Current limitations, the promise of artificial intelligence, and the challenges of modelling human expertise

Increasing legal and regulatory requirements in the fight against money laundering have led to a surge in reporting obligations and a strengthening of human and technological resources deployed by financial institutions. In this context, the use of augmented analytics technologies, such as machine learning (ML)¹ and deep learning (DL)², is now widely promoted as a lever for efficiency.

However, it is essential to recall that the primary objective is not regulatory compliance (compliance), but the effective identification of money laundering operations. In the absence of variables allowing for a direct measurement of detection effectiveness, it is impossible to assess the relevance of the deployed means—whether legal, technical, or operational.

This methodological gap is all the more concerning as cases of non-detection, revealed during investigations into other offences, also affect institutions fully compliant with their obligations. This situation leads to two main issues:

Financial institutions are held accountable for a perceived lack of rigour and respond by adopting excessive formalism, generating adverse side effects such as procedural complexity, increased delays, or unnecessary training.
Regulatory accumulation continues, even though none of these new rules can be correlated with improved effectiveness.

The conceptual shift – from operational performance to procedural compliance – has already been observed in other areas of combating integrity violations, such as corruption or influence peddling. The European research project ANTICORRP (2013–2017) clearly illustrates this:

“Simple bivariate relationship between control of corruption and the respective scores [of public accountability mechanisms] indicates that the extent of regulations is not significantly associated with better controlof corruption […] on the average, countries with less control of corruption tend tohave a more comprehensive and strict regulatory framework.“(Mungiu-Pippidi & Dadašov, 2017, p. 394).

Contributions of machine learning in anti-money laundering: A conditional promise

Reite, Karlsen, and Westgaard’s article, “Improving client risk classification with machine learning to increase anti-money laundering detection efficiency“ (2024), examines the contributions of ML and DL in the American financial sector. Faced with rapidly evolving threats, they argue that integrating these technologies is a necessary response to the increasing burden of compliance obligations. These tools enable faster processing of large datasets and facilitate adherence to reporting requirements.

Furthermore, the underlying ambition is that these algorithms will identify subtle patterns or anomalies that may reveal suspicious behaviour. However, this hypothesis remains largely speculative to date.

Current Limitations

Beyond the authors’ observations, it is crucial to emphasise that the hopes placed in AI have not yet translated into tangible improvements in detection performance. Entire segments of money laundering activities – particularly those linked to drug trafficking, cultural goods, or artefacts – are only detected following the identification of the perpetrators of these crimes. With significant variations depending on the country, investigative services, and further criteria, three factors warrant particular attention :

Prosecutions for drug trafficking are not systematically followed by prosecutions for money laundering, despite the intrinsic link between these activities. This may be explained by the specialisation of investigative services and an implicit prioritisation of prosecuting the primary offence, with money laundering perceived as secondary.
Even when the injection point³ of funds to be laundered is identified, reconstructing the laundering process (actors, layering, and extraction methods) remains challenging.
Investigations into money laundering primarily begin with human intelligence. While transactional data used by financial institutions are useful for reconstructing laundering processes, they appear ineffective in detection.

Building on Reite, Karlsen, and Westgaard: Methodologies, challenges, and perspectives

The authors briefly mention a central point: the importance of expertise acquired by financial sector professionals. This expertise, often intuitive, is grounded in a nuanced understanding of the field, extending far beyond objective data alone.

Can an AI model be trained to replicate the expertise of these professionals?

Methodologies for acquiring and modelling this expertise exist, including qualitative studies through interviews, structured analysis (e.g., analysis of competing hypotheses, morphological analysis), and the integration of these insights into AI models. However, their implementation faces obstacles: difficulty in identifying reliable experts, confidentiality constraints, and professional risks for participants (Labic, 2024).

The European project ENSEMBLE proposes a framework for leveraging advanced technologies within a consortium closely linked to expert investigators from multiple European law enforcement agencies. While accessing experts from exposed financial institutions remains a challenge, ENSEMBLE offers a modular toolkit combining machine learning, graph analysis, and collaborative mechanisms (such as federated learning or the MISP platform) to correlate heterogeneous data. By cross-referencing these technologies with qualitative investigations involving investigators and industry professionals, it becomes possible to capture and amplify intuitive expertise while making it scalable.

This synergy between qualitative methods and technological tools could address the challenges raised by Reite, Karlsen, and Westgaard:

Interviews could enrich AI models with field knowledge often absent from pure transactional data.
ENSEMBLE’s tools (crypto-asset tracing, dynamic incident reconstruction, interactive training) provide a framework for structuring and analysing these insights while ensuring secure sharing between institutions.
Training investigators through hybrid modules (serious games, asynchronous learning) would facilitate the adoption of these technologies while preserving professional critical judgement.

Such an approach, if deployed, could transform current limitations into opportunities: human expertise would guide algorithms toward more relevant signals, while ENSEMBLE’s tools could detect complex patterns invisible to the naked eye. Preliminary results from the project, particularly in detecting illicit flows and cross-border collaboration, suggest that this synergy is not only possible but necessary to move beyond the sterile opposition between procedural compliance and operational effectiveness.

Written by Paul Labic. Laboratory for theoretical and applied economics (BETA), CNRS UMR 7522. Associate researcher at the research lab of the French police academy (ENSP)

REFERENCES

Labic, P. (2024). Interroger la corruption – Menaces liées à la collecte et à l’analyse. In Faire face à un terrain sensible (pp. 95–106). L’Harmattan.

Mungiu-Pippidi, A., & Dadašov, R. (2017). When do anticorruption laws matter? The evidence on public integrity enabling contexts. Crime, Law and Social Change, 68, 387–402. https://doi.org/10.1007/s10611-017-9693-3

Reite, E. J., Karlsen, J., & Westgaard, E. G. (2024). Improving client risk classification with machine learning to increase anti-money laundering detection efficiency. Journal of Money Laundering Control, 28(1), 93–107. https://doi.org/10.1108/JMLC-03-2024-0040

NOTES

Machine Learning (ML) is a branch of artificial intelligence focused on developing algorithms that learn from data, improving their performance as they are exposed to more information. It is widely used in classification, regression, and clustering applications. ↩︎
Deep Learning (DL) is a subset of ML that employs deep neural networks to model complex patterns in data. These networks are particularly effective for processing unstructured data such as images, videos, and text, enabling systems to learn multi-level data representations and perform complex tasks like image recognition and natural language processing. ↩︎
The general model of money laundering comprises three phases: placement (injection), layering, and extraction. Layering can be iterative, with the ultimate goal being the full enjoyment of laundered funds (or sanctification). Money laundering should not be viewed solely as a cost but also as a profit multiplier (e.g., real estate speculation, investments). ↩︎