Anti-money laundering (AML) compliance is one of the most labor-intensive functions in banking. Financial institutions spend billions annually on compliance staff who review transactions, write Suspicious Activity Reports (SARs), and document their reasoning. Fine-tuned LLMs offer a transformative opportunity to automate these workflows while maintaining the rigor that regulators demand.
The AML Automation Opportunity
A typical AML analyst spends 60-70% of their time writing narratives and documentation, not making judgments. Fine-tuned LLMs can automate the narrative generation process, draft SARs from structured alert data, and provide consistent reasoning documentation for every disposition decision. The result is faster case processing, lower operational costs, and more consistent documentation for regulator review.
Fine-tuning is particularly well-suited for AML because the domain has specialized language, regulatory requirements, and output formats that general-purpose models handle poorly. A fine-tuned model learns the specific terminology of AML — from OFAC screening codes to FinCEN reporting requirements — and produces outputs that compliance officers can review and approve rather than rewrite from scratch.
Fine-Tuning Use Cases in AML
SAR narrative generation is the highest-value use case. AML analysts must write detailed narratives explaining why a transaction pattern is suspicious, what investigation was conducted, and why a SAR filing is warranted. These narratives follow specific formats required by FinCEN and other regulators. A fine-tuned model can generate a complete SAR narrative from structured alert data, reducing writing time from hours to minutes.
Transaction classification is another powerful use case. Fine-tuned models can classify transactions into AML typologies — structuring, layering, trade-based money laundering — with higher accuracy than rules-based systems. The model learns from historical classifications to identify patterns that rules alone might miss.
Alert disposition reasoning automates the documentation of why an alert was closed as false positive or escalated for investigation. This is critical for audit trails and regulator review. A fine-tuned model generates consistent, well-documented disposition rationales that satisfy examiner requirements.
Regulatory report generation extends beyond SARs to include OFAC compliance reports, suspicious transaction reports, and AML program effectiveness reviews. Fine-tuned models ensure these reports follow regulatory formatting requirements and include all required elements.
Regulatory Requirements for Fine-Tuned AML Models
SR 11-7 requires that any model used in banking — including fine-tuned LLMs — undergo validation before deployment. For AML models, the validation must cover model design, training data quality, output accuracy, and ongoing monitoring. The model's limitations must be clearly documented, including known failure modes for AML-specific scenarios.
FinCEN guidance on AI in AML requires that automated AML systems maintain audit trails, allow human review of all decisions, and not reduce the institution's overall AML effectiveness. Fine-tuned models must be explainable — the institution must be able to articulate why the model made a particular decision.
OCC expectations for AI compliance tools emphasize risk-based validation, ongoing monitoring, and clear governance. The OCC expects banks to understand their AI models well enough to explain them to examiners, including the training data, methodology, and limitations.
For EU institutions, the AMLD6 and EU AI Act impose additional requirements. AML models are likely classified as high-risk AI systems under the EU AI Act, requiring conformity assessment, risk management, and human oversight. Data processing for AML training must comply with GDPR Article 5 requirements for lawfulness, fairness, and transparency.
Dataset Requirements for AML Fine-Tuning
AML training data includes historical SARs, alert dispositions, transaction classifications, and investigation notes. All data must be de-identified to remove personally identifiable information (PII) before training. Data privacy requirements under GDPR and local banking regulations may restrict the use of certain data for model training.
Synthetic AML data generation is a practical approach for institutions with limited historical data. Using templates from FinCEN SAR filing guidance and public AML case studies, you can generate diverse training examples that cover the full range of AML typologies. Combine synthetic data with a smaller set of real, de-identified cases for best results.
Minimum dataset size for AML fine-tuning depends on the technique. For LoRA, 500-1000 high-quality SAR narratives and disposition examples can produce usable results. For full fine-tuning, 5,000+ examples are recommended. Quality is critical — errors in training data will be replicated and amplified by the model.
Model Selection for AML Fine-Tuning
Claude offers the highest accuracy for AML tasks but requires API-based deployment, which may not be suitable for institutions with strict data sovereignty requirements. It is ideal for institutions that can use cloud AI services and need maximum accuracy on complex AML narratives.
Llama 3.3 70B is the best choice for on-premise AML deployment. It can be deployed entirely within the bank's security perimeter, fine-tuned with LoRA, and audited by regulators. Its performance on AML tasks approaches frontier model quality when properly fine-tuned.
Mistral is recommended for EU institutions that need GDPR-compliant AI deployment. Mistral's European-based hosting and open-weight model family provide a strong foundation for AML compliance in the EU regulatory environment.
Validation and Audit Requirements
Model documentation for AML fine-tuning must include the purpose and scope of the model, training data description and provenance, methodology and fine-tuning technique, validation results including accuracy and hallucination rates, limitations and known failure modes, and ongoing monitoring plan. This documentation should be maintained alongside the model artifacts for regulator review.
Explainability for AML models requires the ability to trace any output back to the training data and model parameters that produced it. For fine-tuned models, this may require techniques like attention analysis, SHAP values, or counterfactual explanations to demonstrate that the model's decisions are based on legitimate AML indicators rather than spurious correlations.
Ongoing monitoring must track model accuracy, drift, and fairness over time. AML models must be revalidated when regulatory requirements change, when new money laundering typologies emerge, or when model performance degrades below acceptable thresholds.