Introduction to AI based AML

The amount of money laundered each year is estimated at 2 to 5% of global GDP, up to $2 trillion annually, yet traditional detection systems in financial services detect only a fraction of this. Recent efforts by banks pioneering AI in financial crime have shown more impressive results.

HSBC’s launch of Google AML AI reduced alert volumes by 60% while finding 2 - 4 times more suspicious activity. That means for every 100 alerts HSBC now works, they find 5 - 10 times more suspicious activity than they did before they deployed AI.

Everyone is talking about AI, but what makes AI so much more effective than traditional rules based methods, and why should you trust it?

What are the problems with traditional methods?

Detecting financial crime is a complex problem space with vast amounts of data, varied criminal strategies, a hugely diverse set of honest customers and a constantly evolving risk landscape. With this backdrop it is easy to see why systems developed using rules have limited success, they are inevitably either too prescriptive and miss lots of risk or too broad and trigger many false positives which are expensive to process. Often both can be true simultaneously, and many false positives are generated while large numbers of suspicious events slip through undetected.

Rules based systems rely on risk stewards, developers and data scientists creating:

1. Red flag indicators drawn from previous suspicious events
2. Scenarios and thresholds which attempt to identify basic patterns of suspicious activity

Both of these steps introduce information loss which reduces the resulting system's effectiveness. Creating red flags from different suspicious events loses relevant information and nuance from those events as they are mapped to common indicators. Coded scenarios and thresholds rely on staff correctly interpreting and programmatically replicating events, a challenge in many areas of software development and incredibly difficult in financial crime. This process takes considerable time as new threats emerge, with new scenarios taking months or years to filter through the first line to be physically implemented in the transaction monitoring systems.

The result is that a rules based system:

• Considers limited ‘red flag’ data in its assessment of risk.
• Is simplistic in its consideration of that data, since it is limited by a human’s ability to consider, process and develop them into usable software.
• Misses large amounts of risk and entire typologies, as there are no accurate scenarios developed for them.
• Generate many false positives, since money launderers attempt to disguise patterns as genuine activity, and scenarios aren’t sophisticated enough to differentiate between a normal customer and a money launderer trying to look like a normal customer.
• Lags identification of new risk significantly, since new typologies need to have red flags, scenarios and thresholds developed, tested and deployed.

Why is AI based anti-money laundering more effective?

AI systems are more effective than traditional methods in most industries because of two main advantages; computers can consider far more data and identify more complex patterns than humans can. A well engineered AI system can review the complex interplays between risk indicators and contextual data for each customer, providing far greater accuracy and agility than a rule designed to check for a simple scenario.

Supervised AI systems also build on the decades of experience operational teams and the results rules have accrued. Supervised models use previous AML events (such as SARs and FC Exits) to identify the customer behaviours that led to those events, by doing this AI systems leverage historic successes while learning to avoid the mistakes.

Specifically, in AI based AML detection:

• A broad range of information about the customer is codified into ‘features’, which the system can understand and process. These features can include data covering red flag activity as well as general and contextual information about the customer and their transactional behaviour.
• Complete examples of money laundering customers and non-money laundering customers are provided to the system.
• The system uses the examples and the data to identify complex patterns indicating which customers are engaging in suspicious or honest activity.
• Every customer, alerted or not, receives a holistic financial crime risk score based on their activity which can be used in decision making across the enterprise - in perpetual KYC and credit scoring.

With features, more information is considered when assessing a customers money laundering risk, and all features are assessed and used for their ability to detect the criminal examples provided. By providing real examples of money laundering to the system, the system itself determines what money laundering ‘looks like’, rather than relying on a developers interpretation. And examples of honest customers reduce false positives since the system understands and looks for patterns which are not concerning, as well as those which are.

What’s next?

There is no doubt the technology works. At HSBC, their use of AI for AML more than halved alert volumes while doubling-to-quadrupling suspicious activity detected.

Financial institutions have the opportunity to revolutionise their detection capabilities, but this comes with a new set of challenges, such as:

• Understanding AI systems
• Developing models robustly and ethically
• Model governance
• Demonstrating risk coverage
• Integration with operational teams
• Guarding against data leakage
• Monitoring systems in production
• Future proofing systems and processes

At Groundtruth, we have experience working with risk stewards, business owners, engineers and data scientists and have successfully implemented AI solutions in financial crime across the financial services industry.

To learn about how AI could benefit your organisation, contact Groundtruth today.

Useful Links

 WSJ - Google Cloud Launches Anti-Money-Laundering Tool for Banks, Betting on the Power of AI

 Fighting money launderers with artificial intelligence at HSBC