The established applications.
WHERE BANKS ALREADY USE IT
Fraud detection, screening transactions for unusual patterns Anti-money-laundering screening Credit risk modelling Customer service, for routine enquiries Document processing Compliance monitoring
WHY FRAUD DETECTION WORKS WELL
High volume, clear feedback when a decision is wrong, and patterns that shift faster than rules can be written.
WHAT IS HEAVILY REGULATED
Anything affecting a customer's access to credit or services.
Explainability, fairness testing and human review are frequently required.
THE EXPLAINABILITY REQUIREMENT
A customer declined credit is generally entitled to know why.
"The model scored them below threshold" is not an explanation.
That constrains which techniques may be used for those decisions.
WHAT THIS MEANS PRACTICALLY
Complex models may be used to flag. Simpler, explainable models are used where a reason must be given.
FOR SMALLER INSTITUTIONS
The same obligations apply at smaller scale.
Buying a system does not transfer the obligation.
WHAT TO ASK ANY VENDOR
How does it explain a decision, and how was it tested for fairness?
