Supporting Adoption of Artificial Intelligence in Fighting Financial Crime

Supporting AI adoption in AML/CFT 1 Supporting A.I. Adoption in Fighting Financial Crime June 2026

Supporting AI adoption in AML/CFT 2 3 5 7 7 9 10 11 12 14 16 19 22 23 Contents 01 Foreword 02 Hong Kong – The Changing AML Landscape Fighting Financial Crime: A.I. Adoption – What To Expect

Point of View
Call to Action
03 Internationally Aligned Perspectives 04 Case Studies in A.I. Adoption in Fighting Financial Crime
Case Study 1
A Global Bank’s Approach to Dynamic Risk Assessment (DRA)
Case Study 2
A Regional Bank’s Approach to Dynamic Risk Monitoring Model (DRM)
Case Study 3
Face Watch List in a Digital Bank
Case Study 4
Detecting and Disrupting Money Mules throughout the Client Life Cycle 05 Where to From Here? 06 Glossary Supporting A.I. Adoption in Fighting Financial Crime

Supporting AI adoption in AML/CFT 3 Over the past year, I have spoken frequently about the rapid evolution of financial crime risks and the increasing importance of technology in reducing the harm which it causes while also safeguarding the integrity of our financial system. That message bears repeating – standing still in the face of such rapid change is not an option. Digitalisation has fundamentally altered the way in which financial services are delivered as well as the speed, scale and sophistication of criminal activity. Fraud networks operate across borders and leverage technology to adapt quickly. The rise and impact of Artificial Intelligence (A.I.) driven scams are making detection harder and more scalable. At the same time, synthetic A.I. powered identities have the potential to challenge the effectiveness of existing approaches to verification. In this environment, traditional approaches to transaction monitoring (TM) and static, rule-based controls are under growing strain. To remain effective, our defences must evolve at an equal pace. 3 01 Raymond Chan, Executive Director (Enforcement and AML) Foreword Supporting A.I. Adoption in Fighting Financial Crime Digital Payments Tripled in Four Years:

FPS Transactions Index: 329 (+229%)
FPS Registrations Index: 206 (+106%)
Remote account opening reached 74% of all accounts
opened in Q4 2025 Data source of FPS transactions and registrations:
https://fps.hkicl.com.hk/eng/fps/about_fps/statistics.php

Supporting AI adoption in AML/CFT 4 A.I. is no longer an emerging concept in anti-money laundering (AML) frameworks. It is already embedded in many institutions’ monitoring, analytics and investigative processes. We have seen tangible benefits where A.I. and network analytics are used to identify suspicious accounts earlier, uncover hidden connections, and intervene before harm escalates. When this data is shared through Hong Kong’s maturing framework of public-private partnerships (PPPs), significantly improved results and disruption are being achieved. These are meaningful advances. However, innovation must be accompanied by accountability. Where A.I. systems influence risk assessments or decision pathways, the responsibility remains firmly with senior management and Boards. Technology does not replace governance. Institutions must be able to explain how models operate, how outcomes are validated, and how risks such as bias, drift or over-reliance are managed. Supervisory tolerance for experimentation exists within well-defined guardrails — not outside them. Supporting A.I. Adoption in Fighting Financial Crime 4 The next phase of development will see a more consistent sector-wide move beyond efficiency gains. A.I. should not be confined to accelerating documentation or optimising workflows. Its real potential lies in strengthening early detection, enhancing intelligence-led monitoring, strengthening complex decision making and enabling more proactive intervention against bad actors. To achieve this, the ecosystem must ensure that innovation is underpinned by strong governance frameworks, supported by cross-line collaboration, and aligned with clearly articulated risk appetites. Our objective is clear. To foster responsible innovation that materially improves the effectiveness of Hong Kong’s anti-money laundering and counter-financing of terrorism (AML/CFT) regime. By combining technological capability with disciplined oversight, we can strengthen trust in our financial system and better protect the community we serve. Our aim in this programme is to encourage the industry to set ambitious goals and move towards them collaboratively. Standing still is not an option. Banks need to adapt and constantly strive for greater efficiency and effectiveness, adopting proven technologies at scale and being prepared to explore and test new technologies as threats develop. Otherwise, they will be overwhelmed by threats and noise from systems that do not adequately mitigate those threats. Raymond Chan Executive Director (Enforcement & AML)

