Anti-Money Laundering (AML) professionals are looking for ways to tackle the rising threat of global money laundering and terrorist financing attempts more effectively and efficiently. Strong AML and counter terrorist financing (CTF) programmes require robust industry collaboration, to develop defences that protect individual financial institutions as well as societies and the global economy.

June 14-15 saw the ACAMS Europe chapters come together in Brussels to discuss best practice, new industry developments, and collaborate across the AML practitioner community. What was clear at the event is that the sheer scale of money laundering attempts requires more effective identification of typologies to reduce wastage associated with investigating false positives.

Does Crime Still Pay?

During their keynote, Burkhard Mühl, Head, European Financial & Economic Crime Centre (EFECC) at Europol, emphasised the growing challenge of technologically savvy criminal networks. “The scale and complexity of ML activities in the European Union (EU) have been previously underestimated,” explained Mühl. According to Europol Public Information Stats:

  • 80% of the criminal networks active in the EU use legal business structures for their criminal activities – showing their infiltration of the legal economy
  • 68% of the criminal networks active in the EU use the services of professional money laundering syndicates
  • 60% of the criminal networks active in the EU use corruption as a preferred way of conducting businesses

Europol previously estimated annual turnover from money laundering in the EU to total 110 billion Euros in 2016, but in the subsequent years of investigation the actual total appears to be much closer to 140 billion Euros. 2.2% of this was seized and 1.1% ultimately confiscated, according to the Europol report: Does Crime Still Pay? Criminal Asset Recovery in the EU.

Money Laundering and Cryptocurrency

According to Europol’s research, cryptocurrency is increasingly leveraged within money laundering activity, with the criminal use of cryptocurrency no longer primarily confined to cybercrime activities, but now related to all types of crime that require the transmission of monetary value. Use of cryptocurrencies as part of crime schemes is increasing, including in money laundering typologies. Now is the time for financial institutions to get ahead of the impending challenge of crypto crime, given that the current volume and value of cryptocurrency transactions related to criminal activities still represent a limited share in the criminal economy.

Using Technology to Improve Transaction Monitoring

I joined a panel of leading AML experts alongside Andrea Bielska, Head of Financial Crime Audit, Nordea; Dan Benisty, Head of Compliance Northern Europe, Western Union; Nitzan Solomon, Head of Transaction Monitoring, AML & Fraud, Revolut; and Christophe Ferrand, Head of Group AML, Société Générale, to discuss turning technology to our advantage as AML practitioners.

The conversation covered creating transaction monitoring protocols to detect and resolve potential indicators of fraud such as atypical spending or anomalous movements or transfers of assets; capturing efficiencies and effectiveness of emerging technologies such as artificial intelligence to strengthen transaction monitoring and enhance Anti Financial Crime (AFC) discipline across enterprise; and crafting formal response processes to transaction monitoring alerts to ensure comprehensive reviews of red flags and standardize expectations in areas such as escalation of investigations.

Real-Time AML and Machine Learning

The speed of financial flows is increasing across borders, and digital assets and real-time payments are on the rise around the world. We as consumers are so used to our real-time digital experiences we don’t even realise it has become the new norm. This real-time ecosystem is supported by advances in cloud computing, and new standards in data exchange through open banking and APIs. This has enabled transaction monitoring now to sit inside the payment flow and provide an immediate risk assessment at the point of processing. So that means transactions can be blocked, held, forwarded for review, or escalated as they occur, or in the case of fraud can be interrupted before the funds are cleared. Real-time or same-day processing in fraud systems (which tend to operate faster) has been known to be very successful at identifying money mule activity for vulnerable populations, and hence stopping customers from becoming an unwitting accomplice to money laundering. But for many AML teams their data is still processed in batch, rather than real-time. This inhibits AML practitioners’ ability to action insights from their fraud teams into their own strategies. There is a growing realisation that for certain crimes like terrorist financing, human trafficking or modern slavery, a timely reporting and speedy intervention by relevant authorities can literally save lives.

Nitzan Solomon emphasized the importance of applying machine learning techniques against a clear data strategy within your organization. “I’d encourage shifting the dialogue from a binary ‘Artificial Intelligence (AI) is good or bad’, to the collective recognition that it’s a very wide spectrum of tools and capabilities to be used across a variety of use cases, with very different risk and reward for each. From what I’ve seen, using AI or Machine Learning (ML) in-house doesn’t work very well if the organizational culture, data, and technology strategies aren’t aligned. If yours is the only department applying ML and your organization doesn’t have a good understanding of the required tools, talent, etc. then it’s not impossible, but it’s much harder to succeed.”

Converging Fraud and AML Insights

Christophe Ferrand explained that his team at Société Générale, is already successfully utilizing the capabilities of machine learning Transaction Monitoring for fraud detection, and has identified how the convergence between both fraud and AML can be effective.

