “audit rights” can sound confrontational, but the underlying principle is simple: the healthcare organization needs a credible way to verify that material contractual data protections are being followed.
Joint Commission–CHAI guidance recommends that healthcare organizations consider reserving rights to audit third-party vendors for compliance with data-use agreements [5].
That is an international governance recommendation, not a universal audit clause required by Vietnamese law.
Audit rights also do not necessarily mean unrestricted access to a vendor's systems. Depending on risk, assurance may come from independent audit reports, security or privacy assessments, certifications, compliance attestations, relevant logs, current subprocessor information, incident documentation, deletion confirmations or targeted review after a significant event.
Higher-consequence AI may justify stronger assurance rights than a low-consequence administrative service.
The important question is whether the contractual mechanism is usable enough to provide meaningful assurance.
Assurance should also extend beyond cybersecurity where relevant. A healthcare organization may need evidence that prohibited model training is not occurring, that downstream processing remains consistent with the agreed arrangement, that required deletion has occurred, that processing locations are accurately represented, or that local performance information has not been used for unauthorized benchmarking or commercialization.
Incident cooperation should be agreed before something goes wrong
If personal data are exposed, used outside the agreed purpose or transferred to an unauthorized recipient, the healthcare organization may have regulatory and patient-safety responsibilities of its own.
The contract should therefore define how the vendor will notify the organization, preserve relevant evidence, investigate the event, provide appropriate root-cause information, contain the problem, implement corrective action and cooperate with regulatory or organizational response.
Contractual notification timing should be designed so that the healthcare organization can meet any applicable statutory reporting deadline.
This should not be limited to classic cybersecurity breaches. Unauthorized model training, unapproved disclosure to a new processor or unexpected reuse of health information may create contractual data-governance concerns even where the issue does not present as a conventional security incident.
Data incidents and AI safety incidents can also overlap. If a model is retrained on inappropriate or corrupted local data and a resulting update affects clinical output, the problem crosses data governance and patient safety. If identifiable information unexpectedly appears in generated content, both privacy and clinical workflow may be involved.
Organizational processes should therefore allow important contractual data incidents to reach privacy, cybersecurity, AI governance, quality and patient-safety functions as appropriate.
Retention, termination and model training should be designed together
Data retention should not be governed by a generic phrase such as “for as long as necessary.”
Source clinical information, temporary processing copies, logs, prompts, AI outputs, local performance data and incident records may have different purposes and legal or operational retention requirements.
The contract should identify the important categories the vendor retains, why they are retained, for how long and what happens when their purpose ends.
Termination also requires more than closure of the customer account.
The healthcare organization should understand what happens to source data, backups, prompts, embeddings, derived information, local performance data, organization-specific model configurations and other relevant information when the relationship ends. The contract should distinguish what is returned, deleted, retained because of applicable law or retained under another separately authorized purpose.
Where deletion is required, the organization may need an appropriate form of confirmation.
Model training adds another complication. Once data have influenced model parameters, deleting the original records used for training may not necessarily remove their influence from the model, and technical mechanisms for reversing that influence may be limited or highly context-dependent.
Healthcare organizations should therefore not rely on a generic termination-and-deletion clause as a substitute for deciding beforehand whether their data may be used for model training.
Where training is authorized, contractual governance should consider what happens to organization-specific fine-tunes or model improvements after termination, whether continued use is permitted and what relevant technical limitations exist.
Material vendor changes should not silently expand data rights
AI vendors and their technology stacks evolve quickly.
A company may adopt a new foundation-model provider, add cloud infrastructure, change processing locations, introduce a new analytics function or create a new product-improvement program.
Those changes can materially alter the data relationship even when the clinical interface appears unchanged.
The contract should therefore identify which changes require notice and, where appropriate, renewed organizational review. Material changes may include new processing purposes, changed model-training practices, new significant subprocessors, new cross-border locations, different retention arrangements, new monetization practices or system changes that materially alter data use.