Supporting AI adoption in AML/CFT 5 Industry A.I. Adoption Detecting, deterring and disrupting financial crime risks remains a cornerstone of financial stability and market integrity. Advances in technology have seen financial crime risks evolve in scale, complexity and sophistication requiring banks to adapt their risk-management and control frameworks. The innovation of new technologies has also presented banks with opportunities to strengthen their riskmanagement playbook by deploying network analytics, A.I. and advanced behavioural analytics to reshape how they tackle different financial crime threats. A further catalyst for how banks are reshaping core financial crime prevention capabilities is the closer collaboration between regulators, banks and law enforcement (for example through PPP like FMLIT 1 and private-private platforms such as FINEST 2 ), which is helping to improve financial crime risk mitigation effectiveness. The ability to prevent financial crime, driven by enhanced collaboration, prediction and detection capabilities, has never been stronger. This collaborative innovation has also supported the acceleration of A.I. adoption and is challenging the way the industry thinks about transforming how financial crime risk is managed. Supporting A.I. Adoption in Fighting Financial Crime 5 0 5000 10000 15000 20000 25000 2019 2020 2021 2022 2023 2024 2025 Number of new suspicious accounts associated with crimes under investigation by law enforcement agencies shared via FMLIT 1 Established in 2017, Fraud and Money Laundering Intelligence Taskforce (FMLIT) is a collaboration between Law Enforcement, the HKMA and the banking sector together with the Hong Kong Association of Banks (HKAB) under the leadership of the Police Force. 2 Established in 2023, the Financial Intelligence Evaluation Sharing Tool (FINEST) is a bank-to-bank information sharing platform initiated by the HKAB with the guidance from the HKMA and operated by the Police Force to enable timely sharing of financial crime risk information among banks. 02 Hong Kong – The Changing AML Landscape Number of New Suspicious Accounts

Supporting AI adoption in AML/CFT 6 The financial services marketplace in Hong Kong has witnessed an expansion of banks (traditional and digital), payment providers and Stored Value Facility Licensees, and this in itself is fuelling greater investment in A.I. solutions – with varying levels of maturity – as these operators look to manage financial crime risk more effectively. Against this backdrop, the HKMA has been promoting a multi-year industry initiative to drive the responsible adoption of A.I. technology. This has formed an integral
part of the regulator’s forward-looking vision for Hong
Kong banks to continue their digital transformation journey under its Fintech 2030 strategy 3 . The HKMA asked banks with significant operations in Hong Kong to undertake a feasibility study 4 into adoption of A.I. in their AML/CFT monitoring systems and to formulate implementation plans based on the outcome of those reviews. The results showed broad alignment on the continued adoption of A.I. – and new technologies more broadly – to strengthen financial crime risk detection. Supporting A.I. Adoption in Fighting Financial Crime 6 The outcome of the feasibility studies, alongside a number of live use cases in the Hong Kong market, were showcased at an industry workshop in November 2025. The key message from industry leaders at the event was that banks should not adopt A.I. for A.I.’s sake but deploy targeted solutions that are aligned with business strategy, operating models and levels of readiness. A number of these solutions are outlined in the Case Study section of this publication. 0% 5% 10% 15% 20% 25% 30% 35% 40% 45% 50% Investigation Alert Triaging Risk Detection Type of A.I. Use Cases in TM workflows* Adopted by 2025 by 2026 by 2027 A.I. Adoption Trends in TM Workflows *One bank may have more than one A.I. use case 3 HKMA Press Release on “The HKMA Unveils “Fintech 2030” at the Hong Kong FinTech Week 2025” issued on 3 Nov 2025. 4 HKMA Circular on “Use of Artificial Intelligence for Monitoring Suspicious Activities” issued on 9 Sep 2024.

Supporting AI adoption in AML/CFT 7 From control to curiosity The prevailing mindset within many financial crime environments has historically been one of compliance and control: minimising regulatory exposure, limiting model risk, and constraining operational variance. While these disciplines remain essential, the next phase of A.I. adoption requires a shift towards structured curiosity. Institutions will need to create safe environments to test generative artificial intelligence (GenA.I.), advanced analytics, and network intelligence – accepting controlled failure as part of innovation. Regulatory expectations are evolving in parallel, with supervisors increasingly distinguishing between unmanaged experimentation and well-governed innovation. Early A.I. initiatives have often been deployed as point solutions: TM enhancements, name screening optimisation, or case summarisation tools. Over the coming two years, competitive advantage will emerge from holistic architectures that integrate data strategy, model governance, workflow management, and human oversight. A.I. adoption should be designed end-to-end — from data ingestion to alert disposition — rather than bolted onto legacy rule engines. This requires aligning data standards, model validation, explainability, and operational processes into a unified framework. Buildingblock experimentation remains valuable, but it must feed into an enterprise A.I. blueprint. Holistic architecture over building blocks 01 02 Fighting Financial Crime:
A.I. Adoption – What To Expect Supporting A.I. Adoption in Fighting Financial Crime 7 Point of View A.I. adoption in financial crime functions will accelerate from controlled experimentation to enterprise-wide integration. Institutions that succeed will not simply deploy new tools; they will adopt new mindsets, operating models, and governance frameworks that position A.I. as a strategic capability rather than a technical enhancement. Financial crime risk does not originate in the compliance function; it originates in products, customers, channels, and geographies. A.I. programmes must therefore be coowned by business lines and control functions. Over the next 24 months, we expect to see more financial crime subject matter experts embedded in frontline teams, joint model governance committees, and shared accountability for A.I. performance metrics. Effective adoption will depend on breaking down silos between first and second lines of defence. When business leaders understand how A.I. improves customer onboarding, reduces friction, and enhances detection quality, secure and scalable business growth can be better supported. 03 Risk resides in the business A.I. models are only as strong as the data and intelligence that inform them. Financial institutions are increasingly looking beyond internal data sets, leveraging PPPs, law enforcement typologies, financial technology collaborations, and cross sector intelligence sharing. The expansion of information sharing frameworks and privacy enhancing technologies will enable richer network analytics and cross-institutional risk signals. Leveraging public–private partnerships and the wider ecosystem 04

