AI · fraud detection
Readable rules generated by a model
A learning model that produces auditable rules, alongside a market engine; the false-positive ratio is set by management. Fraud reduced by 20 to 30% on historical data.
Context
An online booking platform processed thousands of transactions a day. A market anti-fraud engine blocked part of the fraud but let new patterns through and blocked too many good customers. Management wanted to decide the trade-off between tolerated fraud and refused customers itself.
A model that writes rules, not a black box
Rather than an opaque score, the learning model produces readable rules: combinations of signals (country, payment method, timing, history) an analyst can read, challenge and disable. Every rule is auditable.
The false-positive ratio set by management
The cursor between accepted fraud and refused good customers is a business decision, not a technical parameter. Management sets it; the model conforms. The result is measured on history before any go-live.
Alongside, not instead
The market engine stayed in place for what it does well; the model covers what it misses. Both results are logged and compared, which measures the real contribution of each.
Result
Frequently asked questions
Why rules rather than a score?
Because a rule can be read, challenged and disabled. An opaque score neither explains a refusal nor justifies it to a customer.
How was the figure measured?
On the transaction history, by replaying the model on past periods and comparing with what the existing engine had blocked.
Does the model adapt to new fraud?
Yes, through periodic retraining on new transactions and analysts’ decisions. New rules are reviewed before activation.
Related services
A comparable case on your side?
Tell me which process costs you time; I get back within 24 business hours with a first read, no obligation.