Crew legality violations pose significant operational challenges for airlines. Introducing buffer times in crew flight duty periods during the planning phase can mitigate these violations but often lead to higher planning costs. This paper presents a machine learning–driven framework, coupled with a simulation-based analysis, to balance this trade-off. We develop CatBoost models that accurately predict flight delays, translating these predictions into optimal buffer time allocations within crew schedules. Our findings highlight the critical role of delay prediction profiles and conservatism levels in achieving this balance. A case study using American Airlines data demonstrates how our approach helps decision-makers identify buffer levels that minimize legality violations while controlling costs.