AI implementation in AML at HSBC sees a considerable reduction in compliance costs

Current AML Techniques

In today’s world, we tend to combat money laundering efforts with regulation-based account surveillance and ordinary KYC systems, when in all honesty these techniques will not hold up in the ever-changing world of financial delinquency. For one, the Risk Based Approach (RBA) used by financial institutions for AML compliance purposes causes a high rate of false positives. Due to the introduction of Enhanced Due Diligence (EDD), focus on customers that are truly high risk has increased, and false positive rates have consequently gone down. But it is still too high and is a waste of significant resources. Another issue with these outdated AML techniques is that they aren’t capable of learning and getting smarter. In fact, they only find issues that they are specifically looking for. They have no way of predicting new scenarios and only rely on data and information that has been learnt of in the past. Finally, current AML compliance techniques are extremely expensive and still seem to fail on a regular basis. We’ve all seen very large banks having failed AML processes, leading up to enormous regulatory fines and legal issues. These banks are spending a lot of money employing hundreds of people to monitor transactions and investigate accounts, but they do this based off of pre-defined rules and patterns. AI can do away with this and with machine learning implementation, unique, more complex patterns can be detected and analyzed.

Introducing AI to AML Efforts

AI based AML systems can detect patterns and trends that haven’t been seen before.  Detection of previously unknown patterns will inevitably lead to the discovery of more convoluted money laundering patterns and can be used for reference in the future. Through continuous training, a machine learning based system can look at its past analyses and use that information to better prepare itself for future money laundering threats.

Also, AI can reduce the false positive rate discussed earlier. This can lower costs of compliance and allow investigators to dedicate more time to customers deemed high-risk. When HSBC hired a firm to implement AI in AML investigations, they saw the number of investigations drop by 20% without reducing the number of cases referred for more scrutiny. Hopefully, the promise of a more cost-effective AML compliance system will motivate other financial institutions to implement AI in their AML/KYC techniques.

The automation of compliance based tasks will allow for analysis of data in extremely high volumes and high frequencies without the need for over-hiring of employees.


Machine Learning (Unsupervised)

Machine learning has been a proposed AML aide for as long as the phrase artificial intelligence has been thrown into the mix. It is imperative however to understand the difference between supervised and unsupervised machine learning to comprehend the potential that unsupervised machine learning could have on the AML landscape. To briefly differentiate between the two:

  1. Supervised Machine Learning – This is the regular, more common form of machine learning where the software is provided with the data, the objective and expected output of the data. The software is then permitted to identify algorithms and patterns to get to the predicted outcome.
  2. Unsupervised Machine Learning – In this less adopted form of machine learning, the AI software is provided with the data and the objective but not the expected output. This allows the AI to detect never before seen patterns and anomalies in the data. The AI can then learn from this data, adapt, and become increasingly more efficient at its job.

This will be a huge proponent in reducing the false positive ratio that we discussed earlier. It does this by observing anomalous data patterns such as fraudulent transactions without knowing before that moment what a fraudulent transaction is. It can ensure that the activity is indeed illegal by identifying the causes of this activity through patterns that are much too complex for a rule-based AML system.



Clustering is a technique adopted in BIG DATA analysis to find data groups that have similarities between them. In an AML system, an AI software can be used to form clusters based on customer behavioural patterns. Each cluster will have its own level of risk and customers that do not naturally fall in any clusters would be considered suspicious immediately. The further, a customer drifts from a cluster, the higher their risk rating would be, allowing for these ratings alone to be used as basis for investigative compliance.

The Future of AML Systems

It is evident that AI based AML techniques have already proven to be effective at reducing compliance and legal costs. As AML regulation gets tighter and more demanding, risk-based/rule-based AML systems will become obsolete because the cost would easily outweigh the benefit. AI systems that can learn about new patterns and automate account surveillance will no doubt be more cost effective. With banks like HSBC already investing in these technologies and reaping the rewards, it will be interesting to see who follows, in this constantly shifting landscape of AML compliance.