# Start here: AI governance without blocking innovation

Every AI governance programme I have seen fail, failed the same way: it started with a policy and ended with a register nobody maintained. The ones that hold up start with a narrower question — what will we have to show, to whom, and by when — and build backwards from that.

Vietnam has made the answer specific. Law 134/2025/QH15 took effect on 1 March 2026, Decree 142/2026/NĐ-CP took effect on 1 May 2026, and Decision 33/2026/QĐ-TTg put the high-risk list into force on 15 August 2026. This series is six questions, answered one at a time, without a maturity model.

## The six questions

### 1. Which systems are actually in scope?

Classification is three-tiered and applies to complete systems, not to a model on its own. The list in Decision 33/2026/QĐ-TTg covers six sectors, and banking is one — with two examples that a consumer lender is likely to be running already: automated execution of electronic transactions, and automated credit-granting decisions. Embedded AI in purchased tools and model APIs are inside the boundary, not outside it.

### 2. Who is accountable when a decision is automated?

The law defines four roles — developer, provider, deployer, user — and one organisation can hold several. *Provider* includes cloud services, APIs and embedded components, so a vendor model is provided and the bank running it is deploying. Article 29 places compensation with the deployer where a high-risk system was lawfully used and damage still occurred, with recourse only where a contract provides it. An agent does not absorb liability; the institution that deployed it does.

### 3. What must be registered, and by when?

Medium and high-risk systems must be notified to the Ministry of Science and Technology through the one-stop portal before use, and the system receives an identifier automatically. If risk rises, notification is due within 15 working days. For systems running before 1 March 2026, obligations must be complete within 18 months in finance — around 1 September 2027 — and within 12 months elsewhere.

### 4. What does human oversight look like when it has to survive an audit?

Article 4 says AI does not replace human authority and responsibility, which is a stronger statement than requiring a reviewer. In practice, oversight means a logged override, a measured intervention rate, and a stop control that has been tested at least once. A committee that reviews monthly reports is governance theatre; an override nobody can retrieve afterwards is the same thing with better stationery.

### 5. What evidence has to exist before an incident, not after?

The clock starts when the incident does: a preliminary report within 72 hours for urgent incidents, 5 working days for other serious ones, and a formal report within 15 days, with data and system logs retained. Inspection under Article 28 can require the technical file, trace logs and training data. As of 1 October 2026 no dedicated AI penalty decree had been confirmed — but Decree 330/2026/NĐ-CP already fines personal-data processing in AI systems at VND 20–50 million, so the gap is not a holiday.

### 6. How do you do this without stopping delivery?

Put one AI gate in the project lifecycle: role, risk class, evidence owner. Classify at intake, reuse the same technical-file template, and let the sandbox handle genuinely novel solutions — Article 21 makes it voluntary, registered through the national public service portal, and its results are a basis for the authority to consider recognising conformity or reducing obligations. Completing a sandbox does not automatically earn either.

## What to read next

The long-form piece in this series is [What a bank must be able to prove about AI](/what-a-bank-must-prove-about-ai): a seven-criterion control map, three delivery options with their three-year cost shape, and a 30/60/90 plan. If your next deadline is a signed credit contract rather than a model, start instead with [E-signature is an evidence problem](/e-signature-is-an-evidence-problem).

## Sources

- [Law 134/2025/QH15](https://vanban.chinhphu.vn/?docid=216334&orggroupid=1&pageid=27160) — AI Law, in force 1 March 2026
- [Decree 142/2026/NĐ-CP](https://chinhphu.vn/?pageid=27160&docid=218029) — classification, notification, incidents
- [Decision 33/2026/QĐ-TTg](https://chinhphu.vn/?pageid=27160&docid=218658) — high-risk list, in force 15 August 2026
- [Decree 330/2026/NĐ-CP](https://xaydungchinhsach.chinhphu.vn/nghi-dinh-330-2026-nd-cp-ve-xu-phat-vi-pham-hanh-chinh-trong-linh-vuc-an-ninh-mang-119260824172446407.htm) — penalties
- [LuatVietnam](https://luatvietnam.vn/tin-van-ban-moi/danh-muc-he-thong-tri-tue-nhan-tao-co-rui-ro-cao-tu-15-8-2026-186-110045-article.html) — high-risk list summary

---

*Personal views. Client and employer details are anonymised. Figures are third-party estimates, not vendor quotes, and are dated November 2026.*
