How Vietnam law treats your company AI tools
Tech Ops Asia (Công ty TNHH Tech Ops Asia) is a Da Nang tech and cyber consultancy. We map where personal data and AI tools actually move, then give you a fix queue. We are not a law firm. Confirm filings and legal basis with Vietnam-licensed counsel before you act.
Clients keep asking the same thing after our Decree 330 briefing: what does this mean for the AI already in the stack? This is that answer, in plain English.
What laws cover company AI in Vietnam?
Three instruments, not one. Most teams only hear about the fine decree. The duties sit in the other two.
| Layer | Instrument | What it covers for AI |
|---|---|---|
| Personal data in the tool | Law 91/2025/QH15 (PDPL), from 1 Jan 2026, plus Decree 356 | Article 30 is the AI article: only process what you need, lock access, classify risk, do not use personal data in AI to harm people. Consent, transfers, and impact files still apply. |
| The AI system itself | Law 134/2025/QH15 (AI Law), from 1 Mar 2026, plus Decree 142 from 1 May 2026 | If the company uses an AI system in its work, it is usually a deployer. Risk class, human control, labelling, and incident duties sit here. |
| What happens if the data piece fails | Decree 330/2026/ND-CP, from 19 Aug 2026 | Fines, deletion orders, and the power to stop processing or overseas transfers. This is the enforcement layer for personal-data misses, including misses that happen inside an AI tool. |
The AI Law does not replace the PDPL. ICLG’s 2026 Vietnam chapter says the same thing in lawyer language: the AI Law manages the system; personal data in that system still follows the PDPL. If a provider uses personal data in an AI system, Decree 142 even lets them fold a PDPL impact file into the AI risk dossier so they do not write the same story twice.
Does this apply if we only use ChatGPT or Gemini?
Yes, for almost every company that uses those tools at work on real people. The brand on the tab does not decide the law. The data and the job do.
The AI Law (Article 3) splits people into four roles:
- Developer: builds or fine-tunes the model and controls training data or parameters.
- Provider: puts the system on the market under their own name, even if someone else built it.
- Deployer: uses an AI system under their control in professional, commercial, or service work. Home use does not count.
- User: talks to the system or uses its output.
A staff member pasting a customer email into ChatGPT is a user. The company that bought the seats, set the policy, or looked the other way is a deployer. OpenAI or Google is the provider of that public model. If you wrap the model in your own product and sell it as yours, you can become a provider too.
You do not need to train a model to be in scope. “AI system” in the law is any machine that infers from inputs and produces content, predictions, recommendations, or decisions. A helpdesk bot, an HR screener, a coding assistant on company source, and a Zalo auto-reply all sit in that definition.
What does the personal data law require for AI?
Article 30 of the PDPL is short and it is already in force. English translations (LuatVietnam, VN Law Firm) say the same five things.
- Process personal data in AI only for a real purpose, and only as much as you need.
- Follow the rest of the PDPL and other laws, plus Vietnamese ethical standards.
- Build in confidentiality, login, identity checks, and access rights. A shared personal ChatGPT login is the opposite of that.
- Classify AI processing of personal data by risk, then match the protection to that risk.
- Do not use personal data in AI to harm defence, security, social order, or another person’s life, health, honour, dignity, or property.
The rest of the PDPL still applies on top. Consent must be specific to each purpose. Silence is not consent. If you collected a phone number for delivery, that is not permission to feed the same file into a marketing AI. People can ask for access or deletion. You need a clock for that. If the tool is outside Vietnam, you also have a transfer question (next section).
One extra trap if you sell AI that processes other companies’ personal data: Decree 356 treats some automated personal-data processing on AI as a licensed service. ICLG and Chambers both list “automated personal data processing based on big data, AI, blockchain or virtual reality” as a service that can need a Certificate of Eligibility. That is not the same as a café using ChatGPT to draft a menu. If you sell scoring, scraping, or AI processing of other people’s data, stop and talk to counsel.
