Use of Artificial Intelligence for Monitoring of Suspicious Activities

Our Ref.: B10/1C
B1/15C
9 September 2024
The Chief Executive
All Authorized Institutions
Dear Sir/Madam,
Use of Artificial Intelligence for Monitoring of Suspicious Activities I am writing to share with the industry how the use of artificial intelligence can help improve the effectiveness and efficiency of monitoring money laundering and terrorist financing (ML/TF) risks, and actions that the Hong Kong Monetary Authority (HKMA) is taking to further support and accelerate the use of artificial intelligence in Authorized Institutions’ (AIs) ongoing monitoring processes. In light of the increasingly complex ML/TF risk landscape, including global and regional money laundering cases, growing use of mule account networks for laundering of fraud proceeds as well as increasing geopolitical risks, the HKMA has been encouraging AIsto enhance their monitoring and mitigation of these risks, while maximising the intelligence being shared within the industry. Part of these efforts has targeted AIs’ adoption of technologies to make their ongoing monitoring systems more effective and risk-based, thereby providing the opportunity to prioritise resources towards higher-risk activities. The HKMA shared a list of high-level principles on artificial intelligence (https://www.hkma.gov.hk/media/eng/doc/key-information/guidelines-and￾circular/2019/20191101e1.pdf) with AIs in 2019. Since then, some AIs have adopted artificial intelligence to enhance their monitoring systems and controls. The HKMA observes that systems powered by artificial intelligence take into account a broad range of contextual information focusing not only on individual transactions, but also the active risk profile and past transaction patterns of customers in determining whether the activity of a customer should be flagged for further investigation. These enhanced systems have proved to be more effective and efficient than conventional rules-based transaction monitoring systems commonly used by AIs. More details regarding the benefits of using artificial intelligence to enhance suspicious activity monitoring can be found at the Annex. In view of the above, the HKMA is taking the following initiatives to further support and accelerate the use of artificial intelligence in suspicious activity monitoring:

• 2 -
  (a) Sharing experience and success stories of AIs – in November 2024 the HKMA will organise an experience sharing forum with speakers from the industry and technology firms on how artificial intelligence is being deployed to enhance the effectiveness and efficiency of suspicious activity monitoring; (b) Providing targeted guidance and support – the HKMA will establish a dedicated team, supported by an external consultant, to provide supervisory feedback and technical guidance to assist AIs in applying artificial intelligence in enhancing their monitoring processes, through the existing Fintech Supervisory Sandbox and Chatroom; and (c) Creating a conducive environment for AML/CFT innovation – the HKMA welcomes the use of artificial intelligence in AML/CFT work, and will continue to gauge the interest of AIs in applying artificial intelligence to the monitoring of suspicious activity and provide suitable guidance to the industry where appropriate. AIs with significant operations in Hong Kong should give due consideration to adopting artificial intelligence in their ML/TF monitoring systems to enable them to stay effective and efficient. A feasibility study based on the AI’s current ML/TF risk profile and possible evolvement in the near future should be undertaken. Based on the outcome of that review, AIs should formulate an implementation plan, taking into account dependencies, including as related to data requirements outlined in the HKMA’s recent guidance1 . The feasibility study and implementation plan, should be signed off at the board level and submitted to the HKMA by the end of March 2025. The HKMA will communicate with AIs on an individual basis regarding the exact timing for the feasibility study and implementation plan and the format in which they should be provided, and will consider further engagement and follow up in due course. AIs may approach their usualsupervisory contact in the HKMA’s AMLand Financial Crime Risk Division on any questions relating to this circular. Yours faithfully, Raymond Chan Executive Director (Enforcement and AML) Encl. 1 HKMA Circulars “Report on AML/CFT Regtech: Case Studies and Insights Volume 1” issued on 21 January 2021, “Report on AML/CFT Regtech: Case Studies and Insights Volume 2” issued on 25 September 2023, and “Thematic Review of Transaction Monitoring Systems and Use of Artificial Intelligence” issued on 17 April 2024.

---

1
Annex
What are the benefits of artificial intelligence in monitoring suspicious activities? In recent years there has been an increasing trend towards adoption of artificial intelligence to help prevent and detect financial crime. Machine learning is one of the most common forms of artificial intelligence which some Authorized Institutions (AIs), taking into account the size and complexity of their businesses, have used to optimise the effectiveness and efficiency of anti-money laundering (AML) systems. These AIs have found that effective transaction monitoring requires finding the right balance between technology and human expertise to optimise processes. This note illustrates some of the benefits AIs can expect to see from using artificial intelligence, and should be read together with other guidance provided by the HKMA1 , which details how some of the challenges inherent to implementing artificial intelligence for AML can be overcome.