Healthcare organizations should also understand the contractual hierarchy where cloud services incorporate online terms, privacy notices or acceptable-use policies by reference.
If negotiated healthcare data protections conflict with a vendor's general online terms, the agreement should make clear which terms prevail. It should also address whether material data-use rights may be changed unilaterally and how the healthcare organization will be informed about relevant changes.
For clinically important AI, relying entirely on modifiable public terms can create weak governance.
Local validation and negative evidence require clear contractual boundaries
Healthcare organizations may generate local validation datasets and results before routine deployment. Those results can have significant clinical and commercial value because they demonstrate how an AI system performs in a new population or healthcare environment.
The contract should therefore address how local validation information and findings may be used.
The healthcare organization may need those results for internal governance, quality improvement, research, regulatory or other legitimate purposes. A vendor may want to use the same evidence for product development, marketing, benchmarking or regulatory submissions elsewhere.
Those uses should be discussed explicitly rather than assumed.
Contractual arrangements should also avoid creating incentives to suppress unfavorable evidence. If local validation or post-deployment evaluation identifies poor performance, subgroup differences, workflow problems or patient-safety concerns, the healthcare organization needs sufficient rights to use that information for appropriate governance and corrective action.
A responsible healthcare AI contract should protect learning even when the evidence is uncomfortable.
The Data Law may also be relevant
Healthcare AI contractual governance should not be reduced entirely to personal-data law.
Viet Nam's Data Law No. 60/2024/QH15 has been effective since 1 July 2025 and regulates digital data more broadly. The current consolidated text is Consolidated Document No. 42/VBHN-VPQH dated 16 March 2026, incorporating subsequent amendments, including amendments under Law No. 116/2025/QH15 that took effect on 1 July 2026 [4].
The Data Law includes provisions concerning categories such as important data and core data and their cross-border transfer and processing.
Whether those provisions apply depends on the particular data and context.
A healthcare AI contract should therefore be assessed against the full applicable data framework, not only personal-data protection rules.
A practical Contractual Data Governance Framework
Healthcare organizations in Viet Nam can structure contractual data governance around the following sequence:
map the data → map the parties and roles → define primary purposes → define permitted secondary uses → govern model training → govern outputs and derived data → control subprocessors and onward transfer → map processing locations → define retention and deletion → preserve access and portability → establish incident cooperation → establish assurance and audit rights → govern material changes → plan exit
Map the data means identifying the important categories entering, leaving and being generated through the AI service.
Map the parties and roles means understanding which party determines purposes and means and which parties process each category of information, then reflecting those responsibilities accurately in the contractual arrangement.
Define primary purposes means identifying what processing is genuinely necessary to provide the contracted service.
Define permitted secondary uses means deciding which additional activities, if any, are authorized and restricting use outside those boundaries.
Govern model training means distinguishing ordinary service provision, local configuration, organization-specific fine-tuning and broader model improvement.
Govern outputs and derived data means defining practical access and use rights for AI outputs, clinician feedback, local performance evidence and other derived information.
Control subprocessors and onward transfer means maintaining visibility and appropriate protections across the downstream processing chain.
Map processing locations means making cross-border arrangements visible early enough for legal and governance assessment.
Define retention and deletion means deciding how long relevant data categories persist and what happens when their purpose ends.
Preserve access and portability means ensuring that the healthcare organization can obtain the information and evidence needed for care, monitoring, assurance and exit.
Establish incident cooperation means defining notification, evidence preservation, investigation and corrective-action responsibilities before an event occurs.
Establish assurance and audit rights means creating practical mechanisms for verifying material contractual commitments.
Govern material changes means preventing important expansion of data use, processing location or downstream access from occurring silently.
Plan exit means dealing with return, deletion, residual information, model-training consequences, transition support and continuity before the commercial relationship ends.
This Contractual Data Governance Framework is proposed here as a practical organizational approach. It is not an official Vietnamese contract template, a mandatory contractual structure or a Joint Commission or CHAI model agreement.