What does the AI Law require if we are a deployer?
You inherit the provider’s risk class, you keep the system safe in use, and you do not quietly turn a low-risk tool into a high-risk one.
Decree 142 and the AI Law use three classes:
| Class | Plain meaning | What a company usually owes |
|---|---|---|
| High risk | Can seriously affect life, health, rights, public interest, or national security. Health, education, and finance sit under extra care. | Human oversight, security, logs, transparency, incident handling, and the ability to explain the system to the state. Providers also do conformity assessment. A hiring or credit bot that decides people’s outcomes is the everyday version of this class. |
| Medium risk | People may not realise they are talking to AI, or AI-made content can confuse them. | Label it. Tell people when a bot is a bot. Mark AI audio, image, or video that could pass as real. Providers notify the Ministry of Science and Technology before they switch the system on. |
| Low risk | Everything else. | You can still be asked to explain the system if something goes wrong. You still owe the PDPL if personal data is inside it. |
Article 4 of the AI Law is the sentence boards actually need: AI must not replace human authority and responsibility. Keep a person in the loop for decisions that affect staff or customers. A model that auto-rejects CVs with no human review is the kind of design the law is written against.
Systems that were already running before 1 March 2026 get a catch-up window (Article 35): 18 months for health, education, and finance, 12 months for other fields, unless the state orders a stop because the system can cause serious harm. “Catch-up” is not “ignore.” New systems do not get that holiday.
Is pasting staff or customer data into ChatGPT a cross-border transfer?
Often yes. Treat it as a transfer until counsel maps the actual path and says otherwise.
PDPL Article 20 lists three transfer cases. The third one is the ChatGPT case: a person in Vietnam (or overseas) uses a platform outside Vietnam to process personal data collected in Vietnam. Public ChatGPT, Gemini, Claude, and most vendor copilots sit on servers outside Vietnam. That is the same family of risk as Google Workspace in Singapore, which we already covered in the Decree 330 post.
There is a real exemption for some employee data stored on cloud services (Article 20.6.b). That is about storing HR files on a cloud drive. It is not a free pass to paste those files into a public chatbot. Do not stretch the exemption yourself.
If it is a transfer, the PDPL wants a transfer-impact file within 60 days of the first transfer, unless an exemption applies, and updates when the stack changes. Decree 330 is what turns a missing file, or a leak on that path, into a fine or a freeze. Serious cross-border failures that cause a leak can reach a percentage of last year’s Vietnam-market revenue. That is why the AI paste is not a side note.
How does this show up on a normal Tuesday in Da Nang?
This is the map we use with companies here. A privacy policy never reaches these rows.
| Everyday AI use | How the law can land |
|---|---|
| Staff paste a customer email or a CRM export into a free ChatGPT or Gemini tab | Processing plus, usually, an overseas transfer. No access control. No log of what left. Hard to honour a deletion request. |
| HR drops CVs or a biometric timeclock export into an AI screener | Employee data is still personal data. Biometrics are sensitive. An automated reject with no human in the loop fights the AI Law’s human-control rule. |
| A Zalo or website bot answers guests without saying it is AI | Medium-risk transparency duty: people should know they are talking to a system. The chat log is also personal data under the PDPL. |
| A founder turns on “train on our data” on a vendor AI | You may have given a foreign provider a new processing purpose. Consent and the contract are probably both wrong. |
| You ship a product that scores or scrapes other companies’ personal data with AI | That can be a licensed personal-data processing service, not only an AI-Law deployer problem. |
The pattern is the same as the rest of Decree 330. The fine is not always the thing that stops the business. An order to delete a dataset you cannot rebuild, or to stop sending data to the overseas model, can pause support or hiring while the lawyers argue.
How should we treat our AI systems this month?
Start with the tools people already use, not a 40-page AI policy.
- List every AI tool in the company, including free personal accounts on work laptops, Gemini inside Workspace, GitHub Copilot, and the Zalo bot.
