The Difference Between Rules and AI in AML Technology
Artificial Intelligence (AI)-based AML (anti-money laundering) technology is essentially rules on steroids. In other words, AI leverages massive amounts of data to intelligently filter transaction and customer data to detect suspicious behavior. Using this state-of-the-art technology requires taking a new approach to AML Risk Detection, training with quality transaction and customer data, and explaining the risks detected in an interpretable manner.
At a recent webinar with ACAMS, we shared experiences of using AI for AML purposes and explained the benefits AML professionals see from using AI. We also discussed how to overcome the challenges that come with implementing AI for AML purposes.
Watch the full webinar. (Access with free ACAMS registration.)
Key Topics
- Taking a New Approach to AML Risk Detection
- Training AI with Quality Transaction and Customer Data
- Explaining the AML Risks Detected by AI
The Expert Panel
- Moderator: Tobias Schweiger, Co-founder & CEO, Hawk AI
- Speaker: Michael Shearer, Chief Solutions Officer, Hawk AI & Former Group Head of Compliance Product Management, HSBC
- Facilitator: Sarah Runge, Executive Managing Director, K2 Integrity
Taking a New Approach to AML Risk Detection
The Rules-Based Approach
The rules-based approach to detecting financial crime has traditionally been led by a risk steward and has been informed by data. With this method, an external authority (risk steward) dictates sets of rules and thresholds. The risk steward says that if these thresholds are reached, or if these events occur, you must raise a case. You program your machine to do that. The machine examines the data and generates cases, and then investigators look at those cases. This approach should feel familiar to most AML professionals.
The AI Approach
With AI, the traditional approach is inverted; it’s led by data and informed by the Risk Steward. First, the AI learns from what the investigators do. It watches what cases the investigators label as suspicious. Then, it looks at what behavior caused investigators to label customers as suspicious. In this process, AI allows AML professionals to reverse-engineer desired outcomes from transaction and customer data. The AI learns from the investigator rather than applying a set of rules that were dictated to it upfront. The AML risk steward’s role evolves from setting rules and thresholds to deciding what the AI needs to know.
“AI is really good at mirroring the behavior of your best investigator,” said Shearer. “If the investigator is seeing things that they think are of concern, then the machine can copy that. That applies to any type of behavior that the investigator sees where the machine also sees the same type of data.”
Training AI with Quality Transaction and Customer Data
To successfully implement AI for AML risk detection, you must understand that its needs are different from those of a rules-based system. Most of these requirements boil down to training AI with quality transaction and customer data. Here are a few ways AI detects AML risk differently from rules-based technology:
- AI learns just like we do…
- By Example: AI needs sample cases of suspicious behavior. The more cases an AI model sees, the better it learns how to respond to different situations.
- By Pattern: An AI model needs to review lots of normal behavior to spot the abnormal.
- AI needs to see the bigger picture via…
- More Attributes: AI needs a fuller view of the customer. AI can’t ask questions to fill in gaps, so it only works with the attributes you give it.
- Stability Over Time: AI needs a stable period of historical behavior to establish what’s normal behavior and what's unusual or anomalous.
- AI needs precision and correlation via…
- Data Quality: AI needs consistent, precise, and complete data. In other words, the data must be well organized. AI needs clear lines between Category A and Category B.
- Cause & Effect: AI needs to see causal relationships between input data and case outcomes (“If X, then Y”). If investigators are making different decisions based on data that the machine doesn't have, it can't learn that the behavior is bad. There must be a correlation between the behavior that you teach it and the outcomes you show it.
Helping AI learn to detect AML risk depends on using quality data to train AI models.
If your data quality is poor, you can clean it up. Start by selecting the data that you do trust. Train your AI on this data instead of throwing everything you have at it. It takes some skill and time to identify the good data and discard the bad data, but it pays dividends. When you do this, the machine can work with the good data, and you can move forward as you clean the lower-quality data.
Explaining the AML Risks Detected by AI
As part of their FI’s risk-based approach, AML investigators need an AI model to explain why it flagged behavior as suspicious. With manual rule configuration, this has been relatively straightforward. You would employ a relatively small number of rules, which were readable by a human, used a limited number of data points, and operated on a simple on/off binary. In contrast, AI will employ a large number of rules, utilize rules that are not easily readable by a human, use many features, and derive insights via statistics. This would normally make the AI model less accessible to both AML professionals and regulators. However, we have made great strides in Explainable AI technology. Now, AI models can deliver explanations, both in natural language and visually, of why they flagged a given behavior for investigation.
In the example below, we see that the AI model has assigned a risk score of 97% to a particular case of suspicious activity. We also see the individual risk factors that contributed to this risk score, such as “large amounts have been transferred in quick succession” and “transactions happen mostly between 12am and 6am”.
Hawk AI’s AML Technology
Hawk AI’s solution for AML empowers FIs to detect more suspicious activity using fewer resources than they could with rules-based technology alone. Our technology is designed with data science and AML expertise at its core to support a data-driven approach to financial crime risk detection. Our AI is fully explainable, providing natural language explanations and probabilities based on specific risk factors.
Request a demo today to see it for yourself.