Vietnam sets AI banking risk thresholds, payment caps and rights to human review

Automated payments face tighter controls from November 15

Vietnam's new banking AI rules classify systems that automatically make personal payments above VND10 million per transaction, or above VND20 million for one customer in a day, as carrying high risk. For organisational customers, the corresponding thresholds are above VND50 million per payment or above VND100 million a day.

The State Bank of Vietnam's Circular 53/2026/TT-NHNN, signed on September 30, 2026, takes effect on November 15, with some requirements introduced later. Systems automatically deciding whether to extend credit of VND50 million or more also enter the high risk category, bringing continuous monitoring and stronger controls over human intervention.

Customers gain specific protections alongside those thresholds. Banks and payment providers must obtain agreement before allowing AI to make automatic payments, provide tools to withdraw that authority, and arrange human review when customers challenge automated decisions. Providers remain responsible for decisions made or supported by their AI systems.

The circular reaches beyond banks. It covers foreign bank branches, payment intermediaries such as electronic wallet providers, Mobile Money providers and credit information companies. Its scope also includes Vietnam Asset Management Company, the National Banknote Printing Plant and Deposit Insurance of Vietnam.

Which amounts trigger the high risk category?

Article 4 sets separate tests for individual payments and accumulated daily payments. Crossing either threshold brings the relevant automated payment system into the high risk category. Several smaller transfers can therefore trigger stronger oversight even when no individual transfer exceeds the limit.

For example, three automatic personal payments of VND8 million each total VND24 million. Each payment is below VND10 million, but their combined value exceeds the daily threshold of VND20 million. A provider must monitor both the amount of each payment and the running total for that customer.

The wording matters at the boundary. The payment thresholds apply to amounts above the stated figures. The credit threshold starts at VND50 million, so an automated decision on credit of exactly that amount falls within the stated high risk category.

Before deployment, providers must classify AI systems as high, medium or low risk. The circular develops banking requirements within the wider framework of the Law on Artificial Intelligence, No. 134/2025/QH15, and references the Prime Minister's Decision 33/2026/QD-TTg on systems carrying high risk.

Daily payment ceilings serve a different purpose

Article 5 sets an upper limit on the total value of payments that AI may automatically execute for one customer in a day: VND100 million for an individual and VND500 million for an organisation. A provider's internal limit also applies, meaning the actual allowance may be lower.

The risk thresholds determine the level of oversight; the daily ceilings determine the maximum permitted payment value. For individuals, the VND100 million ceiling is five times the VND20 million daily risk threshold. For organisations, the VND500 million ceiling is likewise five times the VND100 million threshold.

A personal customer whose AI payments total VND30 million in a day therefore falls within the high risk payment category while remaining below the national daily ceiling. The provider must still respect the customer's agreement and its own internal limits.

Existing online payment safeguards also remain relevant. Current rules require the account holder's biometric authentication for online transactions above VND10 million each. Biometric checks use physical characteristics, such as a face, to verify identity. Providers introducing automatic payments must account for those authentication requirements alongside the new AI controls.

Customers must agree and retain a way to stop

Banks and payment providers must agree the terms with customers before AI begins making automatic payments. The agreement must identify the permitted transactions, the limits for each payment and each day, its period of validity, applicable conditions and the customer's right to end it.

Customers must receive tools to suspend or cancel the AI system's payment authority themselves. They may terminate the agreement at any time, and the provider must act immediately on that request. A customer's ability to stop automatic activity becomes part of the service rather than a matter left entirely to internal bank procedures.

Providers must also disclose when AI replaces a person in direct communication. The notification must come at the beginning of the session or conversation. This covers a practical concern as banks introduce automated assistants to answer calls, provide advice and handle service requests.

These provisions give customers control at two stages: permission before automatic payments begin, and withdrawal of that permission afterward. The agreed payment scope and limits give both sides a concrete record of what the system was authorised to do.

Human supervisors need real powers

High risk systems require continuous monitoring, supported by staff with the authority and ability to override, adjust or suspend AI decisions. Supervisors must also be able to activate an emergency mechanism that stops or isolates the system.

During a serious incident, the provider must activate that emergency mechanism, take steps to limit damage and restore safe operations. Manual operation must be available when needed. The practical requirement is to preserve banking services while preventing a faulty or compromised AI system from continuing to act.

Each intervention must leave a record showing what happened, when it happened, who intervened and why. Those logs allow a provider to reconstruct decisions and investigate failures, including whether a warning was handled promptly.

Pham Anh Tuan, director of the State Bank's Payment Department, explained the reason for assigning responsibility to people and institutions. In an interview published on November 4, he said, in translation:

AI is not an entity that can take responsibility.

Tuan said the circular addresses model testing before deployment and the responsibilities of boards and executive management in approving systems. He also said human involvement becomes mandatory when AI exceeds the value limits within which it is authorised to decide.

Complaints must reach a human reviewer

A customer who complains about an automated decision must have that decision reviewed by a person. The result of the human review becomes the final basis for resolving the complaint through the provider's process. Providers must establish a department and procedures to receive and handle complaints involving AI.

Decisions affecting a customer's financial obligations, legal responsibilities or access to services also carry a recordkeeping requirement. The provider must retain the basis and main factors behind the decision so it can explain them when asked.

For a customer facing an automated credit decision, that means the institution needs a usable account of the factors that drove the outcome. The explanation requirement allows providers to protect source code, detailed algorithms, model parameters and business or technology secrets.

Security decisions have a further qualification. When action concerns fraud, scams, money laundering, terrorist financing or cybersecurity, providers may give reasons suited to managing those risks and protecting customer assets. Detailed disclosure could otherwise expose how security controls work.

