In 2020, I started looking into Anti-Money Laundering within a collaboration I had with Intesa Sanpaolo that involved real bank transaction data. I started out with what I still think is the right formulation of the anti-money-laundering problem:

How do we catch the bad guys, without bothering the good guys?

That deceptively simple question is a way of expressing the fundamental statistical tradeoff in AML: false negatives allow criminal activity to pass through the financial system; false positives consume enormous investigative resources and impose real costs on innocent customers.

At the time, the imbalance was striking. The literature I examined reported that 75–99% of AML alerts could be false positives, with important operational and social consequences.

Six years later, the field looks remarkably different due to the advent of AI. A major question is how will AI change the behavior of the adversary? If it was reasonable to assume in 2020 that the adversary had limited access to advanced ML and data science methods, now this assumption makes no sense. With a small monthly subscription, an adversary has automated intelligence that was hard to even imagine a few years ago. FATF’s 2025 horizon scan explicitly examines emerging AML/CFT risks from AI and deepfakes. The implications go well beyond deepfake videos. AI can potentially lower the cost of: generating plausible business documentation, producing synthetic identities, coordinating networks of accounts, creating personalized social-engineering attacks, adapting transaction patterns to known detection rules among many other capabilities.

That will change the economics of financial crime, sooner or later, if not already. The defender has machine learning. But now so does the attacker.

The natural theoretical framework may eventually look less like static binary classification and more like a repeated game:$$\text{Detector} \leftrightarrow \text{Adaptive Launderer}.$$

The detector changes policy. The adversary observes the consequences. The adversary changes strategy. The detector retrains. This is an online, partially observed, adversarial optimization problem. One may argue that this not a new phenomenon in AML — and indeed, concept drift has always existed and this is why the temporal aspect plays a major role in the problem. Once one corridor becomes risky, another road is used by launderers. When banks introduce thresholds for flagging transactions, criminals respond by lowering their amounts etc. However, one should expect that this type of concept drift will speed up. And that may be where some of the most interesting AML research lies.

The basic problem has not changed

Money laundering is not primarily a classification problem. A transaction does not become suspicious simply because its amount, destination country, time of day, or customer profile looks unusual.

Laundering is typically a process.

Money enters a system. It moves through entities. It may be split, recombined, converted between assets, routed through intermediaries and jurisdictions, or mixed with legitimate commercial activity. Eventually its provenance becomes sufficiently obscured that the funds can be reintegrated.

The classical terminology is:

1. Placement — getting criminal proceeds into the financial system.
2. Layering — creating transactions whose purpose is largely to obscure provenance.
3. Integration — bringing the funds back into apparently legitimate economic activity.

This immediately suggests why a detector looking at transactions independently is solving the wrong abstraction of the problem.

Suppose I see a transfer$$u\rightarrow v.$$

Its amount might be completely ordinary.

What matters may instead be that, during a short period,$$u\rightarrow v_1,\ldots,v_k,$$

followed shortly afterward by$$v_1,\ldots,v_k\rightarrow w,$$

where the intermediary accounts retain almost none of the money.

The individual edges may look harmless. The shape does not. This distinction — between anomalous records and anomalous structures — is, in my view, the intellectual center of modern AML.

Stop asking only whether a node or edge looks strange. Ask whether a collection of transactions forms a suspicious financial structure.

That view has aged very well. Elliptic2 (see also the Kaggle Bitcoin dataset) represents this important conceptual transition: from node classification to subgraph classification I find this direction particularly convincing. What investigators actually care abou is not that a certain transaction has probability 0.873 of being suspicious. They want something closer to: These 17 entities and 31 transfers constitute a suspicious movement-of-funds pattern; here is where the money entered, how it was divided, which intermediaries were used, where the paths reconverged, and why this differs from the entity’s normal activity. That is much closer to subgraph discovery plus explanation than ordinary classification. The Elliptic2 authors make essentially this argument explicitly: AML is naturally a subgraph problem, whereas much previous graph learning operated at the node level.

It is worth outlining that many consider ML a panacea and tend to replace the previous abstraction whenever a more powerful learning architecture appears. I don’t think that is the right way to think about AML. For instance, a GNN is not a replacement for graph algorithms. For example, suppose laundering produces a structure with unusually high financial flow concentrated among relatively few entities. A simple stylized objective could look like$$\max_{S\subseteq V} \frac{\sum_{(u,v)\in E(S)} w_{uv}} {|S|^\alpha}.$$

That objective explicitly encodes the hypothesis:

a suspicious group moves unusually large amounts of money relative to its size.

