Machine learning (ML) is an application of AI that allows machines to extract knowledge from data and learn from it autonomously. AI/ML (artificial intelligence and machine learning) and rules are decisioning capabilities that reside in the decision engine component of fraud platform solutions.
While each fraud platform solution may approach decisioning their own way, all of them have a decision engine that enables the solution to evaluate the risk of an event or transaction.
Some companies use weighted rules-based decision engines. Weighted rules work like this:
- The practitioner or the solution provider creates rules with common conditions that are associated with either bad actors or legitimate customers. An example of a condition could be “new customer.”
- Being a new customer is not risky on its own. But that might change when the “new customer” condition is connected to other set conditions, like “known bad IP” or “flagged email.”
- The practitioner adds a weighting to the condition in the form of points
- When an event or transaction is evaluated, rules that meet pre-existing conditions will deploy and assign points that contribute to the total risk score. This score is mapped to decision recommendations like approve, challenge, or reject.
Benefits of Weighted Rules
- Rule conditions are very easily understood because they support human written description fields.
- Users of these systems know exactly why an event or transaction received the score it did.
Challenges of Weighted Rules
- Rules are static, and remain unchanged until the practitioner manually updates the rule’s conditions or weighting.
- Optimizing rules can be very time consuming. Because rules execute alongside other rules that contribute to an overall risk score, it’s difficult to truly know how effective one rule is on its own, or how much that rule contributed to the final outcome of the transaction or event.
Many companies that rely on weighted, rules-based decision engines have recently added available AI components; but, the AI model score is typically combined with the weighted rule scores, and essentially acts as an additional rule.
If the rule set isn’t updated on a routine basis, the power of the AI model could be diluted. An additional challenge is overlapping rules—e.g., ML “features” (conditions that are very similar to existing, static rule conditions) that cause the same action or outcome to occur more than once. This results in overweighting, adding double the risk to the same condition, and leading to excessive manual review.
Companies that have AI-first decision engines, like Sift, take a different approach. Rather than trying to individually account for hundreds of distinct conditions and applying a fixed weighting to each one, Sift’s AI models contain upwards of 20,000 conditions. And, unlike static, weighted-rule frameworks that require human updates, these AI model features are automatically, dynamically weighted per each event, based on underlying algorithms.
Sift also separates custom rule logic from AI model logic, instead of blending rule scores with AI scores. After the AI executes and computes a score, customer rule logic can be applied to use that score alongside additional conditions, or even override the AI’s recommendation.
While overriding the AI score is not something that is typically necessary, we stand on the principle that clients of Sift should remain in full control of their decision strategy. Our platform features Clearbox Decisioning, giving users total transparency and control over artificial intelligence and machine learning models. Sift transparently exposes the logic, signals, and insights behind every decision being made, helping practitioners validate actions and align with business objectives at every stage of growth. This clarity is crucial in all industries, particularly where AI and ML decision-making materially impacts revenue and customer retention.
Schedule a demo with us to see Sift in action, or explore Sift’s features and benefits.