Vietnam Sets Limits and Customer Safeguards for AI Bank Payments From November 15

AI payments face daily caps and stronger human control

Vietnam will cap payments automatically executed by artificial intelligence at VND 100 million a day for an individual customer and VND 500 million for an organisation from November 15, 2026. Banks and other providers must first obtain customer permission, offer controls to withdraw it and arrange human review when an automated decision is challenged.

The State Bank of Vietnam issued Circular 53/2026/TT-NHNN on September 30, 2026, establishing safety, risk management and deployment requirements for banking AI. Its rules combine payment limits with requirements covering identity checks, explanations, cybersecurity, data protection and responsibility when something goes wrong.

Article 4 brings stronger oversight much earlier than the daily caps. An AI system automatically making a personal payment above VND 10 million, or payments above VND 20 million for one person in a day, enters the high risk category. For organisations, those thresholds are above VND 50 million per payment or above VND 100 million daily.

Customers who authorise AI to move money therefore face two different controls: thresholds that trigger closer supervision and ceilings that set how much the system may pay each day. Customers making transfers themselves through banking apps remain subject to the rules for those services.

The circular builds on the payment caps and safeguards outlined in October. Its practical requirements reach beyond transfer amounts: institutions must keep people able to intervene, explain consequential decisions and remain accountable even when an outside company supplies the technology.

Which amounts trigger closer supervision?

Article 4 requires providers to assess both individual payments and the accumulated daily total for each customer. Crossing either threshold brings the relevant automated payment system into the high risk category.

• Individuals: above VND 10 million for one payment, or above VND 20 million in total for one day.
• Organisations: above VND 50 million for one payment, or above VND 100 million in total for one day.
• Automated credit decisions: VND 50 million or more.

Three automatic personal payments of VND 8 million would total VND 24 million. Each payment would be below the individual transaction threshold, while the daily total would cross the VND 20 million threshold. Providers must therefore track the running total as well as the size of each transfer.

The wording at the boundary matters. Payment thresholds apply to amounts above the stated figures. The credit threshold includes exactly VND 50 million. An AI system automatically deciding whether to grant credit of that amount falls within the specified high risk category.

Before deployment, institutions must classify their AI systems as high, medium or low risk and obtain approval of the classification from the appropriate authority. They must maintain an inventory of systems and review and update it at least annually. The classification determines how much scrutiny and control a system needs.

Daily ceilings set a separate limit

Article 5 limits the total value of payments that AI may automatically execute for each customer in a day. The maximum is VND 100 million for an individual and VND 500 million for an organisation, subject to the provider's internal limit.

A bank can set a lower allowance. If its internal daily limit for a personal customer is VND 60 million, that lower figure governs the customer's AI payments. The national ceiling gives institutions an upper boundary rather than requiring them to make the full amount available.

Both national daily ceilings are five times their corresponding daily risk thresholds: VND 100 million compared with VND 20 million for individuals, and VND 500 million compared with VND 100 million for organisations. A system can therefore require stronger supervision while still operating within its permitted payment allowance.

The thresholds and review rights detailed in November form part of the same framework. Providers need to monitor permission, transaction amounts, daily totals and system behaviour together, rather than treating the daily ceiling as their only control.

Customers must authorise AI before it pays

A bank or payment provider must establish an agreement with the customer before allowing AI to execute payments automatically. The agreement must define the permitted transactions, limits for each transaction and each day, its duration, applicable conditions and how the arrangement can be suspended or ended.

Providers must verify that the person establishing the agreement is the account holder or a legally authorised representative. For electronic agreements, biometric verification must be combined with at least one additional authentication method under Circular 50/2024/TT-NHNN, as amended by Circular 77/2025/TT-NHNN.

Biometric verification checks a physical characteristic associated with a person, such as their face. The additional authentication requirement adds another check when the customer gives a system authority to move money.

Within the scope of an agreement confirmed in this way, certain transaction confirmation requirements under Circular 50 need not be applied to each automatic payment. The initial authorisation is therefore a consequential step: it defines what the AI can do without returning to the customer for those confirmations.

Customers may terminate the agreement at any time, and providers must act immediately on that request. They must also supply tools allowing customers to pause or revoke AI payment authority themselves. Before signing, customers should check the permitted payments, limits, expiry and suspension terms, and use official banking apps.

Customers can seek explanations and human review

Banks must tell customers at the beginning of a session or conversation when AI replaces a person in direct communication. That disclosure lets customers know whether they are dealing with a human representative or an automated system from the outset.

Providers must retain the basis and key factors behind automated decisions affecting a customer's legal responsibilities, financial obligations or access to products and services. Customers can request an explanation of those factors.

An explanation need not disclose source code, detailed algorithms, model parameters or trade secrets. The requirement focuses on the basis for the decision affecting the customer. For decisions involving fraud prevention, money laundering, terrorist financing or cybersecurity, the reasons disclosed must be consistent with risk controls and protection of customer assets.

Institutions must establish departments and procedures for receiving and handling AI complaints. If a customer challenges an automated decision, a person must review it, and the human review becomes the final basis for resolving the complaint.

The State Bank of Vietnam states the governing principle in the circular:

AI does not replace human authority and responsibility.

For customers, this means the provider remains answerable for decisions made or supported by its systems. A complaint must reach a process with human responsibility, rather than ending with the software's original answer.

Staff need authority to stop a system

High risk AI requires continuous monitoring and people with enough authority and technical capability to intervene. Supervisors must be able to override, adjust or pause decisions, and activate an emergency mechanism to stop or isolate a system.