- Mark which of those ever see a name, phone, ID, CV, health note, or customer file.
- Turn off training on your data where the vendor lets you. Block the rest with a written rule and, if you can, a technical control.
- Give staff one approved tool with a company login. Kill shared personal accounts for work data.
- Write one sentence for each tool: purpose, who may paste what, and who is the human who checks the output.
- If a bot talks to customers, label it as AI.
- If the tool is outside Vietnam and it sees personal data, add it to the transfer map you already owe under the PDPL.
- Ask counsel whether the DPIA and transfer files need an update because you added AI after 1 January 2026. Article 22 expects updates when the processing changes.
Want a systems read of the AI already in the stack?
We inventory the tools staff already use, mark personal-data paths, and give you a two-week fix queue. Bring your counsel if you have one. We will not pretend to be them.
Who should I hire for this?
Hire the person who can close the gap you actually have. Paper is a legal product. Logs and access are a systems product.
| Option | What they actually do |
|---|---|
| Vietnam-licensed law firm | Legal opinions, DPIA and transfer dossiers, and the line between “internal ChatGPT” and a licensed processing service. They will not sit in your Workspace and turn a policy into logs. Best when you need a filing or a disputed legal basis. Limit: many health checks stop at documents. |
| Tech Ops Asia, systems workshop (Da Nang) | An inventory of the AI already in use, a personal-data path map, access and retention that actually run, and a two-week backlog, in English or Vietnamese. Best when the company cannot yet list its tools. Limit: we do not file A05 dossiers or tell you that you are compliant. |
| Owner or IT lead, DIY | Fine for a tiny shop that only uses AI to draft a Facebook caption with no customer list in the prompt. Limit: the moment you have a CRM, HR files, or a parent company, DIY usually produces a rule nobody follows. |
FAQ
Does Vietnam law cover ChatGPT or Gemini used at work?
Yes. If the tool processes data that identifies a person, the PDPL applies. If the company uses the tool in its work, the AI Law usually treats the company as a deployer.
Is pasting customer data into ChatGPT a cross-border transfer?
Often yes. Using a platform outside Vietnam to process personal data collected here is a listed transfer case. Ask counsel before you assume an exemption.
What is the difference between the PDPL and the AI Law?
The PDPL and Decree 330 govern the personal data. The AI Law and Decree 142 govern the system: risk class, deployer duties, labels, and incidents. Most companies sit under both.
Do we need a licence just to use AI internally?
Usually no. The Certificate of Eligibility is aimed at people who sell personal-data processing as a service. Internal use is still processing. Counsel draws the line if you sell scoring or scraping.
Can a workshop make our AI stack compliant?
No. It can stop you guessing. A one-day “compliant” stamp is not a thing anyone can honestly sell. Tech Ops Asia does the systems side from Da Nang.
Sources
- VN Law Firm: unofficial English of Law 91/2025/QH15 (Article 30 on AI; Article 20 on transfers)
- LuatVietnam: Law 91/2025/QH15, Article 30
- DFDL: Vietnam personal data protection 2026 (duties layer from 1 January 2026)
- Chambers: new PDPL framework (sector rules for AI and cloud)
- ICLG Data Protection 2026, Vietnam (Article 30 risk class; licensed AI processing services; AI Law vs PDPL)
- VN Law Firm: unofficial English of Law 134/2025/QH15 (roles, risk class, deployer duties, Article 35 timelines)
- Tilleke & Gibbins: Decree 142/2026/ND-CP
- GVW: Vietnam AI Law and Decree 142 overview
- Viet An Law: Decree 142 risk classification
- Government portal: Nghị định 330/2026/NĐ-CP
Related reading
- Part 1: Decree 330 personal-data fines
- How teams actually buy ChatGPT and Claude subscriptions
- Part 3: lawyer, workshop, or both (next)
- Tech Ops Asia: compliance and AI-tech assessment