Supporting AI adoption in AML/CFT 8 Controlling the narrative to the Board Board engagement will be decisive in securing funding, visibility, and sustained momentum. A.I. in fighting financial crime must be positioned not merely as supporting compliance obligations but as a strategic enabler of improved customer experience and growth. Reporting to the Board should balance opportunity and risk: clear articulation of measurable outcomes (fewer false positives, more efficient investigation, better detection), transparent risk mitigation strategies, and forward-looking investment roadmaps supporting capabilities that are delivering the most improved outcomes. Institutions that control this narrative will unlock sustained capital allocation and cross-enterprise prioritisation. The most significant shift will be away from static rule-led frameworks toward dynamic, risk-based intelligence. traditional rule-based engines, while transparent, are reactive and threshold-driven. A.I. enabled systems will increasingly leverage behavioral analytics, network link analysis, anomaly detection, and adaptive learning to prioritise risk. The future state is not the elimination of rules, but their integration within layered intelligence models. Supervisors expect explainable, well governed A.I. that demonstrably improves effectiveness rather than merely efficiency. From rule-based to risk-based intelligence 05 06 Supporting A.I. Adoption in Fighting Financial Crime 8

Supporting AI adoption in AML/CFT 9 Call to Action The next 24 months will define how transformative A.I. in financial crime can become in Hong Kong with many banks already driving innovation programmes forward to deliver material benefit and measurable outcomes. Institutions that continue to embrace curiosity, integrate holistically, collaborate across lines of defence, leverage ecosystem intelligence, and secure Board sponsorship will move beyond automation toward intelligence-led risk management. Deliver measurable effectiveness, not experimental activity Programmes must deliver measurable improvements in financial crime risk management outcomes. Boards should expect clear evidence that A.I. is:

Improving detection while reducing unnecessary
alerts
Accelerating investigations without compromising
judgement
Making regulatory reporting clearer and more useful
Prioritising genuinely higher-risk activity through
intelligence-led models
The objective should not be automation for its own sake. It must demonstrate improvement in risk identification, prioritisation and escalation. While more than half of financial institutions report A.I. deployment within risk functions, fewer than one third operate fully integrated model governance frameworks across lines of defence. Institutions should:
Establish formal cross-line A.I. governance
committees
Define Board approved A.I. risk appetite statements
Implement real-time monitoring for model drift, bias
and explainability
Conduct annual scenario testing aligned to enterprise
test frameworks
Embed cross-line governance and clear accountability 01 02 It is important to understand that the presence of A.I. tools is not enough. There needs to be sufficient governance, effectiveness and strategic oversight surrounding them. Board-level sponsorship, measurable outcomes, and crossecosystem intelligence integration will distinguish between compliant adoption and transformative risk management. Static rule sets are increasingly insufficient against adaptive and networked financial crime threats. A.I. adoption should enable:
Dynamic, behaviour-based risk scoring
Integration of external intelligence and typologies
Continuous model recalibration in response to
emerging threats
Transparent explainability to supervisors and internal
audit
03
Transition from rule-based monitoring to intelligence-led risk management In Hong Kong and in other jurisdictions, regulators are playing a key role in supporting and encouraging innovation to identify foundational and next-generation technology capabilities which will support strategic transformation to financial crime risk management. Through this programme of supervisory support, the HKMA has formalised how it encourages innovation in identifying A.I. solutions to better manage the harm caused by financial crime. In addition to tracking progress of each bank’s implementation plan, over the next 24 months, Boards and senior management are expected to demonstrate progress across these supervisory priorities, anchored to institutional plans. 04 Collaboration across the AML/CFT ecosystem In Hong Kong and in other jurisdictions, regulators are playing a key role in supporting and encouraging innovation to identify solutions to better manage the negative effects of financial crime.
Hong Kong has made great progress in integrating
PPPs into the AML/CFT eco-system. Technology, including A.I. is recognised as a key enabler that will allow PPPs to move from incremental level improvement to system level impact
The behavioral shift will frame the next phase of
AML/CFT transformation, enabling banks to foster collaborative learning and deliver sustained resilience Supporting A.I. Adoption in Fighting Financial Crime 9

