Skip to content
ContestabilityA research
publication

Research note / 2026-11-17

Vietnam’s bank-AI circular promises a human review of complaints. What will the reviewer have in hand?

What takes effect

On Sunday 15 November 2026, Circular 53/2026/TT-NHNN of the State Bank of Vietnam takes effect. It was issued on 30 September and sets safety, risk and deployment conditions for AI systems used by credit institutions, payment intermediaries, mobile-money providers and credit-information companies. I read the full Vietnamese text on a legal database; the translations below are mine, and I have not yet seen the official gazette copy.

Three provisions bear on contestability. Article 20(4) requires each institution to run a complaints process in which an automated decision that is complained about is re-assessed by a human, and that human result becomes the final basis for resolving the complaint. Article 20(5) says that where an automated decision directly affects a customer’s legal liability, financial obligations or access to a service, the institution must keep the basis and main factors behind it, to provide and explain on request. That explanation need not reveal source code, detailed algorithms, parameters or trade secrets. Article 5(2) requires the oversight function to be able to override AI decisions, with each intervention logged by content, time, person and reason. Systems already in operation have until 1 March 2027 to comply.

What I make of it

This is commentary, not the circular’s own words. A record kept when the decision is made, explained once the person asks, resembles what Paper A calls contemporaneous preservation and inquiry-activated access. The override-and-log duty speaks to reviewer authority, one of the conditions I score highest.

The text leaves open the things a complainant would need most. As I read it, it sets no deadline for the review. It does not say the reviewer must be independent of whoever made or configured the original decision. The training duty in Article 5(3) is framed around oversight staff, not complaint handlers. The explanation is “on request” and limited to main factors, and the text does not say whether the person can see the stored record. It does not say what happens to the decision, or to the record, when the human disagrees with the system.

A proposed illustration

This is a hypothetical, not a case. A customer applies through a bank’s app for a loan of 60 million dong, above the 50 million threshold at which Article 4(2)(b) treats an automated credit system as high-risk. The system declines. She complains and receives a human re-assessment, and asks for the main factors. The answer is insufficient repayment capacity, because the system counted a settled account as outstanding. She can contest that data point. Whether she can contest the way repayment capacity is measured depends on whether “main factors” reaches the standard applied, which the text does not settle. If the reviewer sees only the system’s output, the review may repeat the error. Article 10’s logs would let a reviewer test the decision, but the circular does not say the reviewer must use them.

What the source does not say

It does not report any complaint handled under the rule, since the rule has not yet taken effect. My reading of Article 20 as covering AI decisions generally, not only high-risk ones, rests on its wording and should be checked by a Vietnamese lawyer.

If a human review is to be the final word, what must the reviewer be given for the word to be worth waiting for?