The State Bank of Vietnam has issued Circular 53/2026/TT-NHNN setting safety, risk management and deployment conditions for artificial intelligence in banking, Cafef and Banking Times report. It covers credit institutions, foreign bank branches, payment intermediaries, credit information companies and mobile money providers, and takes effect on November 15, 2026.

The central principle is that AI does not replace human authority or responsibility. The institution stays answerable before the law and before customers for any decision made or supported by an AI system. It cannot cite the technical character of a model, its autonomy, or work done by a vendor to reduce that responsibility.

Banks must classify systems as high, medium or low risk before deployment. Automated payment execution becomes high risk when a single retail transaction exceeds 10 million dong or retail orders in a day pass 20 million dong, with thresholds of 50 million dong per transaction or 100 million dong a day for organizations. AI that proposes credit decisions is also high risk when the loan is 50 million dong or more.

Volume caps sit on top of the classification. Daily automated transactions cannot exceed the provider's internal limit or 100 million dong for individuals and 500 million dong for organizations. Customers must agree in advance to any autonomous payment scope, with stated limits and the right to end the agreement at any time, and they must have tools to pause or revoke it. Supervisory staff must hold the authority to override, adjust or suspend AI decisions and to trigger an emergency stop, and every intervention must be logged with content, time, person and reason.

Systems already running get a grace period. Operators must bring AI live before the effective date into compliance with the safety, risk and deployment conditions by March 1, 2027, while parts of the risk management framework, security checks and staffing rules follow a phased path. Banks must also tell customers at the start of a session when they are speaking with an AI, and a human review stays the final basis for settling complaints against automated decisions.