The scale of illicit money circulating around the world’s financial system is almost beyond comprehension. Some $3.1 trillion a year of dirty money is laundered annually through the world’s banks and other institutions, according to the Nasdaq Verafin Global Financial Crime Report. This includes an estimated $782 billion from drug dealing, $346 billion from human trafficking, $485 million from scams and frauds – and possibly up to €11 billion to finance terrorism. Such huge sums, despite three United Nations conventions, in 1988, 2000 and 2003, that provide frameworks for global efforts to stamp out money laundering.
Those efforts have clearly failed. But now, at last, artificial intelligence (AI) can help us turn the tide.
Historically, banks and other financial institutions have largely relied on hundreds of rules built into their money transfer systems to track customer behavior, triggering alerts about any dubious transactions. When there is a suspicion of money laundering these trigger alerts that can be investigated by compliance officers, who can either approve or block the transaction.
Many of us have encountered AML rules when innocently moving funds to pay for a child’s education abroad, repatriating a legacy and so on. But so have the criminals. And therein lies the problem. The rules and parameters are mainly designed to stamp out the kinds of money laundering exposed by past investigations.
Rules-based payment monitoring systems are programmed to red-flag all funds over X amount, moving to Y country, from Z kind of account. Three rules. But over time the criminals wise up, shuffling money across multiple accounts, using smaller denominations or luring harder-to-spot student ‘money mules’ into laundering cash for pocket money. Three rules become 70 or 100, and compliance officers are overwhelmed by checks to perform. Thus, while legitimate transactions get wrongly red-flagged because they seem suspicious, the actual money laundering activities tend to remain under the radar.
The rollout of instant payments, though great for customers, has exacerbated the problem. The speed and scale of these instant transactions has turbo-charged the AML challenge.
Artificial intelligence offers a new, complementary way for banks to identify money laundering in the transaction system. We explored this shift in depth in our recent webinar, AI in AML: Navigating the New Frontier in Banking Compliance. These two approaches, rules-based and AI, when running in tandem, can identify money laundering more accurately, while reducing the number of false positives and the costs of compliance. How so?
Traditional systems work by highlighting specific transactions that contravene a set of rules, which experience has shown to be markers of possible money laundering. You can learn more in our detailed guide on all about AML transaction monitoring. But these rules-based systems miss transactions that are, for example, above or below the specified parameters. And, of course, such systems raise red flags when regular customers do something out of the ordinary, like transfer an inheritance to relatives abroad. The result is that numerous false positives have to be painstakingly sifted out by large teams of compliance officers before dodgy transactions can be blocked and reported to regulators.
The need for human reviews also limits the effectiveness of these systems. Criminals may strive to get around the parameters, for example, by engaging in many more, smaller transactions. But if parameters are lowered, the number of cases for review becomes unmanageable.
AI models can identify atypical behavior outside AML parameters by flagging the behavior of particular accounts over time. They can swiftly process a much broader range of signals, uncovering subtle anomalies that could point to money laundering. And they do so with great accuracy.
As regulatory compliance requires banks to flag specific transactions, AI cannot replace existing methods. Rather, it can be used in tandem, in either of two ways. The first is to use an AI model to review red-flagged transactions to identify which are most likely to be false positives. This allows compliance teams to prioritize the transactions that are most likely to be criminal. The second is to first use AI to identify accounts that may be party to money-laundering transactions. The transactions of these accounts can then be reviewed using parameters to identify any likely criminality. The advantage of this second method is that it enables banks to explain to regulators which transactions the AI model has identified and why.
Both of these strategies enable compliance departments to focus their teams on the most suspicious transactions – the true positives. This is revolutionary. It has the potential to increase the proportion of dubious transactions identified and reported to regulators for further action, while also reducing the burden on the compliance team by significantly lowering the false positive alerts.
Using AI models in tandem with rules-based models can simultaneously reduce the investigatory burden and help to better identify and catch the most active criminals.
Combining AI models with parameter-based rules helps the AI learn from its own hits and misses. This will steadily improve its accuracy and reliability. But it will also learn to spot new patterns of criminal behavior in real time.
Initially, it makes sense to deploy this technology alongside rule-based systems. However, in the long run, artificial intelligence will fully diverge from rules and ultimately render them entirely obsolete.
For example, suppose a bartender deposits their week’s tips as cash every Tuesday morning. Then suddenly the account shows eight or 10 deposits in a month – they may have become a money mule. The AI model will spot the new trend and flag it.
The two AI for AML risk models we have developed at NetGuardians analyze and report transactions of a given customer account over a given period. By comparing what has happened over past periods, or to a peer group, they generate ‘hits’ on these accounts when they spot a dubious pattern.
This works well, but to be of practical use the AI has to be able to scan a day’s worth of transactions by the 500,000 customers of a typical bank within an hour, as the bank does its nightly transaction processing.
So, the overriding challenge becomes to work out the precise set of transactions that together form the potential money laundering activity while discarding the legitimate or usual ones from the alert. We have developed a groundbreaking heuristic aimed at solving this combination puzzle. It is a unique, reliable and effective AI tool for AML applications.
The magic of the system is that we let the model decide which are the transaction features that constitute a new fraudulent pattern. As experience with AI improves, banks may be able to lower the thresholds that trigger alerts, enabling compliance teams to block more small-scale money laundering while at the same time reducing further the number of false positives as the system improves its understanding of actual money laundering signals.
This is good news for banks, economies and account holders. But most of all, it’s a boon for the world we live in as we leverage AI to fight financial crime and safeguard customers. It’s past time we started to reduce the money-laundering flows of the evildoers preying upon the most vulnerable in our societies.