In an earlier article entitled “The unquenched longing for a transformed KYC-AML solution” I had talked about the key challenges
that financial institutions (FIs) have been facing with regards to their current Know Your Customer (KYC) and Anti-Money Laundering (AML) operations. In order to overcome these considerable and lingering challenges, it has now become imperative that FIs leverage
new-age smart technology solutions. In this regard, I believe innovative artificial intelligence (AI) and machine learning (ML) enabled solutions can be a game changer for FIs.
AI/ML: Setting the context
Artificial intelligence (AI) allows IT systems to imitate the cognitive ability of human – for example “problem solving”, “reasoning”, “planning” and “learning”. AI enabled systems possess inbuilt intelligence to sift through, aggregate, blend, and identify
patterns and relationships that are buried within mountains of data - spanning large number and types of data sources.
AI comprises various branches such as machine learning (ML), affective computing, artificial neural networks, concept mining; natural language processing (NLP), semantic technology and fuzzy logic. Machine learning (ML), a type of AI, leverages sophisticated
algorithms to allow systems to “think”, and over a period of time, automatically train and enhance the systems’ outcome prediction accuracy. These systems don’t need to be explicitly programmed for their learning and evolution.
ML can be broadly grouped into supervised and unsupervised learning. In supervised learning, the system learns via feedback loop with the human user. So for example, when an FI’s staff receives a transaction alert, he tells the system why this particular
transaction is not suspicious. This feedback is then utilized by the system to label the data set and identify/classify similar alerts. Through such ongoing feedback mechanism, the system is able to evolve its pattern recognition ability.
Unsupervised learning, on the other hand, involves learning and drawing inferences from unlabeled data sets. This type of learning does not require algorithms to be trained by human with the anticipated data outcome. Rather, the system leverages iterative
deep learning techniques to continually learn and evolve. Unsupervised learning allows systems to detect out-of-the-norm behaviors - that otherwise could not have been caught through human reviews or knowledge-based rules.
AI/ML: Key capabilities
Following are some of the key capabilities of an AI/ML enabled system.
- Self-learning and predictive: a) contextual awareness; b) automatic and continual
feedback mechanism (to allow systems to learn and adapt to the changing environment, behaviors, regulations etc.); c) system can
generate own rules (and make independent judgment), and d) accurate
predictive algorithm (that considers high number of variables, numerous levels of non-linearity, and large data volume to automatically predict future scenarios and uncover sophisticated and hidden insights and patterns).
- Data processing: a) high-speed processing (of large data volume); b)
statistical analysis (for effective pattern identification); c) intelligent
ontology-based information extraction (OBIE); d) information classification and scoring (to reveal hidden trends, patterns and relationships); e)
probabilistic data matching (from multiple sources, to enhance the matching accuracy); f)
master data management (MDM, to provide single view of data from the disparate data sources); g)
central and unified data pooling (from numerous internal and external sources); h)
data annotation (in the unified data pool, to enable context-sensitivity); i)
linguistic analysis; h) NLP (for recognizing, interpreting and synthesizing the unstructured data); and j) graph databases (for higher performance, flexibility and agility, and to easily establish intensive data links).
- Automation: a) high-speed and scaled-up execution (of narrow but complex data-intensive tasks, that otherwise cannot be automated through traditional rule-based solutions); b)
workflow automation; c) robotic process automation (RPA, that automates repetitive structured data related activities); d)
intelligent process automation (IPA, that automates manual and highly effort intensive processes); and e)
cognitive automation (that uses advanced decision algorithms to automate complex multi-dimensional activities).
Illustrative AI/ML use cases in KYC/AML operations
- Customer onboarding: FIs’ existing rules-based onboarding systems enforce rigid, static and one-size-fits-all checklists for the KYC information collection. An AI/ML enabled system, on the other hand, can provide dynamic questionnaire that can adapt
as per the customer responses. Further, the system’s IPA capabilities would help significantly reduce FIs’ KYC profile update backlogs. Additionally, it would enable ability for real-time transaction-based KYC anomaly detection. Finally, the AI/ML enabled
system would facilitate robust automation that would improve process effectiveness and significantly reduce FIs’ staffing expenses related to the client onboarding and customer due diligence (CDD) processes.
- Link analysis: Leveraging AI/ML enabled system would allow for sophisticated link analysis. Linkages between entities can be established much more efficiently and effectively. In addition to the structured data sources, unstructured data sources
(such as online news and publications, social media, and public databases) would be swiftly analyzed for linkages identification. AI/ML solution’s Bayesian network and dynamic graphical modeling capabilities would facilitate easy understanding of the financial
relationships amongst large number of entities. System would be able to unearth highly complex, opaque and previously unknown connections between entities or transactions; and their remote linkages with suspect actors, ultimate beneficial owners (UBOs), politically
exposed persons (PEPs), prohibited states etc.
- Customer segmentation: Owing to their ability to speedily integrate data from numerous sources, AI/ML enabled systems would allow for intelligent customer segmentation and provide comprehensive 360 degree views of the customer. This would not only
benefit FIs’ KYC/AML operations, but also help them leverage these insights for their cross- and up-selling purposes. The systems’ generative modeling capabilities would facilitate holistic transactions analysis and generate customer behavior archetypes. Further,
by utilizing the systems’ clustering techniques, customer clusters within the KYC segmentation can be formed. This information can then be leveraged for identifying anomalies and high risk customers.