Supporting AI adoption in AML/CFT 10 Internationally, efforts against financial crime are entering a structural phase in which technologies – including A.I. – are central to the debate around effectively implementing the risk-based approach which underpins the global standards on AML/CFT set by the Financial Action Task Force (FATF). The FATF has encouraged the use of innovative technology, which can improve effectiveness and proportionality in applying AML/CFT controls while also maintaining safeguards in the promotion of financial inclusion. The core challenge is no longer whether institutions have compliance frameworks in place to meet legal and regulatory requirements, but whether those frameworks can generate meaningful, network-level intelligence, with the objective of driving stronger outcomes against agreed priority threats in an increasingly complex environment. Criminal networks operate across firms and borders. A.I. — particularly where it supports advanced analytics of shared data in Hong Kong’s PPPs — enables detection that reflects this reality. It allows institutions to identify behavioural anomalies, hidden ownership links, and interconnected actors that would not be visible through rule-based monitoring based on individual entity datasets. At the same time, global transparency reforms — including those relating to beneficial ownership — are producing vast datasets. A.I. is what makes these datasets operational. Without advanced analytics, transparency remains passive. With it, institutions can resolve entities, detect circular fund flows, and identify risk concentrations across sectors. A.I. also offers a more proportionate alternative to blunt de-risking strategies. Instead of withdrawing from higherrisk categories wholesale, advanced modelling allows firms to differentiate genuine risk from contextual noise, supporting both financial integrity and inclusion. However, the international consensus is clear: A.I. is not a substitute for governance. Its deployment must be in line with applicable laws and regulations, be explainable and lead to measurable effectiveness outcomes. The future of financial crime prevention will depend not simply on adopting A.I., but on embedding it within collaborative intelligence frameworks that strengthen trust as well as capability. In this evolving landscape, effectiveness — not volume of reporting — will define success. 03 Internationally Aligned Perspectives Supporting A.I. Adoption in Fighting Financial Crime 10 PPPs have made great progress over recent years, but most PPP models around the world still rely primarily on regular meetings between individuals. The more advanced PPPs have shifted to leverage technology and “data fusion” to support advanced analytics of shared data. Technology, including A.I., is a key enabler that allows PPPs to move from incremental improvement to system level impact. We are seeing scam groups and organised crime use A.I. to attack financial institutions. We now
need to see PPPs maximise the use of technology, data fusion and A.I. - under appropriate data protection safeguards - to improve detection of criminal networks, reduce duplication of compliance efforts, improve suspicious transaction report quality and deliver earlier, more impactful disruption of financial crime. Nick Maxwell, Head of the Future of Financial Intelligence Sharing (FFIS) Research Programme delivering “The new era of private sector collaboration to tackle economic crime” in the 2025 APAC Fighting Financial Crime Conference, hosted by the HKMA.

The case studies shared in this paper are based on the adoption journeys of four banks in Hong Kong and illustrate innovation in three different areas to support their strategies to combat financial crime. Case Studies in A.I. Adoption in Fighting Financial Crime 1 Bank A Dynamic Risk Assessment built holistically using external resources partnership 2 Dynamic Risk Model built holistically using internal resources 3 Mule monitoring and detection using a building block approach for Retail during client onboarding 4 Mule monitoring and detection using a building block approach for Corporate during client onboarding Supporting A.I. Adoption in Fighting Financial Crime 11 Bank B Bank C Bank D 04

Supporting AI adoption in AML/CFT 12

Why it was put in place:
re-engineering financial crime detection Up to 2020, most global banks were operating large-scale, rules-based AML systems to identify suspicious activities which generated large volumes of false positives that required manual investigation, creating operational strain and limiting effectiveness. The challenge was two-fold:

Escalating financial crime complexity, including
networked laundering and rapid movement of funds
Inefficiency in rules-based detection, which struggled
to prioritise true risk
The bank’s top leadership recognised that incremental tuning of rules would not be sufficient, so it pivoted towards an intelligence-led risk-based model aligned with regulatory expectation and best practice. In partnership with a cloud provider, the bank codeveloped a dynamic A.I.-driven platform designed to autonomously identify suspicious behavior patterns rather than relying solely on static rules. The strategic objectives were to:
Improve detection precision
Reduce false positives
Shorten investigative timelines
Enhance identification of criminal networks
Supporting A.I. Adoption in Fighting Financial Crime 12 Case Study 1 A Global Bank’s Approach to Dynamic Risk Assessment (DRA)

A summary of the construction
and technical architecture
Every bank will have its own process of implementing such a programme. However, the core components are likely to include:

A cloud-based A.I. platform that generates risk scores,
uses transaction data, customer profiles, Know Your Customer (KYC) data, and prior suspicious activity that integrates into case-management systems for analyst review
A Domain Training Layer, where the bank contributes
historical financial crime data, customer and transactional datasets, alongside internal typologies and risk expertise
Machine learning models that are trained to identify
suspicious activity including the detection of patterns such as rapid fund movements and abnormal behavior changes
Cloud-native deployment introducing A.I. design for
AML purpose for secure implementation within the bank’s own cloud tenant, addressing model governance and regulatory requirements
Scale of operations to handle vast data volumes, which
in this instance averaged monitoring over one billion transactions per month

