HANOI, 14 AUG 2026 — Tomorrow, Vietnam's list of high-risk AI systems takes effect. Anything on it faces mandatory conformity assessment before deployment, a local presence requirement for foreign providers, and strict liability for damages even where the provider has complied fully.
That last clause is the one to read twice. It is stricter than the EU AI Act on the point where regulators usually flinch.
What comes into force
Decision 33/2026/QD-TTg, issued by the prime minister on 30 June, establishes the List of High-Risk AI Systems under Vietnam's AI Law and Decree 142/2026/ND-CP. It takes effect on 15 August 2026.
Its practical function is to narrow the law. Only systems on this list carry the heightened obligations, so Decision 33 is what turns a framework statute into a compliance question with a defined perimeter.
| Sector | What is captured |
|---|---|
| Education | Automated assessment, learner ranking, behavioural monitoring, and content generation from uncontrolled sources. |
| Healthcare | AI-assisted surgical systems and autonomous surgical robots. |
| Banking | Systems conducting electronic transactions or making credit decisions autonomously. |
| Transport | 31 categories, including autonomous driving, traffic signalling and critical infrastructure control. |
| Judicial proceedings | Large-scale biometric identification in civil cases. |
| Ethnic and religious affairs | Scoring applications, approving or rejecting regulatory submissions, or classifying individuals by ethnicity or religion. |
The sentence that decides everything
Many of these systems qualify as high-risk only where AI outputs are used as the sole basis for decisions without meaningful human review.
This phrase is the core of the instrument, creating both a compliance path and a potential loophole.
On one hand, this looks like good regulatory design. It targets the actual harm — a machine deciding alone — and gives operators a clear route to lower risk: keep a person in the loop who can actually change the outcome.
On the other, it is an escape hatch with a well-worn shape. "Meaningful human review" has been the softest phrase in European data protection law for a decade, because a reviewer who approves ninety-nine per cent of recommendations in eleven seconds satisfies the letter of it. Vietnam has not published a definition, and every compliance strategy in the country will now be built around whatever that phrase turns out to mean.
Strict liability is the sharp end
But the real sting is the liability for damages, which applies even when a provider is fully compliant.
Most technology regulation makes compliance a defence. Do the assessment, keep the documentation, follow the process, and if something goes wrong you argue that you met the standard. Vietnam's framework, as reported, does not offer that trade.
If that reading holds, it changes the calculation for anyone deploying a listed system. Compliance stops being insurance and becomes an entry ticket — necessary to operate, insufficient to protect. The exposure then has to be priced or transferred rather than argued away, which is an insurance question before it is a legal one, and the insurance market for AI liability in this region is thin.
This reading comes with caveats: we are working from legal summaries rather than the statutory text, no liability cap or dispute mechanism has been described in what we could read, and "strict liability" carries different weight across jurisdictions. Anyone with real exposure should read Decree 142 itself and take Vietnamese advice. But the direction is unambiguous and it is not the European direction.
The dates that matter
Note the scope of that last row. Registration applies to all AI systems operating in Vietnam, not only high-risk ones. The list determines who faces conformity assessment and heightened duties; the registry catches everybody.
How this differs from the European template
Vietnam's structure will look familiar to anyone who has read the EU AI Act; the differences are what make it interesting.
The shared architecture is real: a risk tier defined by a published list, conformity assessment before a high-risk system reaches the market, documented risk management and data governance, human oversight requirements, and registration in a central database. Vietnam has adopted the European shape.
Three differences stand out.
First, speed. The EU's obligations phase in over years from a 2024 act; Vietnam passed a standalone AI law and had a sectoral list in force inside a much shorter window, with compliance dates in 2027. Whether this pace proves decisive or merely hasty depends on whether the country has the assessment capacity to match.
Next is the local presence requirement for foreign providers, which has no direct European equivalent for AI systems and reads more like a data-localisation instinct than a safety one. The EU requires an authorised representative; requiring presence is a heavier ask and a different kind of policy.
The final and sharpest divergence is liability. Europe's approach leaves compliance meaningful as a defence, and its attempt at a dedicated AI liability directive was abandoned. Vietnam appears to have gone the other way. If the strict-liability reading is correct, an operator in Ho Chi Minh City carries risk that the same operator in Frankfurt would not, on the same system, having done the same work.
Why this matters beyond Vietnam
Vietnam is the first country in Southeast Asia with a binding, standalone AI law, and the region's other regulators are watching what happens next rather than starting from scratch.
Singapore has governed AI through frameworks and voluntary testing regimes. Indonesia has moved on platform and child-safety rules and is building capacity rather than legislating models. Malaysia has paired online-safety enforcement with a sovereign AI cloud. None has done what Vietnam has just done, which is to publish a list, attach hard obligations to it and set dates.
Whichever way this goes will shape the regional template. If a defined list with fixed deadlines produces workable compliance, it becomes the model. If it produces an unworkable registry and a conformity-assessment queue with no assessors, it becomes the cautionary tale.
For foreign providers, the local presence requirement has the sharpest commercial edge. What constitutes presence has not been defined in the material we could read, and the answer determines whether a global vendor can serve Vietnamese customers from Singapore or must incorporate, staff and take on liability locally.
What to do about it
If you operate AI in Vietnam, the immediate work is a mapping exercise rather than a legal one. Establish which of your systems touch the six listed sectors, and for each one determine whether its output is the sole basis for a decision. That single question separates a registration obligation from a conformity-assessment programme.
Then check the healthcare, education and finance carve-out. If a system in those sectors was already operating before tomorrow, it has until September 2027 rather than March. That is six extra months, and it is only available to systems live before the effective date — which makes tomorrow a meaningful line for anything close to launch.
And start the documentation now regardless. Risk management and data-quality records are the slowest part of any conformity regime, they cannot be produced retrospectively with any credibility, and March 2027 is closer than it reads.
What to watch
Whether "meaningful human review" gets a definition, and how demanding it is. Everything else in this instrument bends around that phrase.
Watch too for whether the conformity-assessment capacity exists. A mandatory pre-market assessment regime needs accredited assessors in sufficient number, and standing one up takes longer than passing the rule that requires it.
Finally, penalties. None has been described in what we could read, and a compliance deadline without a stated consequence is a deadline that firms will price accordingly.
