AI and machine learning

SEON’s AI and machine learning features help fraud and risk teams move from investigation to action quickly. Our AI tools highlight important details such as fraud scores, summaries, network patterns and rule suggestions to aid analysts in decision-making.

SEON’s former “blackbox” score is now called the AI Insights Score. This naming update reflects the intelligence and transparency behind the model. In the API, scoring engine and alert triggers, the field name blackbox_score remains until those fields are updated in 2026.

The AI Summary tool explains in plain language why a transaction or alert was blocked, approved or sent for review. It appears directly on the Transaction and Alert detail pages. The AI Summary highlights the most influential factors — such as device fingerprints, behavioural anomalies, rule matches and list hits, so analysts don’t need to manually piece together information.

This results in faster workflows, consistent decision rationale, easier onboarding and better documentation. 

To generate a summary, open a transaction or alert and click Generate. For a deeper look at how AI Summaries work and the full list of benefits, see the AI Summary documentation.

At its heart, machine learning algorithms learn from good and bad examples. Our machine learning solutions learn from Transaction labels and take states resulting from rules or manual investigations into account. Using positive and negative Transaction labels to give the most accurate feedback and rule out false positives. Provide feedback labels to SEON’s AI and Machine Learning algorithms to optimize the performance and customize to the use case you want to apply AI Insights Score and AI Rule Rules.

Learn more: Read how to provide feedback manually or through the Label API.

The AI?Insights Score is a real-time, AI-generated fraud probability score based on more than 900 data points. The score ranges from 0 to 100; higher scores indicate greater risk. It detects complex fraud patterns and anomalies that rules might miss while continuously retraining to stay up to date.

When you first start using SEON, a base machine learning model assigns an AI Insights Score to each transaction. By labeling transactions, you train a customer-specific model that eventually replaces the base model.

Click Explain beneath the score to see which signals drove it up or down.

SEON’s Advanced Digital Footprint module provides immediate email and phone fraud?likelihood scores through pre-trained base models. As soon as you start using SEON, these models calculate network scores for email addresses and phone numbers using sanitized cross-customer data. The email score considers consortium data, deliverability, domain registration and social?media patterns and has an AUC of 0.94. Set decision thresholds based on your risk tolerance (e.g., 0.85) to minimise false positives.

The phone network score uses consortium hits and carrier information and adapts as new data arrives.

SEON’s machine learning algorithms automatically generate AI Rule Suggestions. These human?readable rule recommendations are trained on your organization’s data.

Once you have 1,000 transactions in SEON (with at least 100 approved and 100 declined), the system starts producing rule suggestions. The algorithm retrains multiple times per day and assesses the accuracy of each suggested rule.

Suggested rules fall into three categories

• Complex rules: Surprising connections across many data points.
• Heuristic rules: Rules focusing on a single parameter.
• Email clustering rules: Grouping algorithmically generated email addresses.

SEON’s AML Screening Agent hit analysis empowers AML teams to rapidly sift through customer screening matches by identifying potential false positives and highlighting discrepancies. This allows analysts to make faster, more informed decisions during the review process.

• Confusion matrix: Use confusion matrices to evaluate the performance of your machine?learning models and rules. A confusion matrix summarises true positives, false positives, true negatives and false negatives and helps you adjust thresholds.
• Labels vs tags: Labels serve as ground?truth outcomes for machine learning. Tags categorise transactions for filtering or rule logic.