Supporting AI adoption in AML/CFT 13
3. Collaboration timeline: what
worked and what did not
The bank had been exploring cloud analytics and machine learning use cases well before DRA’s formal launch. By around 2021–2022, it partnered with the cloud provider to pilot A.I.-enhanced AML detection. The successful trial led to the formal launch of DRA as a product in 2023. As with all large-scale programmes and “first through the door” innovators, the bank faced its own unique set of challenges. The solution worked in areas such as reducing false positives by 60%, increasing the detection of suspicious activities by 2 to 4 times, and faster execution of case investigation (c.50%); all coupled with strong model governance and secure cloud deployment. Some of the challenges of such a large-scale transformation programme included transitioning from compliance-driven, rules-based frameworks to A.I.-first, risk-based models. It involved communicating with multiple regulators as to the value of the shift and demonstrating that risk detection would be at least as good as under the previous systems. Whilst this was daunting, it ultimately paved the way for a paradigm shift in how the rest of the regulated industry would come to view the inadequacies of relying on just rule-based monitoring. Other challenges that were encountered and overcome included ensuring explainability and model validation in a highly regulated environment as well as cultural and operational adaptation within compliance, operations and business teams. Supporting A.I. Adoption in Fighting Financial Crime 13
4. Measurable outcomes and
what’s next
The bank reported filing over 137,000 suspicious activity reports in 2025 globally, underscoring the scale of its financial crime operations. Strategically it has improved signal-to-noise ratios, enabled investigators to prioritise higher-risk cases, enhanced network-level detection and reduced operational compliance burden. It has also positioned the bank as a reference case for its A.I.-driven AML transformation. The bank is currently expanding its deployment capability across more markets and business lines as well as investing further in cloud-based analytics and A.I. It is further strengthening its A.I. governance and validation structures and framework as well. The dynamic risk model represents one of the most prominent large-scale A.I. deployments in financial crime compliance in Hong Kong and globally. It represents a move from rule-based compliance to predictive, intelligence-driven risk detection. It combines A.I. platforms for AML/CFT with domain expertise and data scale. Operationally it has delivered measurable performance gains, and it illustrates for the industry how A.I., when embedded within governance and regulatory frameworks, can materially transform risk management in global banking.

Supporting AI adoption in AML/CFT 14

The strategic pivot
The bank became increasingly convinced that rule-based monitoring was not fit-for-purpose to handle complex and layered money laundering schemes which were becoming increasingly dominant. Instead of single suspicious transactions, the bank saw multi-step, multi-product crossjurisdictional flow deliberately crafted to avoid detection. These legacy rules-based systems struggled to see patterns that unfolded across accounts, products and time horizons. They also generated too many false positives, despite tuning efforts, creating operational strain and alert fatigue and increasing the risk of missing the signal that mattered. Following the analysis of the response in one case that exposed how fragmented systems were, accelerating the change needed to shift from chasing alerts to understanding patterns, thereby moving from rule-based detection to risk-based intelligence. That decision was not about technology or A.I. alone but a strategic pivot to reimagine how compliance, data and analytics could work together. Supporting A.I. Adoption in Fighting Financial Crime 14
The holistic approach
This required being able to see the entire picture of customer behavior. It involved building integrated datalakes, combining customer profiles, transactional activities, behavior patterns and external data into a single environment. A dynamic risk-scoring model was introduced to continuously adjust risk ratings based on behaviors and context, not simply on static thresholds. Network analytics played a role in allowing the identification of links between customers, counterparties and entities that would have otherwise gone unnoticed. This drove a shift from a siloed transaction-only perspective to an intelligence-led, behaviorally informed approach. A shift from “What is suspicious?” to “What is really happening?”. The core objective was to improve detection rather than just efficiency and centred on the predictive detection of bad customers by assigning scores. This differed from other banks’ strategies, which use A.I. primarily to achieve efficiency. O C B C Risk Level 0 DRM score 1 HOLISTIC RISK SCORING Comprehensive risk score on customers, ranging from 0 (lowest risk) to 1 (highest risk) Differentiation from other “DRM”

Pro-active detection of bad
actors
Does not replace Periodic
Reviews
Proactive Risk Mitigation
-Increased Effectiveness
Dynamic monitoring of
customers
Replaces Periodic Review
procedure
Reduces volume of reviews
Increased Efficiency
Other “DRM” Bank B’s DRM
Machine Learning – Name Screening Machine Learning – Alerts Prioritisation PoC Machine Learning – Transaction Screening PoC Agentic AI – Investigation Report Agentic AI – Source of Wealth Report Global Visual Investigation Platform Sanctions & AML Workbench FINAL LAYER OF CONTROLS DRM ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! ! Dynamic Risk Monitoring Model Case Study 2 A Regional Bank’s Approach to Dynamic Risk Monitoring Model (DRM)

Supporting AI adoption in AML/CFT 15
3. Deeper dive into the model
The system operates at a post-onboarding, layered process level and is part of the ongoing monitoring that is continuously scaling customer risk, rather than focusing on pre-boarding and focusing on activities. It is not a binary choice (rules-based vs dynamic) but a two-layered process as it continues to use the transactionmonitoring system in parallel. The first layer of alert prioritisation is crucial to efficiently handling the volume generated by the underlying systems before applying the second predictive DRM layer. This foundational alert prioritisation layer is machine learning based and uses transaction data to define a score and prioritise alerts. The advanced DRM layer then sits on top of the prioritisation model and connects with other models such as Network Risk Model to take into consideration the holistic activities of the customer. This sequencing is vital, especially for smaller banks, as it adopts dynamic principles after it solves the foundational and structural issues of high alert volumes before tackling more advanced problem statements.
4. Navigating implementation
challenges
Breaking the silos – getting old systems to communicate with new ones was the largest challenge, coupled with each business having its own standards of data, definitions and systems. This made integration complex and time-consuming requiring investments around data governance, standardisation and architecture. Supporting A.I. Adoption in Fighting Financial Crime 15

