Rules engine

In the mid-1970s at Stanford, the MYCIN expert system recommended treatments for bacterial infections from a few hundred “if ... then” rules written with infectious disease specialists, and every recommendation could be justified by walking back up the chain of rules that produced it. The business rules engine carries that principle into insurers' policy and claims systems. Experts write explicit conditions (if the loss occurs less than thirty days after inception, if the amount exceeds a cap, if the driver is under twenty-one and the vehicle has more than 150 horsepower), and the engine evaluates them on each file to accept, decline, refer to a human or raise an alert. In underwriting it carries the acceptance guidelines and authority limits; in claims it forms the first layer of fraud detection, ahead of or alongside a learned model. Its strength is traceability. Each decision cites the rule that triggered it, which is what an auditor, a regulator or a policyholder contesting the outcome asks for, and a rule can be changed in an afternoon with no retraining. Its weaknesses mirror that. Rules pile up in layers, contradict one another, and a book ends up with thousands of conditions that nobody reads end to end any more. They see only what someone thought to write, with hard thresholds that fraudsters learn fast: once a ring knows that a claim above 3,000 euros goes to a loss adjuster, its claims start clustering around 2,900. A learned model picks up combinations nobody would have spelled out, but it explains its outputs less well and drifts silently. Many insurers run both side by side, rules holding the regulatory requirements and the clear declines, the model ranking the rest, and the real trade-off is about who has the final say when they disagree.