1. Wider data coverage (Effectiveness)

• In rules-based transaction monitoring (TM) systems, alert generation is mainly
  based on customers’ existing transaction activities within certain prescribed periods (e.g. daily/ weekly/ monthly) without taking into account historical transaction activities outside these periods or other related contextual data from the subject customer.
• In contrast, the higher computing capabilities/ processing power in artificial
  intelligence-empowered monitoring systems enables the collation and analysis of vast amounts of data from various sources, which can greatly enhance AIs’ detection capabilities for suspicious activity.
• Artificial intelligence has the capability to generate a much larger network of
  interrelated rules for different segments of an AI’s customer portfolio. This can facilitate deeper insight, a more comprehensive and accurate assessment of a customer’s ML/TF risk profile, and hence achieve a more effective outcome, providing the ability to identify suspicious patterns that may be difficult to spot in high volumes of data.
• Sources of data covered by artificial intelligence-empowered monitoring
  systems usually include:
• AI’s own data - data regarding the customer profile (e.g. obtained at on￾boarding), transactions, previous alerts and results of any review or
  analysis, and where suspicious transaction reports (STRs) have been made within the AI’s own system.

1 HKMA Circulars “Report on AML/CFT Regtech: Case Studies and Insights Volume 1” issued on 21 January 2021, “Report on AML/CFT Regtech: Case Studies and Insights Volume 2” issued on 25 September 2023, and “Thematic Review of Transaction Monitoring Systems and Use of Artificial Intelligence” issued on 17 April 2024. .

2

• External data - intelligence from financial institutions within the same
  banking group, law enforcement agencies, information sharing platforms (e.g. the FMLIT or FINEST 2 platforms) or watchlist information.

2. Enhanced detection capability (Effectiveness)

• Artificial intelligence-empowered monitoring systems are able to identify new
  typologies and detect complex networks of suspicious customers and their activities more effectively.
• The system also generates less noise and enables AIsto target higher-risk cases
  for investigation, freeing up resources of the AML function to focus on genuine high-risk cases. AIs which have adopted artificial intelligence￾empowered monitoring systems have been able to achieve a higher STR conversion rate without compromising system effectiveness.

3. More timely follow-up action (Efficiency)

• Many AIs which have deployed artificial intelligence have cited addressing
  the overwhelming volume of false positive alerts, a long-standing industry pain point, as the primary driver. These alerts are not truly suspicious but must be cleared out of the queue, wasting valuable time and effort.
• The preliminary assessment results generated from an artificial intelligence￾empowered monitoring system replace the different levels of initial
  assessment manually conducted by staff, thereby reducing overheads and, where implemented effectively with adequate testing and retraining of the model, significantly streamlining operations for alert handling and investigation.
• Artificial intelligence-empowered monitoring systems facilitate
  differentiation and prioritisation of cases of higher probability of financial crime for investigation and STR filing. Investigation of alerts generated from a purely rules-based TM system are consolidated into a holistic review on a customer-relationship basis, which can be conducted by a single investigator. This can shorten the overall time required for investigation, enabling the AI to mitigate the ML/TF risk sooner.

4. Improved sustainability (Effectiveness and Efficiency)

• Compared to rules-based TM systems which adopt static rules and parameters
  and require constant manual intervention for adjustment to cover the latest red 2 FMLIT stands for Fraud and Money Laundering Intelligence Taskforce and FINEST stands for Financial Intelligence Evaluation Sharing Tool.

3 flags and typologies, an artificial intelligence-empowered monitoring system has embedded machine learning capabilities and can be trained on comprehensive data, customers’ transaction activities and other parameters to detect more complex typologies and emerging risks.

• Manual rule-tuning to capture these emerging risks can be challenging,
  because codifying the new typology into quantifiable behaviours and then setting appropriate thresholds and segments takestime. Retraining an artificial intelligence empowered model can be much faster as there is no requirement to go through the whole cycle manually.
• New typologies and hidden patterns can then be identified through the AI’s
  assessment of a customer’s transactional activities. These new typologies, together with those shared by law enforcement agencies and through public private partnerships and intelligence from previous STRs, are fed into the model with a view to identifying similar patterns in the future, thereby enhancing detection capabilities in a more sustainable way.