Shifting mindsets and culture – In addition to reform of
systems, significant re-training of staff was required to enable them to operate as analysts and investigators rather than simply enforcing rules. The shift in mindset from control to curiosity was the key to successful adoption.
Justification of the investment – the narrative to
leadership needed to be re-framed from compliance spend to business value through better service for lowrisk customers, improving efficiencies and building reputational resilience. It had to be a Board-level discussion and once the Board acknowledged that this wasn't just about catching criminals but to also enable better banking, support began to flow naturally.
Impact on the customer – a significant benefit was the
reduced friction for legitimate customers. With more accurate risk scoring, fewer false positives and unnecessary reviews. High-risk activities were being detected earlier and more precisely in the cycle especially prior to issues being escalated. In short; less noise, more focus and better service.
Roadmap for A.I. deployment – the bank has put in
place a roadmap for the ongoing innovation and development of A.I. and Data Analytics (DA). The roadmap sets out its plans for new A.I. and DA tools to be added to further strengthen its compliance risk management. Alerts Prioritisation Model Name Screening Model Global Visual Investigation Platform for Network Analytics Sanctions & AML Workbench A.I. Driven S cam Detection Model DRM Model CURRENT TOOLS Agentic A.I. – Source of Wealth Report Agentic A.I. – Investigation Report Global Visual Investigation Platform UPCOMING TOOLS Agentic A.I. – Client Lifecycle GenA.I. – Advisory GPT Now 2026 2027 A.I.-powered Fraud and Market Conduct Surveillance Looking Ahead: The bank’s A.I. / DA roadmap for 2026/2027

Supporting AI adoption in AML/CFT 16

Background and context
A digital bank had to be more innovative in its pursuit of a target A.I solution to make optimum use of limited resources. Seeing a rise in mule activities industry-wide and individuals opening accounts for illicit fund movement, the bank recognised the needs to overcome these challenges by balancing seamless digital onboarding with stronger fraud detection. Using a building-block approach, they started small, demonstrated value and prepared to scale thereafter. By leveraging the HKMA GenA.I. Sandbox for the initial development and testing of the proof of concept (PoC), the bank was able to build and evolved a number of practical use cases, through multiple iterations and an agile development approach, with limited internal resources (including data engineers and data scientists). By deploying a test-and-learn approach to develop small modular A.I. models, the bank went beyond rule-based transaction to holistic monitoring of customer behaviors. The use case involved detecting third parties attempting to impersonate customers by identifying irregularities in the backdrops of their pictures submitted as part of the remote onboarding process. Supporting A.I. Adoption in Fighting Financial Crime 16 The outcomes of the case were later used to help law enforcement further investigate and disrupt criminal syndicates using this modus operandi.
Why adoption was successful
A. The innovation was practical and executable with strong results:

It has seen mule account detection increase by
30%
Screening times have improved to within seconds.
By identifying malicious actors at the initial stage,
the solution has reduced end-to-end monitoring costs delivering benefits that cascade through the entire AML/CFT risk monitoring process B. It was not difficult to scale the A.I. used:
Implementation was fast. A PoC was initiated in
June 2025 and the first version deployed into production within 3 months
This included training the machine learning models
to learn suspicious background scenarios Case Study 3 Face Watch List in a Digital Bank

Supporting AI adoption in AML/CFT 17

The agile development approach, streamlined
governance and industry infrastructure (Sandbox) enabled the team to secure executive sponsorship, drive forward the development in multiple sprints and accelerate delivery
C. It involved more than financial crime professionals
in the second line:
Incorporated skillsets across the bank and not
solely reliant on financial crime compliance staff
Demonstrated cross-function collaboration and
agility between risk stewards, business lines, technology and data architecture as well as management

How it began
It was initiated by several team members analysing an actual mule-account case. It involved dissecting the case including the account and customer profile, transaction patterns as well as digital footprints including facial images. Several scenarios were identified that had common attributes.
How does it work?
It uncovered the same unknown individual appearing in facial recognition records of multiple fraudulent account holders. It identified that a controller was possibly managing multiple accounts simultaneously, and the images of the suspected controller would be added to the bank’s internal watchlist. Supporting A.I. Adoption in Fighting Financial Crime 17 It was also noticed that there were common backgrounds in some of the images. These included suspicious backgrounds involving alleged gambling or drug dens, illegal casinos and other unusual locations. In addition, specific objects appeared in the same background of different account holders. In one instance, syndicated activity was indicated in over 200 accounts by a similar moving vehicle, an apartment with similar furniture or customers holding on to multiple sets of handphones while logging on to the bank’s accounts.
Steps to using building blocks
With smaller agile banks, problem statements are often broken down into building blocks or bite-sized chunks, while identifying the key components needed to achieve the desired output. Components can often include:

Image database set up to store all onboarding and
step up authentication images as well as an adjacent database to house all suspicious images. The next step is labelling them into different categories to train the A.I. models
Getting the right A.I. model to address the problem
statements that utilise both real-time and batch processing screening for fast, accurate and nimble deployment
Continuous re-calibration including training and
validation of the A.I. models to further enhance accuracy and to build out the library of suspicious images