Institutions must record each intervention, including what changed, when it happened, who acted and why. Those records create a traceable account of the human response when a system behaves unexpectedly or a decision needs correction.

Before a high risk system enters operation, the institution must assess its purpose, scope, operation and possible effects on banking activities, cybersecurity, customers and other affected parties. The assessment must address personal data processing, safeguards, monitoring and incident response. It must be approved, retained throughout operation and updated when the system changes.

The institution's senior governing body has the highest responsibility for AI use and must approve both the proposed deployment and entry into operation of high risk systems. Providers must organise risk management through three independent lines of defence, separating operational responsibilities, risk oversight and independent assurance.

For systems classified as high or medium risk, institutions must report on operation, risks, serious incidents, complaints and trends in bias at least once every six months. Serious incidents require an immediate response, including emergency shutdown and a switch to manual operation when necessary, measures to limit harm, regulatory reporting and notification of affected customers as required.

Data quality and outside suppliers come under scrutiny

AI systems must undergo cybersecurity assessments before entering operation, periodically afterward and when substantial changes occur. Institutions must monitor unusual outputs, deteriorating performance, bias and changes in the model or the data it receives.

Data used for training, testing and operation must have clear, lawful origins. Providers must check its quality, address inaccurate or manipulated information and prevent leaks. Data poisoning means deliberately corrupting information so that an AI system learns or produces harmful results.

Model drift and data drift are also part of the monitoring requirements. In simple terms, a system's behaviour or the information reaching it can change over time, making earlier testing less reliable as a guide to current performance.

Institutions must manage model versions, configurations and significant changes, recording when a version is introduced, changed or retired. Operational logs must capture relevant inputs, outputs, model versions, unusual results, errors and security warnings. Major changes require renewed assessment and updated guidance, response scenarios and backup arrangements before implementation.

Outside suppliers may not use an institution's data to train, fine tune or improve their models without written consent. Banks must assess supplier capabilities and regularly monitor risks arising from dependence on third parties.

The institution using an outsourced system remains responsible for operational safety, customer rights, cybersecurity and personal data protection. If that system causes harm, the institution must take corrective action, handle complaints and compensate affected parties as required by law.

Digital growth makes the controls consequential

Nearly 89% of Vietnam's population aged 15 and above had payment accounts by the end of 2025, according to the central bank. Digital transactions accounted for more than 95% of transactions at many credit institutions.

Cashless payment volumes rose 34.28% in the first half of 2026 compared with the same period a year earlier, while their value rose 12.24%. Transaction numbers were growing faster than the total amount of money moving through those payments, increasing the importance of controls that can follow large numbers of transactions.

Vietcombank launched its One Data Platform project in August to support advanced analytics and prepare for future AI and machine learning applications. TPBank General Director Nguyen Hung has said the bank combines AI, large datasets and analysis of transactions as they happen to identify potential risks before losses occur.

At the Banking Sector Digital Transformation Forum on August 18, 2026, State Bank Governor Pham Duc An called for focused and controlled innovation amid challenges involving cybersecurity, digital fraud, privacy and AI governance.

Ly Hoang Vu, an economist at the French Vietnamese Center for Management Education, known as CFVG, has warned that AI can reproduce voices and faces, making remote verification harder. He advocates independent verification, multiple layers of control and detection of unusual situations.

A 2024 fraud at Arup Group Limited's Hong Kong office illustrates the financial stakes. Arup confirmed losses of about HKD 200 million, equivalent to USD 25.6 million, after a finance employee transferred money following a video meeting featuring fabricated images of the chief financial officer and colleagues. The case gives concrete context to requirements for identity checks, intervention and institutional responsibility.

Who must comply, and when?

The circular applies to credit institutions, foreign bank branches, payment intermediaries, mobile money providers and credit information companies. Its reach also includes the Vietnam Asset Management Company, the National Banknote Printing Plant and Deposit Insurance of Vietnam.

The framework covers AI used to make decisions or support decisions in customer services across banking, payment intermediary services, credit information and mobile money. It sits within requirements on AI, cybersecurity, data and personal data protection, with references to the Law on Artificial Intelligence, No. 134/2025/QH15, and Government Decree 142/2026/ND-CP.

• September 30, 2026: The State Bank of Vietnam issued Circular 53/2026/TT-NHNN.
• November 15, 2026: The circular takes effect.
• March 1, 2027: AI systems operating before the effective date must meet the new safety, risk management and deployment requirements.

Some provisions covering risk management frameworks, cybersecurity audits and human resources follow a separate implementation roadmap. Specific dates for those stages were not given in the descriptions of the circular.

Banks introducing systems must build authorisation, monitoring and human intervention into deployment. Institutions with existing systems face the March deadline for bringing those operations into compliance. Customers gain practical rights to control payment authority, request explanations and challenge automated decisions.

Key Points

• Circular 53/2026/TT-NHNN takes effect on November 15, 2026.
• Personal AI payments above VND 10 million per transaction or VND 20 million daily trigger high risk classification.
• Organisational thresholds are above VND 50 million per payment or VND 100 million daily.
• AI payment totals are capped at VND 100 million daily for individuals and VND 500 million for organisations, subject to lower internal limits.
• Customers must agree before AI pays, and can pause or revoke that authority.
• Challenged automated decisions require human review, while providers remain responsible for systems supplied by outside companies.
• Existing AI systems must meet the new requirements by March 1, 2027.