- Screening: AI/ML enabled system can screen for larger number of risk signals. For example, system would be able to detect money transfer to terrorists, who even though are in news, have yet to be added to the official watchlists. Further, unlike
the simple ‘name match’ approach of existing rules-based systems, in AI/ML enabled system, the screening process would adopt a context-driven ‘identity matching’ approach. In this approach, high number of attributes would be considered to enable a holistic
view and providing true identity match. Additionally, these system’s name matching and linguistic search capabilities would allow for data analysis in multiple scripts and languages. Further, the system’s linguistic search capabilities would help in significantly
improving regulatory (e.g. Office of Foreign Assets Control (OFAC)) compliance.
- Risk management: AI/ML enabled systems’ self-calibrating models can learn money laundering typologies and proactively ascertain specific (geographic, temporal, emerging etc.) risks. So for example, an emerging geographic risk – related to a previously
low-risk country that has of late become high-risk due to questionable activities of its prominent citizens - would be proactively identified. The context-sensitive system would intelligently analyze high number of contextual data attributes (such as account
profile data from CRM, non-transactional behavior from web login activity, and other unstructured data sources) to create a highly refined risk score. Further, the system’s exploratory data analysis (EDA) capabilities – that are leveraged on transaction data,
case management files, customer data, external data etc. – would help uncover the risks in real time. Better integration of the KYC and transaction monitoring systems would also be enabled – which would lead to stronger risk controls. Optimal ongoing feedback
from the transaction monitoring system would be provided into the KYC system. This would help continually evolve the customer’s relationship risk scores. Further, the systems’ sophisticated AML risk scoring capability would help in prioritization of investigation
queues for suspicious activity reporting (SARs). System would also be able to automatically comprehend the verbiages related to regulatory changes, and identify gaps in FIs’ existing KYC/AML operations.
- Transaction monitoring: AI/ML enabled system’s highly refined self-learning models would facilitate advanced and adaptive real-time transaction monitoring. For example, the models would continually learn and evolve the transaction behavior
profile of each customer and the customer categories (grouped by age, business type, geographic location etc.). Further, the system would be able to unearth erroneous invoice numbers, cryptic SWIFT messages, duplicate/linked addresses, originator to beneficiary
information (OBI) messages, and many other forms of hidden money laundering clues within the terabytes of transactional data. Additionally, sophisticated transaction monitoring capabilities for high risk geographies and entities (including those on OFAC and
specially designated nationals (SDN) lists) would be enabled. So for example, system can automatically flag transactions that are linked to regions that have been classified as unstable in recent period.
Further, the system’s NLP capabilities would allow for speedy analysis of unstructured data (negative news, social media etc.) for anomaly detection. Employees’ communication can also be effectively tracked to unearth internal money laundering related collusion
and fraud. System would also be able to detect new AML patterns in real-time. There would be significant reduction in false positives – as system would be able to continually analyze the false-positives and understand and learn the common predictors.
- Alert investigation, reporting and case management: AI/ML enabled system would continually learn and evolve its alerting models, and improve the alerts’ relevance. Further, it would enable robust and user-friendly alert investigation and case management
workflow. Smart dynamic dashboards and graph based database would deliver quality annotated data to the alert investigators. The system’s rich graphical data representation and intuitive UI capabilities would enable intelligent data visualization and easy
referencing. Further, the system’s NLP capabilities would help identify and parse through adverse news (in multiple languages) on the investigation subject. It would then translate pertinent information, and bring it to the investigator’s attention. Furthermore,
system would be able to continually learn the investigation steps from the past cases. It would then leverage these learnings to automatically investigate similar cases in future and provide most plausible scenarios and recommendations to the investigators.
System would also be able to recommend intelligent approaches for resolving new cases. AI/ML enabled system would also facilitate integrated reporting (currency transaction report (CTR), SAR, Compliance Regulatory Reporting (CRR) etc.) and audit trails. FI’s
systems can be seamless linked to the regulators’ systems for automated SAR reporting etc. Additionally, automation of the record keeping and post-investigation case resolution activities would be enabled.
AI/ML adoption for KYC/AML operations is gaining traction…few examples…
FIs – especially the large ones – can gain immensely by leveraging AI/ML in their KYC/AML operations. AI/ML enabled system can massively enhance the overall KYC/AML process efficiency and effectiveness, significantly reduce staffing needs, and bring in huge
cost savings for FIs. Additionally, it would significantly improve customer experience and which would have a positive impact on FIs’ top line.
Given such a compelling value proposition, it is strongly recommended that FIs don’t endlessly dither on adopting AI/ML for their KYC/AML operations. The “first mover” advantage in this case is too huge to ignore. The good news for FIs is that AI/ML enabled
KYC/AML solutions would complement their existing rules-based systems - these new-age solutions can be implemented on top of an FI’s existing systems. Hence, end-to-end replacement of the existing systems would not be required.
Importantly, it is highly advisable that FIs take a strategic and gradual approach towards adopting AI/ML enabled solutions. This is because, although fast developing, the AI/ML technology is still in an evolutionary stage. Hence, it would take few years
before these solutions would reach optimal maturity. FIs should therefore, in the initial period, focus on adopting AI/ML enabled solutions in KYC/AML processes where these would have maximum impact with minimal disruption. Link analysis, watchlist filtering,
alert investigation, name matching, communication surveillance, and linguistic search capabilities are few good candidates for early stage adoption.