Humans are constantly facing decisions that need to be made quickly without access to complete information. Yet, despite the complexity of naturalistic decision problems, most classical decision-making studies rely on artificial settings that might fail to capture these real life demands. To address this issue, we conducted a multi-attribute decision making study to understand human decision making strategies in everyday problems. Large language models were used to generate naturalistic problems under two conditions: one where the order of feature importance was known (ranking condition) and one where the direction of correlation with the outcome was known (direction condition). Participants (N = 123) were not informed about the condition they were assigned to and randomly placed into groups. We compared three computational models (weighted additive, equal weighting, single cue) across three binarization strategies (none, relative, absolute threshold). Model comparison revealed that in the ranking condition, the single cue model characterized human decisions the best, indicating that participants made decisions by focusing on one dominant cue. In contrast, in the direction condition, the equal weighting model dominated, indicating that participants attributed equal importance to both cues. In addition, models without binarization explained the observed behavior best in both conditions. This means participants processed cue values as continuous quantities rather than mentally categorizing them as “high” or “low”. Taken together, these findings imply that human decision strategies are flexible and adapt to task demands, offering new insights into decision-making in naturalistic contexts.