Dr Martin Lukavec, Senior Lecturer at London School of Business and Finance, explores how AI can trace money laundering activities for a more secure future.
Money laundering is the process of cleaning dirty cash. Criminals earn it through drugs, cybercrime, corruption or other assorted rackets, then disguise it as the fruit of honest trade.
They cannot simply deposit sacks of cash, so they route it through shell firms and fake invoices until it appears to be income earned by a hard-working enterprise. A few creative spreadsheets later, the money is respectable enough to buy a yacht.
The first step is entry. Large sums are sliced into small, unremarkable payments. Restaurants, logistics firms and even charities can serve as a front, mixing real money with the tainted. In the digital age, the toolkit has expanded: prepaid debit cards, peer-to-peer apps, and crypto wallets now serve as quiet gateways into the financial system.
When cash won’t do, the funds zigzag across fintechs, online lenders and offshore intermediaries – constantly changing owners and addresses until the trail evaporates.
Banks sift through millions of daily transactions to find the few that reek of deceit. The task is like spotting a droplet of soapy water in the Thames. Most of the time, they don’t notice it.
The rise of the machines
Artificial intelligence, while long talked about, is finally proving its worth. It promises to replace rigid rules with pattern-seeking logic.
By pooling payment data, account details, devices, merchant histories and geographic markers, AI creates a panoramic map of financial behaviour. Traditional systems see dots, AI connects them into shapes. It learns what “ordinary” looks like and flags what isn’t – sudden bursts of transfers, curious timing, or clients who seem oddly well-connected.
Crucially, it views the financial world as a network, not a collection of isolated accounts. Money laundering depends on cooperation – chains of firms, intermediaries and digital wallets passing money like a hot potato. AI can trace those hand-offs in ways human analysts never could.
Mapping the invisible
Graph analytics turn data into a living map of relationships. Each entity – a customer, account, or device – is a node, each transaction a link. Certain shapes tell stories: triangles of quick-fire payments, loops that bring funds back home, or clusters that light up when one participant turns dirty. Risk radiates outward, guilt becomes a statistical probability rather than a hunch.
Field tests suggest that such network approaches catch more illicit flows than rulebooks have ever managed to do.
The data problem
However, even the most sophisticated algorithms stumble on messy inputs. Most alerts are never confirmed, and official reports capture only a small share of real money laundering. In effect, models must learn from rumours, not proof. To compensate, researchers use “positive-unlabelled” learning, treating known crimes as confirmed and everything else as suspicious, or “weak supervision,” which blends partial clues from different sources. Graph neural networks (GNNs) go further, learning directly from patterns of interaction and spotting trouble even in uncertain data.
Early evidence is encouraging. GNNs can flag risky networks with surprising accuracy, even when the truth is fuzzy. In an industry where certainties are rare, progress is measured by probability.
Cat and mouse
Still, no algorithm will end laundering. Criminals adapt quickly, designing transfers that look tediously normal. The contest between regulators and wrongdoers persists, as it has always done.
AI won’t clean up global finance, but it will raise the cost of cheating and shorten the crooks’ lead. For now, the machines are learning faster than the money launderers – but probably not for long.
