Artificial intelligence in anti-money laundering

…Analytics using AI for anti-financial crime processes in banks

The financial services industry faces a significant challenge in preventing money laundering due to manual, repetitive and data-intensive tasks that are ineffective. AI has the potential to enable a quantum leap in AML and provide a means to scale and adapt to the contemporary threat of money laundering. However, the growing awareness and number of AI applications have sparked a debate over the effectiveness of these solutions and the extent to which AI could or should replace human analysis and decision-making.

A survey of businesses from a variety of industries found a correlation between the strength of an organisation’s AI-human partnerships and the organisation’s cost savings and overall revenues. To better explore and realise the potential of AI, the financial services industry must continue to expand its knowledge of AI’s capabilities, risks and limitations and establish an ethical framework through which the development and use of AI can be governed.

The vast majority of transaction monitoring systems (TMS) are based on generic risk typologies with threshold-based rules and other parameters. In practise, these rules generate a significant number of false-positive alerts that must be manually processed. The current TMS, including alert generation, can be significantly enhanced through the application of new technologies that can dynamically and customer-specifically adapt to rules without sacrificing effectiveness. In practise, we have observed the following problems with the effectiveness and efficiency of the banking systems currently in use:

Transaction monitoring

Theoretically, banks can implement ML across the complete AML value chain. However, we believe that transaction monitoring—specifically, the combination of ML with other sophisticated algorithms, such as random forest, gradient boosting and deep learning—is where banks can realise one of the most immediate and significant returns from their AML efforts.

Transaction monitoring — specifically, the combination of machine learning and other advanced algorithms — is one of the most immediate and significant anti-money laundering benefits for institutions. For transaction monitoring, numerous financial institutions currently employ rule- and scenario-based instruments or fundamental statistical methods. These rules and thresholds are determined primarily by red flags in the industry, fundamental statistical indicators and expert opinion.

However, the standards frequently fail to account for contemporary money-laundering trends. In contrast, machine learning models employ more granular, behaviour-indicative data to develop intricate algorithms. Moreover, they are more adaptable to new trends and develop continuously over time. By replacing rule- and scenario-based tools with machine learning (ML) models, the leading financial institution enhanced the detection of suspicious activity by up to 40 percent and increased its efficacy by 30 percent.

Alert and case processing

Frequently, the reasons for the termination of alerts are not documented in a standardised manner and as a result, are not taken into account when determining the next course of action. Additionally, recurring alerts confound alert processing. Even if the behaviour has been previously classified as non-suspicious, identical transaction patterns generate a new alert each time. Most systems lack an intuitive learning function.

Typically, alerts-to-SAR (ATSAR), cases-to-SAR (CTSAR), and other system metrics are not monitored in real-time. Using metrics, such as money laundering risk exposure and total cost of money laundering risk, management is unable to monitor case processing or system changes. In addition, feedback on a suspicious activity report (SAR) is typically unavailable from the Financial Intelligence Unit, and is therefore not utilised to improve the efficacy of the system.

Efficiency of today’s transaction monitoring

The regulatory authorities are increasing their transaction monitoring requirements, necessitating that banks pursue innovative and cost-effective solutions. An example is a bank that runs 10 million transactions per year and generates 20,000 alerts, of which 18,000 can be closed promptly as false positive alerts and 1,000 can be closed as false cases. This bank has alert-to-case and alert-to-SAR ratios of 10 percent and 5 percent, respectively. A rise in false-positive alerts has an immediate effect on personnel and operational costs. Due to system inefficiency, a large number of alerts are generated.

Each month, global institutions manually evaluate millions of financial crimes monitoring alerts, of which approximately 95 percent are deemed “non-suspicious”. Current procedures for alert review frequently rely on large, geographically dispersed personnel teams. A significant proportion of compliance professionals surveyed anticipated further personnel increases in 2019 to maintain current compliance levels. Therefore, investing in and implementing an AI-enabled TMS can generate a rapid ROI and increase productivity.

Current status of AI in transaction monitoring

In the current state of AI integration, transaction monitoring incorporates machine learning (ML). In practise, an increasing number of financial service providers are endeavouring to integrate use cases with the implementation of AI processes and systems. Successful incorporation of AI and ML into rule-based systems will determine the future of transaction monitoring.

In such a scenario, artificial intelligence would determine whether a particular transaction should be labelled as questionable. Similarly, the AI or an alert processor could conduct or assist with the case investigation. Experience has proven to us that the complexity of developing an AI application of this nature is slightly understated. Numerous institutions, especially small and mid-sized banks, have not yet integrated AI into their current systems.

In light of the current perspective of regulators and European institutions on this issue, it is prudent to consider using artificial intelligence to process and prioritise notifications from rule-based TMS. This expedites the identification of false-positive alerts and enhances the system’s performance.

Large international bank

A leading Asian bank recently implemented additional TMS, for instance. The integrated solution is based on the bank’s anti-money laundering (AML) framework, which includes “Know Your Customer”, “Transaction Monitoring”, “Name Screening,” and “Payment Screening”. This solution augments the existing rule-based primarily system with an additional module. The module contains an AI engine and functions as an additional unit to boost the system’s efficiency and productivity.

The objective of the transaction monitoring module is to identify new, previously unidentified suspects. With the help of the AI engine, the TMS rules and thresholds can be optimised by mapping and ranking previously generated alerts. Continuous and typically moderate ATL background monitoring of the AI system within the primary TMS accomplishes this optimisation, whereas BTL false-negative background monitoring boosts productivity.

Private/small bank

Implementing an AI-based module in a bank with fewer transactions can pose certain difficulties. The number of alerts, cases to be processed, and subsequent reports of suspicious activity will decrease as a result of the decline in transactions. Data science-wise, it is difficult to create a machine learning model with a tiny amount of data.

Due to a lack of training examples, the model might struggle to correctly classify alerts. It would also have a tendency to generate measurement errors, meaning that it could identify non-existent problems and dependencies and divert alert operators’ attention away from real areas of risk. A few companies have developed technology based on a variety of data science approaches to resolve the problems caused by limited training datasets:

• Typically, the system learns from vastly more data generated by SARs or studied cases. Using this method, the system is trained further by incorporating the alert processor’s closed alerts.
• Typically, these analyst activities are informal. Therefore, the system must perform process mapping using information that is stored but never utilised, such as analyst activity logs or similar data.
• Classifiers are typically supervised learning algorithms. These are less effective with highly specific data, such as a customer’s activity pattern. Utilising unsupervised learning algorithms to segment customers and transactions into more behaviourally homogeneous groups and then applying classifiers only to those groups.
• The alert scores are made interpretable so that the analyst can view all factors that increase or decrease the risk of money laundering for a particular transaction. Despite the fact that machine learning models used for alert classification initially provide no interpretability to the analyst, this can be achieved by displaying the parameters in the case management system.

>>>the writer is a finance professional and Chartered Accountant with extensive experience providing audit and assurance services to UK and international financial services companies. She oversees engagements to ensure regulatory and reporting requirements are met and client objectives are met on time. I am proficient in reviewing and preparing financial statements, external auditing and assurance, data analysis, and project management. She can be reached via  jek64@jbs.cam.ac.uk