Macao's monetary authority tells every financial institution that its board carries ultimate responsibility for AI and that high-risk AI should be independently validated before launch

Who is responsible for AI at the institution?

The board. The circular says the board "須對AI的應用和有關的風險管理承擔最終責任", which in our rendering is that the board shall bear ultimate responsibility for AI use and the related risk management. For a branch, the circular reads the board as local senior management or the head-office management responsible for the branch.

The same section asks the board to set specific strategic goals before introducing AI, assess the business impact, assign suitably qualified staff and resources, build a data-governance and AI-governance framework, and fold it into the institution's overall risk management. It also asks the board to understand the risks and technical limits of AI and to keep the staff involved in data, development, testing, validation and monitoring trained.

Responsibility does not move with the work. In the vendor section the circular says that even where risk-management duties are divided with a third party, the board still bears ultimate responsibility for AI outputs and potential risks.

What counts as high-risk AI?

The circular asks institutions to tier AI uses by their effect on customers or internal processes, model complexity and reliance on the AI. It then names four characteristics of high-risk AI decisions. Any one marks the class on our reading of the list: a significant effect on customers' rights or property; a significant effect on operations, information security, assets or continuity; use in regulatory compliance, customer due diligence, anti-money-laundering and counter-terrorist-financing or fraud monitoring, detection of cyber attacks, or a significant effect on compliance or critical risk processes; and use of large volumes of sensitive or private data for training, or sending customers' sensitive personal data to an AI application after deployment.

Institutions are also asked to keep a complete AI inventory, updated when things change, recording at least the use case and technical limits, model type and architecture, data sources and governance, the risk tier, and the oversight and control mechanisms.

Human control is tied to the tier. The circular describes human-in-the-loop (a person keeps real control and a veto before a decision takes effect), particularly for high-risk uses, and human-on-the-loop or human-in-command oversight with the ability to intervene or trigger an emergency stop.

What does the lifecycle section ask for?

Governance of data comes first, then a lifecycle mechanism covering development, data pre-processing, training, explainability and stability validation, deployment and monitoring, and post-termination data storage and destruction, with record-keeping, change-management review and contingency measures.

For high-risk AI, validation before deployment should be done by a team independent of model development and training, or by a third party. The scope should cover at least conceptual soundness, data quality, model security, reliability and technical limits, and performance, and the results and the effectiveness of mitigations should be reported to the board. For other AI the circular allows validation to be scaled to business need and done by suitably qualified internal staff.

After launch, institutions should monitor continuously and start remediation when metrics breach preset thresholds or when the circular's listed triggers appear: data drift, concept drift, structural change in the environment, business-scope expansion, or a major change to the model itself.

For high-risk AI the circular asks for fail-safe mechanisms, defined operating boundaries and risk tolerances, and a kill-switch or manual intervention that "須立即啟動" (shall be started immediately) when thresholds are exceeded or a major anomaly occurs, with the ability to switch to a backup system or manual process, and periodic drills. On ending an AI use, and subject to applicable data-retention requirements and retention of data necessary for audits and dispute handling, institutions should securely destroy, clear or de-identify training and validation data, model weights, feature vectors and temporary files so they cannot be restored, and keep proof of destruction where a vendor does it.

What changes for vendors and customers?

Before taking third-party AI, institutions should divide risk-management and monitoring duties with the vendor and assess the vendor's AI and data governance, on top of the existing outsourcing guidelines the circular cites. High-risk AI should get code review, vulnerability scanning and security testing before deployment or a major change, by people independent of the business and the vendor. Institutions should also manage concentration risk from relying on one or a few AI vendors.

On data, the circular asks for contracts that expressly prohibit vendors, without the institution's prior written authorization, from retaining data or using it to train their own models or for other purposes.

For customers, institutions should tell them up front that a service involves AI and explain in plain language its purpose, basic decision logic and limits. For customer-facing high-risk uses the circular asks for an alternative, such as human-in-the-loop or a purely manual route, and for an accessible complaint channel with timely, fair human review. It also asks for regular bias and fairness assessment, with a non-discriminatory business reason and legal and compliance review where sensitive characteristics, or proxies for them, are used.

What does the circular ask institutions to do before launch?

In its closing paragraph, the AMCM says that to ensure AI applications carry controls proportionate to their risk, an institution intending to implement a high-risk AI application should first communicate fully with the Authority. The circular names the Banking Supervision Department as the contact, with the email address dsb@amcm.gov.mo.

On our reading this works as an early-engagement expectation rather than a formal approval step. The text does not say the AMCM approves or refuses anything, and sets no timetable. That reading is ours, and an institution with a live project should ask the department directly.

What we did not verify

What we opened: the full six-page circular PDF from the AMCM's own domain, read in the original Traditional Chinese with Portuguese headings. Quotations are reproduced in the original Chinese characters from that PDF; every English rendering of the circular is ours and is labelled as such. The extraction contains a few simplified-character variants that we have not altered inside quotations.

What we did not open: any Portuguese or English version of the circular, the AMCM outsourcing guideline the circular cites, any AMCM press notice or FAQ, any enabling provision of Macao law, and any guidance on how the AMCM will treat a late or missing pre-launch conversation.

What we refuse to claim: that the circular has the force of law or carries sanctions; that any institution has complied, or that any deadline applies; how the AMCM will apply "should" language in examinations; or that the pre-launch communication is an approval requirement. We make no comparison with other jurisdictions beyond the related links.

Informational analysis for working professionals, not legal advice. Confirm how any rule applies to your situation with qualified counsel.

Related briefings

• Hong Kong HKMA: ART framework for responsible AI in banking
• China NFRA: AI safe development in banking and insurance
• Singapore MAS: AI risk management consultation
• Asia Pacific AI regulation tracker
• AI Regulation News hub