SBV announced Circular 53/2026/TT-NHNN on 7 October. Signed on 30 September, the banking AI safety and risk-management framework takes effect on 15 November 2026. The release emphasises human oversight and intervention, with AI not replacing human authority or responsibility.
Event / source publication date: 30 September 2026The following is Human Fintech’s analysis
An escalation path has to lead somewhere.
For readers following digital lending across markets, Vietnam provides a concrete reason to examine what human control means in practice. A lender may offer an escalation button yet still leave its staff unable to investigate an automated outcome. The customer sees a route to review; the operator sees only a score. Those are very different levels of control.
Consider a hypothetical applicant whose employment details were recorded incorrectly. Re-running the model without correcting the input is unlikely to resolve the underlying problem. A useful review would distinguish a data error from a valid credit judgement. It would also avoid treating a complaint as an instruction to approve. This is our operational interpretation, not a claim that the release prescribes this particular workflow.
Look across the decision chain.
An application can pass through document processing, fraud screening, risk assessment and offer generation before the customer receives an answer. A problem introduced early may appear to be a sound decision at the end. Measuring the final model alone can therefore miss an important source of customer harm. The relevant unit of analysis is the complete decision and the evidence supporting it.
This also changes how an institution should think about accountability when it uses external technology. Identifying a supplier does not, by itself, identify who can resolve an individual case. Staff need a practical way to locate the affected stage, establish which information was used and decide whether a correction requires a new assessment. The discussion is about service design; legal allocation of obligations requires the underlying rules and contracts.
Judge control by a disputed case.
A useful reader’s question is whether a service can explain what happens after an automated answer is challenged. Can the applicant submit corrected information without starting an unrelated request? Does the reviewer retain the original evidence? Is there a clear distinction between acknowledging the complaint and completing the review? These details reveal more than a generic assurance that a human remains involved.
For organisations comparing markets, the Vietnamese announcement should not be treated as an international rule or as proof that a particular product is compliant. It is a prompt to examine whether fast decisions remain accountable decisions. This article uses the official announcement rather than offering a comprehensive legal interpretation of the circular. The broader lesson is that convenience at the front of a service needs a credible route to resolution when something goes wrong.
White Fintech.
Parts II–III of White Fintech cover lending economics, funding and risk management: a foundation for further reading about the lending business.
Complete English edition, PDF + EPUB. Read the first 30 pages free, with no sign-up.