
The decision this quarter is not whether to keep using AI. It is whether to register the automated credit-decisioning engine, produce the file that defends it, and keep it in production — or stop automating decisions until someone can produce that file. The question that follows never reaches a slide: what must we put on the table, and by what date?
Vietnam answers part of that in law, and the answer is procedural: a classification, a file, a log, an override, a report inside a clock. Good news if you have run an audit. Bad news if you thought governance meant a policy PDF.
What changed
1 March 2026 — the AI Law took effect. Law 134/2025/QH15 (passed 10 December 2025, eight chapters, 35 articles) covers development, provision, deployment and use, and Article 33 repeals Chapter IV of the digital technology industry law (71/2025/QH15). Article 4 says AI does not replace human authority and responsibility — further, the source material notes, than the EU model.
15 August 2026 — the high-risk list began to bind, and banking is on it. Decision 33/2026/QĐ-TTg (issued 30 June 2026) covers six sectors. The banking examples are the two a consumer lender is likeliest to run: automated execution of electronic transactions, and automated credit-granting decisions. Press summaries count roughly 46 systems; read the annex.
1 May 2026 — classification became the institution's liability. Decree 142/2026/NĐ-CP (issued 30 April 2026) requires self-classification before use, portal notification for medium and high risk with an automatic identifier, re-notification within 15 working days if risk rises, and incident reports in 72 hours, 5 working days or 15 days by severity.
Since 1 March 2026 — the operator pays, and inspection asks for the file. 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. Article 28 requires the technical file, trace logs and training data.
1 October 2026 — no dedicated AI penalty decree, which is not the same as no enforcement. Decree 330/2026/NĐ-CP already fines personal-data processing in AI systems at VND 20–50 million; the draft figures — up to VND 2 billion for organisations, 1 billion for individuals — remain proposals.
1 September 2027 — the transition ends for finance. Systems running before 1 March 2026 have 18 months in finance, health and education, and 12 months elsewhere.
The decision framework
Four roles exist — developer, provider, deployer, user — and one organisation can hold several. Provider includes software licensing, cloud services, APIs and embedded components, so a model API vendor is a provider and the consuming bank a deployer; fine-tune the model and you are both. I score seven criteria, weighted for a lender whose credit decisions are in scope.
| # | Criterion | Weight | What "high" looks like |
|---|---|---|---|
| 1 | Inventory and boundary | 15% | Every system has an owner and a declared purpose |
| 2 | Classification file | 15% | Reasoning recorded, mapped to the Decision 33 annex, signed by an officer |
| 3 | Human oversight you can test | 20% | Logged override, measured intervention rate, a tested stop control |
| 4 | Technical file and traceability | 15% | Any decision reproducible: inputs, model version, thresholds, timestamp |
| 5 | Conformity readiness | 15% | Assessed before go-live; third-party cases identified; file maintained |
| 6 | Transparency and marking | 10% | Customers know they face a system; content carries machine-readable marking |
| 7 | Contract and recourse | 10% | Notice, audit, incident and indemnity duties; a Vietnamese contact point |
Two criteria are traps. On 3, a reviewer who approves 99% of outputs in two seconds is a signature, not oversight. On 5, the standards and the designated certifiers were still open points in November 2026 — you can prepare the file, not book the certificate.
Options and trade-offs
| Option | Strength | Risk | Three-year cost shape |
|---|---|---|---|
| Comply in place — register, classify, document what already runs | Fastest route to a register; no delivery freeze | Retrofit evidence; oversight claims may fail | Low in year one, steady run cost |
| Build the evidence layer — registry, AI gate, standard logs | Compliance becomes a reusable control | Needs product, legal and data to agree; a slow gate gets bypassed | Moderate in year one, falling cost |
| Replace or retire the high-risk decision systems | Removes the obligation at source | Expensive, loses decision speed already banked | High in year one, smallest surface after |
Most lenders need all three, applied per system.
The numbers
Effort is where estimates earn the label. A 2027 planning exercise I reviewed — anonymised, an estimate, not a quote — put the load at roughly 0.5 FTE in IT governance and architecture, 0.2 in legal and compliance, 0.2 in data: about one FTE across three functions, plus an optional VND 150–300 million for external legal review and conformity support. Certification was deferred a year, and in November 2026 nobody can price it: standards and certifiers are still pending.
What is law rather than estimate: 15 working days to notify a risk increase; 72 hours, 5 working days and 15 days for incidents; VND 20–50 million under Decree 330/2026/NĐ-CP; roughly 1 September 2027 for finance.
The measurable control targets: 100% of use cases registered with a risk level, zero unclassified production systems, 100% of high-risk systems assessed before go-live, 100% of IT staff trained.
What I would do in 30, 60 and 90 days
First 30 days — count systems, not vendors.
- Inventory every system, including embedded features and every model API in use. Shadow use is inventory.
- One owner per system; one sentence of boundary — what it decides, for whom, with what data.
- Start the classification file; the reasoning you record now is the evidence later.
- Open a contract gap list with vendors: notice, audit, incident reporting, indemnity, a contact point in Vietnam.
Days 31–60 — test one control, not seven.
- Override a decision on your riskiest system, then ask someone who did not build it to retrieve the override, the input, the model version and the time.
- Build the technical-file template from that exercise. If it cannot be filled for one decision, it will not be filled for a thousand.
- Table-top a Saturday incident: who reports what, to whom, inside 72 hours.
Days 61–90 — gate the lifecycle, price the rest.
- Add an AI gate to the project process — role, risk class, evidence owner — and retrofit it to live systems.
- Decide which systems you defend, rebuild or retire, in that order.
What I would not do
- I would not buy a governance platform before the evidence set is defined. The tool encodes whatever file you describe.
- I would not let a vendor's risk statement, or the Ministry's self-classification tool, become the classification. Both are inputs; liability stays with the institution.
- I would not keep the register outside delivery. If a use case can reach production without touching it, it is already wrong.
- I would not read the missing penalty decree as an absence of enforcement, or assume Article 27 stops at state agencies without a written opinion.
For the non-technical sponsor
The law does not ask whether your AI is good. It asks what you can show: that you classified the system and said why, that a person can overrule it and the overrule is recorded, that incidents are reported in days not months, and that the file survives an inspection. Fund one person across three functions, name an owner per system, and treat September 2027 as a delivery date.
Sources
- Law 134/2025/QH15 on Artificial Intelligence, Cổng thông tin điện tử Chính phủ
- Decree 142/2026/NĐ-CP detailing the AI Law
- Decision 33/2026/QĐ-TTg — list of high-risk AI systems
- Decree 330/2026/NĐ-CP — administrative penalties in cybersecurity
- "Danh mục hệ thống trí tuệ nhân tạo có rủi ro cao từ 15/8/2026", LuatVietnam
Personal views. Client and employer details are anonymised. Figures are third-party estimates, not vendor quotes, and are dated November 2026.
Comments
Join the discussionNo comments yet. Be the first to comment.



