Many experienced compliance professionals feel that the reality of Artificial Intelligence in the field of Anti-Money Laundering has been, at best, disappointing. Current AI-based AML solutions tend to focus on the single application of transaction monitoring, to the detriment of both money laundering detection and well-trained and user-friendly AI.
However, by taking a wider view of AML compliance and of real-world implementation of AI, in particular by better leveraging the role of the AML investigator and by staying laser-focused on the fact that AML is meant to catch criminals who are attempting to hide, there are in fact considerable gains to be made.
This article introduces an alternative approach that looks at financial institutions' AML programs as a combination of three primary activities, and proposes, at a high-level, ways to properly implement an AI-based solution that will significantly improve each of those three activities. A deep dive into the implementation of such a solution will be the subject of future articles.
The three primary activities that make up customer management stage money laundering programs include detection, investigation, and resolution.
The monitoring of transactions, customer demographic information, and customer interactions with financial institutions for suspicious activity, which includes most of the traditional transaction monitoring systems
The work performed by AML investigators to clear or verify the alerts from suspicious activity detection
The effort to write, approve, and submit SARs, as well as conduct auditing of the system to ensure compliance
The application of AI to transaction monitoring has traditionally focused on reducing the number of false positive alerts by improving the statistical models used to generate these alerts. While these new models can introduce improvement through the use of more complex mathematical methods than simple business rules, this improvement is well-known to be accompanied by the issues of limited explainability, limited knowledge of the subject matter -- criminal activity -- by model builders, and limited experience with the techniques required by rare event detection with partially-labeled data and high class imbalance. In order to combat these problems, an AI-based transaction monitoring system must be designed with the following in mind:
Clear and comprehensive explainability
In particular, an emphasis on feature engineering that makes business sense is a must and goes far in explaining even complex mathematical models.
Easily customizable business rules
Business rules are often the best choice given other considerations, including staffing, explainability, and ease of use, so the system must make reviewing, optimizing, and updating these rules easy to implement. This allows SMEs to contribute to the detection, while providing model builders concrete indicators of criminal activity that they can use to improve their modeling.
Extensive use of semi-supervised machine learning techniques
Criminal activity, while impactful, is rare and intentionally hidden by an adversaries, making detection more akin to finding hackers and terrorists than to the more familiar machine learning techniques used in credit modeling and advertising. Success requires applications of graph theory and network analysis, with a high degree of input from subject matter experts, together with storage technologies such as semantic data hubs and a strong focus on data hygiene.
Differential Privacy capabilities
To combat the limited AML data available, financial institutions would see great benefits from working together to discover and share the patterns and methods of money launderers; however, currently this is nearly impossible without revealing confidential or proprietary information. Differential privacy has the potential to provide this capability, with novel techniques from areas like zero-knowledge protocols, semi-homomorphic encryption, and the shared distributed ledger technology of blockchain. The first steps towards realizing this potential include the implementation of high standards of data management and consistent semantic data models.
Reducing the negative time and cost impact of false positives can come in two ways: a reduction in the overall number of false positives, and an improvement in the clearance velocity. High false positive rates are a reality due to the difficulty in modeling rare events, so using AI to improve the tools used by investigators to clear alerts can significantly reduce the overall time and costs dedicated to satisfying compliance requirements -- in other words, large gains can be realized by modeling the AI on the investigators as well as the money launderers. Furthermore, increasing the number of cleared alerts provides additional data that can then be fed back into the transaction monitoring AI training, improving the AI and further reducing the overall numbers of future false positives.
A well-designed AI-based investigator tool should be developed with the following considerations
Use of Design Thinking techniques
Often used in the design and development of customer-facing apps and tools, design thinking techniques are indispensable to developing easy and intuitive interfaces; by focusing on investigators as the end user, these techniques can produce tools that investigators want to use, and that improve their workflows over the current state of searching diverse and unconnected databases and third-party data sources, resulting in more efficient and productive use of expensive resources.
Comprehensive logging of investigator actions
With full instrumentation that will gather data on the actions taken by investigators in the course of their work, tool improvements become easy to identify and implement.
Continuous learning capabilities
Using the logged investigator actions as input to pattern detection algorithms promotes the fast & continuous identification of areas in which repetitive or redundant work can be automated for both individual investigators and the group, further reducing the amount of time spent investigating alerts. Additionally, the results of each investigation can be quickly fed back into the detection models to improve their capabilities, as well as create an environment in which changing criminal behavior can be identified and responded to in the short-term.
A significant amount of time is spent by analysts, investigators, and management in writing and approving SARs and dealing with audits and consent orders. Time and costs can both be reduced greatly with a well-designed AI-based system that properly implements the following:
Pre-population of SAR fields
A built-in flagging capability for the data used by investigators in the course of an investigation allows a system to identify and extract the correct information for SAR fields, as well as provide suggestions for the narrative portion using natural language processing techniques.
Prioritization strategy for submitted SARs
This empowers management to streamline the approval process using a scoring & prioritization system based on the outcomes of previous SAR decisions.
Full system transparency
Audits and reviews can be sped up significantly by using advanced data management techniques in lineage and provenance, which provide an integrated end-to-end audit trail on every decision made by a compliance team, including each transaction, customer, interaction, data nugget, and software test that was a part of a decision.
The Future: Augmented Intelligence
Artificial Intelligence can play a major role in reducing compliance costs, but to see the big payoff AI must go deeper than just transaction monitoring: AI should be used as Augmented Intelligence, assisting investigators and management in making the nuanced decisions that AI cannot. AML compliance should be regarded as a system that encompasses detection, investigation, and resolution activities; accordingly, the best AI system for AML will help humans make decisions and then use all the decisions in those areas to train and grow. From alerts to consent orders, all activities should be fully documented, with all decisions fed back into the system to bolster future improvements.
This quick overview aimed to show that while AI can be difficult to get right, it is not all hype. Once past facile solutions, development based on a deeper, more holistic understanding of the issues facing AML compliance can provide an effective, efficient, and elegant AI-based compliance solution.