A learned GNN might eventually rediscover this signal. But if we already believe the signal is meaningful, encoding it explicitly provides several benefits: computational efficiency, interpretability (which is particularly important in this context to justify actions), robustness when labels are scarce, the possibility of theoretical guarantees, and an intelligible candidate-generation mechanism for investigators.

This suggests a hybrid architecture that I find more compelling than “put everything into a giant neural network”:$$\text{graph search} \rightarrow \text{candidate subgraphs} \rightarrow \text{representation learning} \rightarrow \text{risk ranking} \rightarrow \text{human investigation}.$$
The technical landscape is not the only thing that has changed since 2020. Europe has also fundamentally reorganized its institutional architecture for anti-money laundering.

The centerpiece is the new Authority for Anti-Money Laundering and Countering the Financing of Terrorism (AMLA), an EU agency headquartered in Frankfurt. AMLA was legally established on 26 June 2024, following the adoption of Regulation (EU) 2024/1620, and became operational during 2025. Its first Chair took office in February 2025, and AMLA’s 2025 work programme states that the Authority became operational in accordance with its mandate on 1 July 2025. On 1 January 2026, an important institutional transition took place: the European Banking Authority (EBA) transferred its EU-level AML/CFT mandates and functions to AMLA. The transfer included not only regulatory responsibilities but also EBA expertise, supervisory insights, risk assessments, and EuReCA, the European reporting system/database containing information on material AML/CFT weaknesses identified by national authorities. Existing EBA AML/CFT guidelines and standards remain applicable until AMLA replaces them.

AMLA’s mandate is unusually broad. It sits at the center of a new integrated European AML/CFT supervisory system and has three particularly important functions: completing the EU Single Rulebook, promoting convergence of supervisory practices across Member States, and strengthening cooperation among national Financial Intelligence Units (FIUs). AMLA will facilitate cross-border analysis and information exchange among FIUs and will operate the common FIU.net infrastructure.

Perhaps the most consequential change is that AMLA will not merely coordinate national supervisors. It will also become a direct European-level supervisor. During 2027, AMLA is scheduled to select 40 high-risk cross-border financial institutions or groups according to a harmonized risk-assessment methodology. Those selected institutions will come under AMLA’s direct supervision beginning in 2028. This is relevant to data-driven AML research as well. Historically, AML has suffered not only from fragmented datasets, but also from fragmented regulatory frameworks, risk-assessment procedures, reporting practices, and supervisory institutions. AMLA itself describes the challenge as moving away from a landscape of different national frameworks and supervisory approaches toward a more uniform European system. Greater harmonization therefore creates an interesting research opportunity—although this is an inference rather than an explicit AMLA objective. Common supervisory methodologies, standardized reporting, and increasingly coordinated information infrastructure could make it easier to formulate comparable risk measures and, eventually, meaningful cross-institution evaluation protocols and benchmarks for AML algorithms. For machine learning and graph-mining researchers, this could be particularly important: one of the persistent obstacles in AML research has been that algorithms trained and evaluated inside one institution are extremely difficult to compare with results obtained elsewhere.

Whether that opportunity will actually be realized is another question.

The FATF has also revised Recommendation 16, concerning payment transparency. The strengthened standard was agreed in June 2025. It aims to improve consistency in originator and beneficiary information accompanying cross-border payments and introduces stronger mechanisms for preventing fraud and errors. FATF is currently consulting on implementation guidance, with countries expected to implement the changes by the end of 2030. This is technically interesting because better standardized payment metadata potentially changes the information available to AML algorithms.

If structured originator and beneficiary attributes become more reliable, an edge is no longer simply$$(u,v,w,t).$$It becomes a richer object$$e= (u,v,w,t,x_e),$$

where $x_e$ may contain structured payment and counterparty information. That makes heterogeneous and temporal graph learning even more natural. Many knowledge graph systems lag significantly in terms of temporal graph learning and that is another direction that looks promising.

What lies ahead?

A modern AML system should discover suspicious structures, understand how they evolve, explain why they are suspicious, rank them according to the scarce attention of investigators, and continuously adapt as adversaries change their behavior. In other words, the future AML system is not simply a classifier, but a dynamic graph intelligence system.

References

1. AMLA
2. EBA and AMLA complete handover of AML/CFT mandates
3. The Shape of Money Laundering: Subgraph Representation Learning on the Blockchain with the Elliptic2 Dataset
4. Smurf-Based Anti-money Laundering in Time-Evolving Transaction Networks
5. AntiBenford Subgraphs: Unsupervised Anomaly Detection in Financial Networks
6. Flowscope: Spotting money laundering based on graphs
7. Autoaudit: Mining accounting and time-evolving graphs