What is Machine Learning (ML) in AML?

Last updated: 26 October 2026

What is machine learning in AML?

Machine learning (ML) is a branch of artificial intelligence in which algorithms learn patterns from data and use them to make predictions or decisions without being explicitly programmed for each case. In anti-money laundering (AML), ML is used to detect suspicious activity, score risk, prioritize alerts and reduce the manual workload of compliance teams.

Where traditional controls rely on fixed rules, such as “alert on any cash deposit over a set amount,” ML models learn what normal and abnormal behavior look like from historical data, and adjust as that behavior changes.

Why is machine learning used in AML?

Rules-based systems are predictable and easy to explain, but they have well-known weaknesses. They generate large volumes of false positives, they are easy for criminals to learn and avoid, and they struggle to keep up with new typologies. In the FinCrime Frontier 2026-27 report, 37% of respondents reported false positive rates above 50%, and 70.8% said that 5% or fewer alerts lead to escalation or a suspicious activity filing.

ML addresses these weaknesses by finding patterns that rules miss, prioritizing the alerts most likely to matter, and learning from analyst decisions. It does not replace rules, which remain useful for regulatory minimums and for transparent, easily explained controls. Most institutions combine the two.

What types of machine learning are used in AML?

Here are the machine learning approaches and network analytics used in AML detection:

Type	How it learns	AML example
Supervised learning	From labeled examples, such as past alerts marked as suspicious or not	Scoring alerts by their likelihood of leading to a SAR
Unsupervised learning	From unlabeled data, by finding clusters and outliers	Detecting unusual customer behavior or new typologies without prior examples
Semi-supervised learning	From a small amount of labeled data plus a large amount of unlabeled data	Improving detection where confirmed cases are rare
Reinforcement learning	By trial and error, rewarded for good outcomes	Optimizing alert thresholds or investigation workflows
Graph and network analytics	By analyzing relationships between entities	Exposing money mule networks and linked accounts

Choosing a model depends on the data available, the accuracy required, how explainable the result must be, and how quickly it must run. In practice, supervised models dominate AML because institutions hold years of labeled alert outcomes, while unsupervised methods are used to find what rules and labels have not yet captured.

How is machine learning used in AML and financial crime compliance?

Transaction monitoring: detecting unusual patterns across transactions, accounts and customers, and reducing the alerts generated by static rules.
Alert prioritization: ranking alerts by risk so investigators see the most likely true positives first.
Sanctions and name screening: improving fuzzy matching and learning which near matches are cleared.
Entity resolution: linking records that refer to the same person or company across systems.
Customer risk scoring: supporting dynamic customer risk ratings and KYC refresh.
Fraud and mule detection: spotting coordinated behavior, synthetic identities and rapid movement of funds.
Investigation support: summarizing evidence and drafting case narratives, increasingly with generative AI.

What is the difference between machine learning, generative AI and agentic AI?

Machine learning learns patterns from data to classify, score or predict. It answers questions such as “how risky is this alert?”
Generative AI creates content, such as summaries and narratives, from large language models.
Agentic AI combines these capabilities so that AI agents can plan and carry out multi-step tasks, such as investigating an alert, within defined guardrails.

ML is the foundation for both of the others. Modern platforms layer the three together, with ML scoring risk, generative AI explaining it and agents running the workflow.

What are the challenges and risks of machine learning in AML?

Explainability. Regulators and auditors expect institutions to explain why an alert was raised or closed. Black box models are hard to defend.
Data quality. Models are only as good as the data they learn from. Incomplete, inconsistent or siloed data weakens results.
Bias. Models can reproduce biases in historical decisions, producing unfair outcomes or blind spots.
Model drift. Customer behavior and criminal typologies change, so models degrade without monitoring and retraining.
Validation and model risk management. Models need independent validation, ongoing performance testing and clear documentation. See the US supervisory guidance on model risk management (SR 11-7).
Governance. In the 2026-27 FinCrime Frontier report, only 14.1% of respondents said their AI governance framework is fully operational, while 61.1% have none or are still developing one.
Adoption. The same report found that 56.8% of respondents have not progressed beyond exploration or early-stage adoption of AI.

The report’s findings are indicators from 202 respondents, not statistically representative.

Machine learning in AML best practices

Start with a clear use case, such as alert prioritization, and measure the outcome.
Fix the data first. Consolidate sources and resolve entities before modeling.
Keep humans in the loop. Analysts should review model outputs and make final decisions on material risk.
Prioritize explainability. Document features, logic and results, so decisions can be explained to a regulator.
Test before and after deployment. Include below-the-line testing to confirm the model does not suppress genuine risk.
Monitor for drift and retrain on a schedule.
Apply governance consistently, in line with your risk-based approach and model risk policy.
Do not remove rules entirely. Use ML to complement rules, not to replace controls regulators expect.

Machine learning with SymphonyAI

SymphonyAI applies ML across the financial crime lifecycle. SRI Augment for Transaction Monitoring (formerly SensaAI for AML) uses predictive AI to prioritize alerts and reduce false positives on top of existing monitoring, and SRI Augment for Screening does the same for sanctions and name screening. Both are part of Symphony Risk Intelligence, which also brings together agents and unified data for Always-on Compliance. To see ML results in practice, read the Absa case study, or explore why regulators love agentic AI.

Frequently asked questions

What is machine learning in AML?
Machine learning in AML uses algorithms that learn from data to detect suspicious activity, score risk and prioritize alerts. It improves on fixed rules by adapting to new patterns and reducing false positives.

How does machine learning reduce false positives in AML?
Models learn from past analyst decisions which alerts were cleared and which were escalated, then rank new alerts accordingly. This lets teams focus on the alerts most likely to be genuine.

What types of machine learning are used in AML?
Supervised, unsupervised, semi-supervised and reinforcement learning, along with graph analytics. Supervised models are the most common because institutions hold large volumes of labeled alert outcomes.

Is machine learning allowed by regulators for AML?
Regulators generally support responsible use of ML, provided institutions can explain and validate their models and keep appropriate human oversight. Requirements vary by jurisdiction, so check local guidance.

What is the difference between machine learning and rules-based monitoring?
Rules-based monitoring applies fixed thresholds that must be updated manually. Machine learning learns patterns from data and adjusts over time, though it needs stronger governance and explainability.

Does machine learning replace AML analysts?
No. ML prioritizes and informs, but analysts still investigate, make decisions and take responsibility for outcomes.