Data quality and outside suppliers become management duties

Providers must know where the data used to train, test and operate AI comes from and ensure it is lawful. They must check its quality and detect errors, bias, manipulation or poisoning. Data poisoning means deliberately inserting misleading information that can distort what a model learns or how it behaves.

Monitoring must continue after deployment because the data entering a system can change. A credit model built around one pattern of customer behaviour may perform differently as economic conditions or borrowing habits shift. The circular requires attention to those changes alongside model risk, cybersecurity, ethics and supply chain risks.

Hiring an outside AI supplier leaves the bank or payment provider responsible for operational safety, customer rights, personal data protection and risk management. A supplier may use the institution's data to train, fine tune or improve models only after receiving written consent.

If an outsourced AI system harms customers or other affected parties, the institution using it must promptly take corrective action, protect customer interests, handle complaints and provide compensation as required by law. Customers therefore continue to deal with their service provider when technology supplied by another company causes a problem.

Security checks and governance extend beyond launch

High risk AI systems must undergo cybersecurity checks and assessments before operation, periodically afterward and when significant changes occur. Providers must also address data safety, changes to systems, incident handling, logging and human supervision.

Deployment requires an impact assessment, risk controls and documentation submitted to the appropriate approving authority before the system enters operation. Article 14 introduces a framework organised around three lines of defence, separating operational risk ownership, risk oversight and independent assurance.

Banks already use AI for credit scoring, loan assessment, payment fraud detection, automated advice and electronic customer identification, commonly called eKYC. At the Digital Financial Security Vietnam 2026 forum in Ho Chi Minh City on November 6, Dao Minh Tu, vice chairman and secretary general of the Vietnam Banks Association, cited figures showing that more than 70% of credit institutions had adopted AI.

Tu called for clearly defined limits on AI assistants, human approval of large or sensitive transactions, emergency stops and fraud detection while transactions are happening. He also advocated clean, consistently defined business data. These proposals address the speed at which an automated system can repeat an error across many decisions.

Fraud figures explain the urgency

A report by the National Cybersecurity Association and Viettel Cyber Security recorded about 552,000 cyberattacks in Vietnam during 2025, with approximately 41% directed at finance and banking. Applying that share to the total gives roughly 226,000 attacks aimed at the sector.

Authorities also recorded nearly 1,000 online fraud cases in the first six months of 2026, with losses reaching trillions of dong. Colonel Nguyen Ba Son, deputy director of the Ministry of Public Security's cybersecurity and high technology crime prevention department, warned at the November 6 forum that criminals were increasingly using AI.

Ly Hoang Vu, an economist at the French Vietnamese Center for Management Education, described how copied voices and fabricated faces weaken familiar methods of verifying identity. A case involving engineering company Arup's Hong Kong office illustrates the risk: in 2024, an employee transferred about HK$200 million after a video meeting featuring fabricated images of the chief financial officer and colleagues.

Existing fraud warnings show the value of intervention before money moves. Hoang Minh Tien, deputy director of the State Bank's Information Technology Department, said the SIMO payment fraud monitoring system had issued 5.8 million warnings by October 2026. Customers suspended or cancelled suspected fraudulent transactions on 1.9 million occasions, protecting nearly VND7.2 trillion in transaction value.

Tien also reported more than 21 billion cashless payments worth over VND260 quadrillion in the first eight months of 2026. At many credit institutions, digital channels account for more than 95% of transactions. Reliable oversight therefore affects the main way customers use banking services.

Implementation runs into 2027

Providers face several implementation dates rather than a single deadline. Systems operating before the circular takes effect must meet its safety, risk management and deployment requirements from March 1, 2027. Some governance and staffing provisions follow separate schedules.

• September 30, 2026: The State Bank signs Circular 53/2026/TT-NHNN.
• November 15, 2026: The circular generally takes effect.
• January 1, 2027: Article 14 on the AI risk management framework takes effect.
• March 1, 2027: AI systems already operating before the effective date must meet the required conditions.
• September 1, 2027: Article 21 on staffing takes effect, subject to an exception involving institutions with high risk AI systems.

The precise staffing timetable for that exception requires confirmation from the legal text. Institutions operating high risk systems need to establish which personnel requirements apply to them before relying on the general September date.

Security planning extends beyond Circular 53. A separate draft replacing Circular 09/2020/TT-NHNN proposes stronger encryption, data loss prevention, backups protected against unauthorised alteration and closer control of technology suppliers. It proposes that payment intermediaries and Mobile Money providers operate security operations centres and centralised security event monitoring from January 1, 2028.

Le Quang Ha, deputy director of Viettel Cyber Security, has advocated Zero Trust architecture, which repeatedly verifies access and limits movement between parts of a network. Intel Vietnam country director Phung Viet Thang identified openness, security and rapid recovery as core computing requirements. Tu also called for preparation for cryptography designed to withstand future quantum computing threats.

The State Bank and Ministry of Public Security have a nearer operational target under Coordination Plan 513, signed in June 2026. Tools to trace and urgently block fraudulent money flows are scheduled for deployment at most large commercial banks during 2026, with broader use across the system targeted for the first quarter of 2027.

Key Points

• Automatic personal payments above VND10 million each or VND20 million a day trigger the high risk category.
• For organisations, the corresponding thresholds are above VND50 million each or VND100 million a day.
• AI systems automatically deciding credit of VND50 million or more also carry high risk.
• Daily automatic payment ceilings are VND100 million for individuals and VND500 million for organisations, subject to lower internal limits.
• Customers must agree to automatic payments and receive tools to suspend or cancel that authority.
• Human supervisors must be able to override decisions and activate emergency stops; customer complaints about automated decisions require human review.
• The circular generally takes effect on November 15, 2026, with a March 1, 2027 compliance date for systems already operating before it takes effect.