Supporting AI adoption in AML/CFT 18
6. Hurdles and Outcomes
The adoption of new technology brought some challenges. The key to approaching challenges is to break them down into smaller blocks and tackle them one at a time. In this use case there were two main hurdles:

Matching speed
The initial PoC took five to seven days to run a match against a million images; this evolved to between 1 to 3 seconds with results delivered in real-time within the onboarding process. Certain scenarios such as identification of suspicious backgrounds in images take slightly longer.
Training and accuracy
Initial results showed between 65% to 70% accuracy, which was later enhanced after studying the failure points. It was fixed by labelling the suspicious part of the image, which drastically improved the learning rate of the A.I. model. One of the most measurable adoption outcomes involved reducing TM alerts by 25% largely because bad actors could be stopped at the door, so the bank did not waste precious time and resources detecting them after onboarding. The Human Element The human-in-the-loop factor is critical for implementation and outcomes. Subject matter experts are constantly needed to teach and train the A.I. model on how to identify suspicious images, including translating those requirements for the technology professionals involved in the training and application. Last and most importantly, support from senior management was critical to deliver the outcome the business required. Leveraging Public-Sector Support Seeking support from the public sector (i.e. through the HKMA GenA.I. Sandbox) to take an initial idea from baseline to “Minimum Viable Product” was invaluable in helping the bank generate the required internal support and funding to drive and scale up the A.I. solution. Supporting A.I. Adoption in Fighting Financial Crime 18

Supporting AI adoption in AML/CFT 19 A large bank became aware of increasing payment fraud volumes within its Hong Kong commercial banking portfolio, where shell companies were used to facilitate transactions and flow funds from illicit scams through the bank’s accounts. The rise in the number of mule accounts and the increasing sophistication of mule networks left the bank with limited visibility into the sources of incoming funds and the risks present at account origination. Although the bank had strong payment fraud detection systems, it needed to strengthen its ability to detect shell companies and mule accounts used for money laundering, as traditional rules-based approaches were less effective in identifying and freezing these funds. The business (First Line) implemented a new “Two Masters Model”, by leveraging prevention tools to freeze payments in near real-time, with investigation teams providing strong support and collaboration. The shared objective was to shift from reactive investigations of money mule transactions and shell companies to a proactive mode that would:

Prevent the onboarding of money mules based on
various DA (e.g. device biometrics, addresses etc.)
Identify money mules based on transactional and
behavioural analytics
Use the banks fraud risk controls to freeze funds in
near real-time
Two New Detection Tools –
Application Engine and In-Life Monitoring To respond to the problem, the bank developed two new detection capabilities to monitor customers at onboarding (account application) and throughout the customer life cycle (In-Life). Supporting A.I. Adoption in Fighting Financial Crime 19 Prevent: Application Engine (Onboarding) An “Application Engine” was built with real-time application programming interface (API) to enable onboarding assessment and controls leveraging data sources including:
Device intelligence
Cross-application analytics (against historic onboarding
applications)
Business information (KYC data – internal and
external).
The engine scored a customer at onboarding against a specific set of risk indicators and behavioral identifiers with the application receiving:
Pass / Reject / Review (escalate for human
intervention and assessment)
Review cases were used as part of the tuning and feedback loop to train and strengthen the solution. By iteratively testing and tuning the analytics models, the bank trained its application engine to reject suspected money mules at onboarding, helping to reduce exposure earlier in the client life cycle. Disrupt: In-life Monitoring (Customer Life Cycle) A second “In-Life Monitoring” capability was established to monitor customer account behavior for suspicious activity across the whole client portfolio combining:
Traditional human-explainable risk indicators
A new machine learning prediction model
Case Study 4
Detecting and Disrupting Money Mules throughout the Client Life Cycle

Supporting AI adoption in AML/CFT 20 The solution leverages data (KYC, transactional and external) and machine learning models to:

Assign a risk score to the customer, or
Evaluate against predefined typologies
Where suspicious behaviour is identified, the bank can intervene quickly through targeted actions, for example:
Initiate a Customer Due Diligence (CDD) review
Raise an Unusual Activity Report (UAR)
Implement payment controls on the account
Delivery: Approach (Agile, Co-Creation, Pace) Both capabilities were developed using an agile execution approach led by the business (First Line) with “cocreation” workshops incorporating representatives from multiple areas i.e.:
AML / Fraud
DA
Technology
Critical to this was the ability to incorporate knowledge, experience and learnings from different functions and subject matter experts to build strong risk models. In addition, to ensure pace of delivery and continued support from executive sponsors (budget owners), the business implemented an iterative development approach to:
Drive innovation
Test prototypes
Fail fast
Deliver quick wins
A Solution Pioneered Through
Data & Analytics
The bank adopted a transformation approach to the design and orchestration of the innovative solutions. A cross-functional team led by the business incorporating colleagues from risk functions, data office, payment operations and Information Technology were brought together to collaboratively evaluate how new technology and risk-detection techniques could be used to solve the money mule detection challenge. Fundamental to this was the support of DA colleagues and data scientists to:
Identify the data sets required to develop the solution
Establish thresholds to deliver appropriate risk scoring
across the different data sets (transactional, behavioral, network, KYC and risk indicators)
Build out the network analytics capability and define
the machine learning models
Ensure adequate levels of data quality
Business and risk colleagues were engaged to validate the model outcomes for explainability based on subject matter experts' knowledge and judgement. In addition, the models were subject to formal model governance. A key takeaway was that while data quality is typically critical to the success of traditional rules-based detection, the machine learning models improved risk detection even when some data attributes were missing or data quality was suboptimal. The machine learning model rated 42% of customers as high risk whereas the rules-based detection model had previously rated these customers as medium risk. Supporting A.I. Adoption in Fighting Financial Crime 20

