Illuminating and Stopping Sanctions Activity with AI

Introduction 

Lynx has been building and applying artificial intelligence and machine learning technology to fraud prevention over the last 20 years and we are now applying this expertise to uncover and stop sanctions evasion. Though the methods for detecting and stopping fraud and sanction screening differ in some ways, there are similarities and best practices that can be leveraged between the two, such as: 

• Utilizing digital data to identify digital attributes of criminals 
• Uncovering organized crime network and associated compromised or synthetic identities 
• Identifying and stopping mass account takeover / onboarding resulting in a network of mule accounts 
• Uncovering malicious actors inside the organization who are both facilitating money laundering and fraud. 

We’re pairing this expertise in big data with practical industry experience to create detection tools that are more effective in identifying illicit activity (and therefore reduce false positives), as well as streamlining the way investigators work alerts and cases.  

From our experience in the field, we know that investigators will not be replaced by AI anytime soon. AI can, however, generate more productive alerts, automate time consuming manual tasks, and give investigators more time to focus on the activity that really matters. This is exactly what Lynx AML aims to achieve through its advanced machine learning and configurable case management solution – the automation of manual activities so that investigators can spend more time on investigating truly risky activity. 

Artificial Intelligence, Machine Learning, GenAI.. What’s the difference? 

First, let’s establish the difference between artificial intelligence, machine learning and generative AI. 

Lynx AML Sanction Screening uses a hybrid of artificial intelligence technologies, most of which are machine learning, but it’s worth establishing what the difference is between these technologies. 

Artificial Intelligence: these computer models perform complex tasks that exhibit intelligent human-like behaviors. As explained by the AI pioneer Arthur Samuel, AI “gives computers the ability to learn without explicitly being programmed.”[1]

Machine Learning: machine learning is a subfield of AI and learns to program itself using historical data. This can be done in 3 different ways: 

1. Supervised Learning: involves training the machine learning model with labeled data sets, after which the model can classify data or make predictions.  
2. Unsupervised Learning: trains the machine learning model using unlabeled data, from which the model looks for patterns and trends. Unsupervised learning is often used to understand trends and relationships within datasets and make it especially useful in AML Transaction Monitoring, where it’s difficult for the human eye or traditional algorithms to make draw trends across such large transaction data sets. 
3. Reinforcement Machine Learning: trains the model through trial and error rewarding the system when it takes the best action.  This method is often used when training autonomous vehicles. 

Generative AI: according to IBM, refers to “deep-learning models that can take raw data… and “learn” to generate statistically probable outputs when prompted. At a high level, generative models encode a simplified representation of their training data and draw from it to create a new work that’s similar, but not identical, to the original data.”[2] 

How Lynx uses expertise and experience to identify and stop sanctioned activity 

Lynx AML – Sanction Screening solution is AI-led, whilst also human-centered. The solution leverages human insight to target models effectively (via the data validation process detailed below) and to continuously receive feedback from experienced investigators to continuously incorporate SME knowledge to ensure continuous learning, iterative model updates and avoid model drift (e.g. model degradation as the data, patterns or attacks change) (e.g. avoid model drift). 

What information is screened? 

First, we need to do some groundwork to ensure our models are utilizing high quality data. To do this, Lynx starts with data validation, to establish accurate data feeds (e.g. who and what information is being screened?) and to target transaction and customer details with the appropriate level of precision. For example, when searching for a sanctioned individual’s name in a transaction, the model should apply an “approximate match” to the name field, rather than an “exact match” to cast a wide net to account for name variations. More on the model approaches are outlined below. 

Before we get to the models though, we need to determine the sources that provide the sanctions information, such as sanctions watchlists or internal lists that outline the individuals, entities, countries, ships, etc., with whom transactions are prohibited. Once this is established, our configurable case management system allows users with the appropriate permissions to create complex rules for these watchlists to ensure specific lists are included or excluded (E.g. filtering for OFAC to ensure US sanctions lists are included).

After the data has been validated, we have established the watchlist sources, and we have established the frequency of screening (we can screen real-time, as in milliseconds. a capability we’ve brought over from our real-time fraud solution), our models get to work screening transaction and customer details.

Lynx’s approach  

Lynx applies a layered approach to name screening. We first “widen the net” to consider multiple variations of the names involved in the transaction, as well as the names on the watchlists. This allows the models to account for nicknames, misspellings, phonetic similarities, adding/removing brackets and whitespaces, etc.. This of course increases the number of possibilities. 

We then apply methodologies to “narrow the net” or focus which name variations actually match the fields from the transaction. We apply multiple approaches here to derive a similarity score, including distance algorithms, phonetic algorithms, and AI models trained on large sets of data with known screening traps. Our models continually learn from new traps introduced by new data sets. The model ultimately assigns a score assessing how closely the transaction names align to the watchlist names. If that score exceeds the threshold set by our client, an alert is produced.

Using this hybrid and layered approach, we have seen decreases in false positives, while also achieving a high level of accuracy (AKA no false negatives).  

Why was Lynx able to reduce false positives?  

Legacy solutions have applied one or two methods to the matching process to ensure no missed matches but result in higher numbers of false positives. We drive down false positives without sacrificing accuracy by applying the right combination of data science and artificial intelligence to both expand and narrow the range of possibilities, without risking false negatives. 

Configurable Case Management 

Finally, we pair our academic expertise with practical insight to deliver alerts into a smart case management system that reduces the burden on compliance managers. We give you the tools to: 

• assign and prioritize alerts based on alert characteristics, 
• configure a pre-generated narrative template, 
• simulate different thresholds and watchlist filters to optimize alerts, 
• provide out-of-the-box reporting dashboards based on industry best practice,  
• provide dashboards configurable to your needs, 
• enable real-time reporting capabilities (No waiting for data refreshes… real-time KPIs.) 

And we’re not stopping there. We are building the next generation of AI-enabled case and quality management to bring the age of AI to compliance operations professionals. 

Ultimately, we aim to optimize the way organizations detect and manage financial crime, using AI-led, human-centered AI technologies to illuminate risk, eliminate the mundane and allow you to focus on what really matters. 

[1] https://mitsloan.mit.edu/ideas-made-to-matter/machine-learning-explained

[2] What is generative AI