What healthcare organizations should avoid
One common failure is purpose ambiguity, where broad contractual language allows information to be used for “improvement” without explaining what that means. Another is training by default, where ordinary use of a service automatically authorizes local data to improve a general model. A third is derived-data ambiguity, where broad terms give the vendor extensive rights over AI outputs, telemetry, clinician feedback or local performance evidence simply because its technology helped generate them.
Organizations should also be cautious about undefined claims that information is “de-identified,” silent addition of important subprocessors, unrestricted onward transfer, limited access to local performance information, indefinite retention, absence of meaningful deletion evidence, overly broad restrictions on safety reporting and unilateral changes to material data-use terms.
The objective is not to make every AI contract longer.
It is to make the important boundaries explicit.
What this means for healthcare organizations in Viet Nam
Healthcare AI increasingly creates relationships in which data continue to move and generate value long after the first clinical transaction.
The healthcare organization may begin by acquiring a diagnostic, documentation or administrative service, but the resulting relationship can involve patient data, derived information, clinician corrections, AI outputs, local performance evidence, cloud infrastructure, downstream processors and model improvement.
That makes the contract part of the AI governance system.
The practical transition is therefore:
from “the vendor may process our data” → to “which data, for which purpose, through which parties, under which controls, with what rights, and what happens afterward?”
Good contractual data governance does not mean preventing legitimate AI innovation. It means allowing innovation within boundaries the healthcare organization can understand, justify, monitor and enforce.
For healthcare AI, those boundaries matter because data are not simply an input to technology. They sit within the relationship among patients, healthcare organizations, technology providers and the care system.
Responsible healthcare AI therefore requires not only sound privacy policies and secure systems, but contractual arrangements that make secondary use, model training, output rights, local performance data, downstream processing, assurance and exit visible before they become disputes, compliance problems or patient-safety concerns.
Professional education notice
This article is intended for professional education and knowledge exchange. It does not constitute individualized legal, contractual, procurement, intellectual-property, personal-data protection, cybersecurity, clinical, medical-device, research-ethics or regulatory advice.
The term contractual data governance is used in this article as a practical organizational governance concept. It is not a defined statutory category under Vietnamese law.
Contractual provisions cannot, by themselves, satisfy all legal conditions for personal-data processing and cannot waive mandatory obligations under Vietnamese law.
The contractual controls discussed in this article, including restrictions on model training, secondary-use limitations, re-identification prohibitions, output and performance-data rights, downstream-party controls and audit or assurance rights, are practical governance approaches. They should be adapted to the specific legal relationship, bargaining position, data, technology and deployment context.
Joint Commission RUHD Certification and Joint Commission–CHAI RUAIH guidance are used as international reference points. They are U.S.-oriented voluntary healthcare governance frameworks and do not create legal requirements in Viet Nam.
Questions concerning ownership, intellectual-property status or other proprietary rights in data, AI outputs, model improvements or derived information should be assessed separately under the relevant legal framework and specific facts. For many operational purposes, contracts should define concrete rights of access, use, disclosure, modification, retention, commercialization and transfer rather than relying solely on broad ownership terminology.
References
National Assembly of Viet Nam. Law No. 91/2025/QH15 on Personal Data Protection. Adopted 26 June 2025; effective 1 January 2026.
Government of Viet Nam. Decree No. 356/2025/NĐ-CP Detailing a Number of Articles and Measures for Implementation of the Law on Personal Data Protection. Issued 31 December 2025; effective 1 January 2026.
Joint Commission. Responsible Use of Health Data™ Certification. Voluntary U.S. certification addressing the responsible secondary use of de-identified health data; available since 1 January 2024.
National Assembly of Viet Nam. Law No. 60/2024/QH15 on Data. Effective 1 July 2025, as amended. Current consolidated text: Office of the National Assembly. Consolidated Document No. 42/VBHN-VPQH dated 16 March 2026.
Joint Commission and Coalition for Health AI (CHAI). Guidance on the Responsible Use of AI in Healthcare (RUAIH). Initial high-level guidance, released 17 September 2025.