Machine Learning in KYC and the Amplification of Social Trade-offs

Abstract

Context and Problem Statement As financial crime grows more complex, governments have tightened Anti-Money Laundering (AML) regulations, with banks serving as the first line of defense. The Know Your Customer (KYC) process, used to identify and monitor clients, is central to this effort. To keep up with the resource intensive process, banks are beginning to turn to machine learning (ML), hoping to improve detection while lowering costs. However, this shift raises new concerns: opaque algorithms may introduce bias, reduce transparency, and unintentionally exclude vulnerable groups from the financial system. In the Netherlands, recent reports have shown how well-intentioned AML measures can lead to discriminatory practices. As the EU prepares to implement the new AML Regulation (AMLR) in 2027, which will further tighten compliance. While policy and technical discourse continues, there is little understanding of how these changes affect society as a whole. Research Gap and Question The literature showes that while AML policies and the integration of ML have been widely studied from regulatory, technical, and institutional perspectives, their broader societal effects remain underexplored. Most research focuses on costs and benefits for banks and governments, mostly neglecting the indirect impact on society. This thesis addresses that gap by systematically analyzing how ML integration into KYC processes shifts the societal costs and benefits of AML compliance. While most evaluations of ML in finance focus on technical or regulatory performance, this research applied a qualitative Social Cost-Benefit Analysis (SCBA) to examine broader social consequences such as exclusion, inequality, and institutional trust. By analyzing current practices and exploring how ML changes these dynamics, the study contributes to a more balanced understanding of innovation in financial compliance. The main research question was therefore formulated as follows: “How does the implementation of machine learning in the KYC process alter the relevant social costs and benefits of the current KYC process under AML regulation?” Methodology This thesis applies a qualitative SCBA framework (Romijn and Renes, 2013) to explore the societal effects of AML regulation through the KYC process, with a specific focus on the impact of ML. The research followed the first five SCBA steps as outlined in Dutch policy guidelines (Romijn and Renes, 2013), including problem definition, establishing a baseline, defining the policy alternative, identifying social effects, and identifying costs. The social effects and costs were not quantified in this research due to the qualitative nature of the identified effects. Due to data limitations the final three SCBA steps were excluded. Primary data was collected through 9 semi-structured interviews with experts from the corporate sector, academia, and human rights advocacy. Interviewees were selected based on their expertise in AML, KYC, and/or ML, and their ability to reflect on broader societal implications. A preliminary KYC process diagram and stimulus texts based on literature were used to structure and enrich the interviews. Interview data was analyzed using thematic analysis in ATLAS.ti, following the six-step method of Willig and Rogers (2017). Emerging codes were grouped into themes and translated into nine key social effects, e.g. inequality, crime reduction, and consumer surplus. These effects were then visualized in a conceptual model showing their interrelations and the impact of ML. Triangulation with existing literature was used to enhance validity. Key Insights The following key findings highlight the most important insights regarding the social costs and benefits of the KYC process and the impact of machine learning within AML compliance. • The KYC Process Has Dual Social Impacts ii iii This thesis examined the social costs and benefits of KYC processes under European AML regulation. Using thematic analysis of expert interviews, the study identified a wide range of interconnected social effects. The candidate positive effects include reduced crime and improved government financial health. Identified potential negative effects include increased administrative burdens on consumers, reduced economic participation, and rising inequality. • Machine Learning Amplifies Both Benefits and Harms The introduction of ML into the KYC process intensifies many of these effects. For the benefits, ML can improves the detection of financial crime and may enhance the efficiency of public enforcement and private compliance. The identified harms that ML can raise were opportunity costs due to legal uncertainty, introduces privacy risks, and may increase stress among consumers. • The KYC Process can Intensify Social Harm and Unequal Burden Distribution A central finding is that society bears the weight of the costs, even though financial institutions and governments control the design and implementation of AML systems. ML expands the surveillance function of banks, blurring the line between private and public roles and potentially reducing accountability. Vulnerable groups can be disproportionately affected through exclusion, algorithmic bias, and stress. This points to a structural imbalance in who benefits from technological innovation and who suffers from its unintended consequences. Conclusions This thesis explored how machine learning reshapes the social costs and benefits of KYC processes under AML regulation. Based on expert interviews and qualitative SCBA, the study found that ML amplifies both the benefits, such as improved crime detection, and the risks, including legal uncertainty, exclusion, and inequality. The effects are interconnected, with societal consequences often falling disproportionately on the public. By framing KYC as a socio-technical system, this research highlights the need for ethical oversight and systems thinking. ML introduces not just technological change; it also creates a responsibility to ensure that its effects are fair and beneficial to society as a whole. Actionable Recommendations This study highlights the need for a more balanced and accountable approach to integrating machine learning in AML compliance. Policymakers should move beyond a narrow enforcement focus by embedding social impact assessments into AML policy design and collaborating with academic experts. Regulation must ensure that ML systems are not only legally compliant but also socially fair and explainable. Financial institutions must take responsibility for the real-world effects of ML in KYC, prioritizing explainability, bias mitigation, and transparent vendor oversight. Close engagement with regulators and researchers is key to setting fair and practical standards. Finally, society, though not directly involved in implementation, bears many of the negative impacts. Public awareness, advocacy, and particip