How AI Transforms AML Thresholds to Reduce False Positives

AML (Anti-Money Laundering) thresholds are a cornerstone of risk-based approaches to combating money laundering, but they come with several challenges:

• Especially in large FIs (financial institutions), diverse and complex customer bases make granular segmentation virtually impossible
• Manually recalibrating thresholds is both labor-intensive and time-consuming
• Finally, AML thresholds are often a significant source of false positive alerts due to their generic nature

However, raising all thresholds to reduce false positives and missing genuine financial crime risk (true positives) is not an acceptable trade-off for FIs either.

The Benefits of Using AI for AML Thresholds

AI and supervised machine learning offer a different approach to AML thresholds, one that tackles these challenges and solves the sensitivity dilemma once and for all.

With AI, FIs can achieve a highly granular, data-driven segmentation of customers, forming context-specific AML thresholds. The result? Fewer false positives and more accurate detection of high-risk activity – all fully explainable and transparent for both compliance teams and regulators. 

How Do AI-Derived AML Thresholds Work?

Here’s a simplified explanation of the process, including model training, implementation, and retraining:

Model Training

To develop AML thresholds that are dynamically tailored to your customer base, an AI model must first be trained. In other words, the AI needs to learn what really constitutes financial crime risk and what does not, based on previously investigated cases. 

In this example, we’re training the machine to learn what level of cash deposits are suspicious for a given level of expected monthly income.

For example, in the customer behavior table above, a simple rule to flag customers whose cash deposits exceed 70% of their expected total income, identifies nine cases of suspicious behavior.

However, after investigation, only four of them were confirmed as suspicious (implicitly the investigators were applying a higher, 80% threshold). In total, five false positive alerts occurred. 

Using this ground truth we can train an AI model to use the investigators' decisions to improve our detection scenario and make more accurate predictions.

During the training, the AI produces a decision tree, identifying criteria that correlate with the investigators' decisions.

What did the AI learn? After analyzing the sample data, it learned to segment the customer population by income and apply tailored cash deposit thresholds as follows:

To prove that our AI model works as it should, let’s apply it to another set of customers:

So, what happened here?

While the simple >70% threshold created eleven alerts, five of them turned out to be false positives. 

Instead, our AI-derived thresholds – applying what the AI learnt from the training data – only created eight alerts, resulting in only 2 false positives after investigation. 

But wait a minute: Why do you need AI to figure this out? You could have just used traditional analytic methods to determine that raising the cash deposit threshold from 70% to 80% of total income – the threshold the investigators were implicitly applying in the training dataset –  would produce the same result. Why use AI instead of traditional methods?

The answer: While this 80% threshold might seem obvious in a simplified example with only one variable (cash deposits) and one customer context (expected total income), the real-world complexity of AML compliance is far greater. Larger FIs often deal with:

• Dozens of attributes per customer
• Millions of customers
• Hundreds of investigators

Manually identifying correlations between these variables is virtually impossible. AI eliminates guesswork by deriving precise thresholds tailored to behavioral data and real risk profiles, significantly outperforming manual recalibration. Once trained, the AI applies optimal thresholds consistently, with updates during retraining cycles (e.g., every quarter).

Implementation

Once AI models are trained, integrating them into existing AML workflows is straightforward. FIs typically start by validating AI-computed thresholds using two common methods:

• Small-Scale Testing: Applying AI-derived thresholds to subsets of the customer base
• Parallel Runs: Running AI-enhanced and rules-based approaches side by side to compare results

An effective and resource-efficient way to get started with AI-supported AML thresholds is Hawk’s AML AI Overlay solution, which is designed to give you all the benefits of AI without the need to replace existing systems.

Model Retraining

To continuously pinpoint the optimum thresholds and drive false positive numbers as low as possible, periodic retraining of AI models is necessary. Most institutions opt for quarterly retraining cycles, though other timeframes may be appropriate depending on the nature of the customer base.

Compared to month-long, resource-heavy manual recalibration processes, retraining Hawk’s AI models is fast and efficient and can be literally completed over a single weekend.

Explainability

“We didn’t know this degree of explainability is technically possible”, is a quote we’ve heard before. 

So yes, just like every part of Hawk’s AI-supported solutions, AML thresholds set by AI are fully transparent and explainable and therefore regulator-friendly. Compliance teams can easily understand how thresholds are applied and why specific alerts are generated. At a case level, teams can review:

• The exact threshold applied
• Actions that triggered the alert
• How much the threshold was exceeded
• Relevant context explaining why a particular scenario is suspicious

All details are given in natural, everyday language – no data science degree needed.

Takeaways From This Article

• Granular segmentation at scale: AI delivers granular thresholds that handle the variety of customer behavior present in large FIs – such detail can’t be achieved by traditional rules-based methods, no matter how sophisticated they are.

   

• Fast, efficient recalibration: Periodic AI retraining is significantly faster and less resource-intensive than traditional manual threshold adjustments.

   

• Regulator-friendly transparency: AI provides clear, explainable reasoning for thresholds and alerts, ensuring compliance with regulatory requirements.

   

• False positive reduction: AI can create precise, behavior-based thresholds tailored to customer risk profiles and play a key role in delivering false positive reductions of 70% or more for banks and other financial institutions.

   

Interested in learning more or have questions? Contact us today!