Advanced AI model set to combat money laundering and financial crime

The model, named Atlas, is a self-supervised solution that learns from data without the need for human annotation. It leverages data structures to create its own training tasks. Under development since February 2024, the model is now complete and ready to deliver insights alongside other Tietoevry Banking products. 

According to Sigurd Roll Solberg, Data Scientist at Tietoevry Banking, the motivation for developing a self-supervised model – rather than a supervised one – stems from the limited availability of documented money laundering cases to use for training. 

“Known cases of money laundering are relatively few, and we know many go undetected. We needed a model that understands how financial behavior works, so we trained one capable of performing a broad range of financial tasks,” explains Solberg. 

Self-learning intelligence

Atlas can predict future behavior across a bank’s customer base and enhance existing fraud detection systems targeting card and account fraud. It also supports customer risk assessments (Customer Risk Rating, CRR) and can forecast potential interactions between customers. 

“What is particularly exciting is that the model gets better at all tasks by learning from multiple assignments. Knowledge gained from one task boosts performance in others,” says Solberg. 

The analogy to large language models like ChatGPT is clear: instead of building separate models for writing code, emails, or poetry, one foundational model is trained to read and write – making it easier to train for specific language tasks later. 

“The same concept applies to financial transactions. Once Atlas understands the financial ecosystem and transaction environments, we need significantly less data to train it for specific tasks,” Solberg adds. 

Once Atlas, our large-scale generative financial AI model, understands the financial ecosystem and transaction environments, we need significantly less data to train it for specific tasks.
Sigurd Roll Solberg
Data Scientist at Tietoevry Banking

Sharper risk assessment 

Anti-money laundering (AML) is a complex area with wide-reaching implications. Traditional systems rely on static rules set by banks to identify suspicious transactions. 

With the combination of Atlas and AI Explore data analytics platform, banks will be better equipped to filter and flag transactions, assessing them based on customer history and existing rules. 

The goal is to significantly improve risk assessments over time, including identifying patterns and emerging trends that might otherwise go unnoticed. 

“We cannot say definitively that a transaction is money laundering, but we can identify deviations from past behavior or similarities to other clients who were previously flagged,” says Anders Birkeland, Product owner at Tietoevry Banking. 

Interpreting alerts with AI

Banks follow a risk-based approach, but there is always a risk of over-investigating irrelevant cases while underestimating complex ones. 

“We wanted to explore a model that predicts the relevance of a case which helps banks allocate their investigative resources more effectively,” Birkeland explains. 

AI is also used to interpret CRR alerts by analyzing the underlying causes, offering a clearer view of a customer's risk profile based on transaction history. 

“Banks can filter by customer groups and compare risk profiles against established flags. This helps evaluate whether current risk assessments are accurate or require refinement,” Birkeland says. 

When a case handler reviews a flagged case, they can, for instance, use the AI Explore interface to assess whether the customer is part of a segment that frequently transacts internationally. 

“The aim is to efficiently answer the question: ‘Is this behavior abnormal?’ These models help provide a clear foundation for that assessment,” Birkeland adds. 

“We wanted to explore a model that predicts the relevance of a case which helps banks allocate their investigative resources more effectively.”
Anders Birkeland 
Product owner at Tietoevry Banking

Promising test results 

Birkeland emphasizes the importance of explainability in AI-driven risk assessments. 

“If a customer is flagged, the decision cannot be based solely on an AI score – you need to understand why. That is why human oversight remains essential. AI is there to guide you and suggest why a customer may be considered high-risk.” 

AI Explore serves as a standalone analytics platform but also enhances all AML applications within Tietoevry Banking. Initially launched as a decision-support tool, it strengthens existing AML workflows. Early testing indicates that AI scores correlate more strongly with flagged cases than CRR alone. 

“This is valuable insight for banks and encourages them to consider AI scores as a complement to CRR. The models we have developed will bring value across all our service domains,” Birkeland concludes.