“We uncovered a complex international fraud scheme. This scam used an organized laundering structure linked mostly to a fake investment scheme. After further investigation including network analysis, we have been able to detect flows involving shell companies going through several layers of banks.

There is clearly some convergence between fraud detection and AML monitoring as fraud is a predicated offence. It is especially the case when some red flags are common, such as companies with no apparent links but sharing the same physical or IP address, or the same directors. However, we are not looking exactly for the same offence as fraud teams, and AML investigations must go much further than fraud.”

Real Life Real-Time AML Operations

Nitzan highlighted that at Revolut the key to success with machine learning in AML has been to continuously self-test and benchmark models. And that the Revolut team does not bucket all AI and machine learning together, but rather focuses on understanding the differing techniques that deliver high explainability and provide clear rationales for AML flags and decisions.

And Dan Benisty detailed how the Western Union AML team uses automation in the filtering of false positive alerts and alert triage: it has been using alert closing for sanctions and Politically Exposed Persons (PEPs), but not yet for transaction monitoring. Automatic alert closing was one of the few areas where the panel really held diverse views on team and technology readiness, with opinions ranging from the belief that it will never be possible and AML examiners will not accept it, through to others believing that it may happen, or that there are other ways to generate the same effect with scoring alerts and setting thresholds.

Challenges in Real-Time AML

That is not to say there aren’t considerable challenges for real-time transaction monitoring, a payment request alone contains insufficient information for monitoring, and hence in order to produce a significant risk decision the transaction monitoring needs to be able to look at the whole customer profile to identify any anomalous pattern. For fraud those lookback periods are typically a lot shorter, whereas AML transaction monitoring is all about historical pattern analysis. Evaluating all of that customer intelligence whilst providing a risk evaluation within milliseconds is a technological challenge.

Despite this, real-time machine learning is gaining traction. “Machine learning techniques are gaining traction in the compliance word,” explained Dan. “Machine learning can be valuable in improving the efficiency and effectiveness of AML program. In most cases, machine learning is augmenting rather than replacing traditional transaction monitoring methods. Indeed, I think that the use of machine learning helps to give more insight to AML investigators and prioritize risk-based investigation.” Dan emphasised pragmatism, in that implementing machine learning into ‘greenfield’ AML operations can be a cost and time intensive project, and therefore teams must be clear on their goals for the technology.

Regulating AI Innovations in AML

Regulatory frameworks will become increasingly important in the successful application of machine learning for AML teams. Both Dan and Christophe explained that regulators are actually using ML already, including during on-site examination. In fact, Autorité de contrôle prudentiel et de résolution (ACPR, the French Prudential Supervision and Resolution Authority) is currently hosting its first tech sprint on the explainability of artificial intelligence. It’s not necessarily the case that regulators are inhibiting AI innovation, Andrea Bielska noted, “sometimes it is harder for models to pass internal model validation than it is to get regulatory acceptance.”

In fact, even regulation pertaining to AI and machine learning more widely seems broadly accepting of its importance for AML operations. As Christophe noted in relation to the upcoming EU AI Act, it will be interesting when it becomes clear how AML is considered within that regulation. It seems likely that banks’ Transaction Monitoring Systems will fall under the ‘acceptable risk’ category for the use of machine learning. Christophe also highlighted the differences between regulator and Financial Intelligence Unit (FIU) attitudes to- and motivations for- machine learning applications. Because of this FIUs are using machine learning when reviewing or scoring suspicious activity reports (SARs) even in geographies where the regulator may not yet take a public and positive stance towards machine learning in AML.

The Future of AI for AML

One thing I am sure of though is that in ten years’ time ACAMS will not host a panel to discuss whether machine learning has a place in financial crime detection – maybe we will be talking about hologram scams, or Web 4.0 asset classes. Or as Dan joked, “maybe in ten years robots will host a panel to discuss whether they still need humans to do AML investigations.” But jokes aside, the statistical techniques used by data scientists in machine learning are mature and analytically sound and fall under a higher level of scrutiny than historic decision models have been subjected to. We also see the wider risk community making active use of machine learning, whether that’s in cyber security, card or payment fraud, or credit risk.

Of course, technology is never the complete answer to any challenge, but it is a tool for AML practitioners to achieve specific goals they are setting. And I would encourage anyone in our community to be very clear and ruthless in defining their AML goals. Statistical techniques are great in that they encourage us to take an outcomes-based view on our AML detection regime. The goal always has to be in line with our primary purpose as AML practitioners – and that is to stop financial crime, and anything that can support us in spending less of our time and focus on activity that turns out to be non-criminal. As such automation serves one of three purposes – to increase capacity, provide time savings and help treat backlogs. We are on a journey to the industry-wide adoption of machine learning, reaching ubiquity in the next 5-10 years, meaning it is imperative not to be left behind in the accelerated adoption of AI for AML.

Discover more about machine learning for AML practitioners, read the white paper: AI Innovation in AML Transaction Monitoring