Supporting AI adoption in AML/CFT Supporting A.I. Adoption in Fighting Financial Crime 21 Of these customers, 70% were escalated for investigation and 40% of these customers were exited for financial crime concerns. By leveraging:

Network analytics for relationships (mule networks)
Machine learning for detection
GenA.I. for explainability (to summarise money
laundering signals, contextualise findings and augment external research) The team were able to design, test and tune an effective detection model for money-mule activity. Adoption by other markets The early success in detecting and disrupting money mule activity in Hong Kong, together with the continued evolution of the solution, generated interest from other jurisdictions and the wider bank helping the team to sustain ongoing funding and executive support. While initially designed to solve a growing financial crime issue in Hong Kong, the solution has subsequently been rolled out in several other markets with commercial banking portfolios. Its application is also being evaluated in other business lines e.g. retail banking. The data-driven and collaborative approach has also been presented as a model of success internally to drive the development of other financial crime risk management projects within the bank; and the techniques established are being leveraged by the wider business. By moving beyond a purely rules-based approach to aggregated risk dimensions, the solution enabled the bank to target shell company and money mule behaviour more effectively across the client life cycle. The Model Data-Led Money-Mule Framework Prevent – Application Engine Real time API enable Onboarding Assessments and Controls Cross Application Analysis to Identify Bad Actors Cross Line of Business / Entities Bad Actor Data Exchange Predict – Data & Analytic Asset Unconventional Data (e.g., device intelligence) Behavioral Data Network Analytics Data Machine Learning Models Highly Refined Rules Disrupt – In-Life-Engine Integrated Payment Control Direct referral CDD / Trigger / UAR Onboarding In-life

Supporting AI adoption in AML/CFT 22 From Automation to Symbiosis:
The Emergence of Agentic A.I.
In any discussion of the strategic transformation of financial crime risk management, the adoption of agentic A.I. is already front and centre; it is no longer conceptual or theoretical. However, deployment remains largely efficiency oriented as A.I. is being used to summarise cases, triage alerts, draft narratives and optimise workflows. It is accelerating processes, but it is not yet materially reshaping detection strategies. A majority of Hong Kong banks reported active A.I. deployment in risk or compliance functions, with a few describing those capabilities as decision-supporting beyond workflow optimisation. Fewer still currently deploy autonomous analytical agents within core detection environments. Most programmes remain tightly constrained to productivity enhancements rather than intelligence-led amplification. This begs a strategic question for the industry:
If agentic A.I. can analyse network relationships, test hypotheses, simulate behavioral scenarios and adapt to emerging typologies, why is it not yet being deployed as a supervised co-pilot in financial crime detection? Financial crime threats are increasingly adaptive, networked, and technologically enabled. Agentic A.I. introduces the potential for:

Dynamic hypotheses generation across linked
accounts and counterparties
Continuous and tireless anomaly detection beyond predefined rule parameters
Real-time investigative assistance during complex
case escalations
Proactive identification of emerging typologies
The near future is not autonomous decision-making without oversight. It is supervised augmentation with A.I. operating as a co-pilot alongside investigators and risk officers. In this model, the human would define the risk, the risk appetite, the judgment thresholds and escalation pathways. The agentic A.I. would continuously interrogate data, unveil unseen connections and challenge assumptions. All surrounded by governance frameworks to ensure transparency, explainability and auditability. This represents a shift from A.I. as a productivity tool to A.I. as a strategic intelligence partner. The use of agentic A.I. will be a future focus area of the programme. 05 Where to From Here? Supporting A.I. Adoption in Fighting Financial Crime 22

Supporting AI adoption in AML/CFT 23 Glossary of Terms Supporting A.I. Adoption in Fighting Financial Crime A.I. Artificial Intelligence AML Anti-Money Laundering APAC Asia-Pacific API Application Programming Interface CDD Customer Due Diligence CFT Counter-Financing of Terrorism DA Data Analytics DRA Dynamic Risk Assessment DRM Dynamic Risk Monitoring Model FATF Financial Action Task Force FFIS Future of Financial Intelligence Sharing FINEST Financial Intelligence Evaluation Sharing Tool FMLIT Fraud and Money Laundering Intelligence Taskforce GenA.I. Generative Artificial Intelligence HKAB Hong Kong Association of Banks HKMA Hong Kong Monetary Authority KYC Know Your Customer PoC Proof of Concept PPP Public-Private Partnership TM Transaction Monitoring UAR Unusual Activity Report 06 23

Supporting Adoption of Artificial Intelligence in Fighting Financial Crime

Enclosure - Supporting A.I. Adoption in Fighting Financial Crime

Similar documents from other regulators