Pretesting provides essential evidence about item quality, but traditional approaches require substantial candidate response data before item performance can be evaluated. This session presents an AI-based simulated administration approach for generating earlier item performance evidence before calibration. The method estimates how synthetic examinees at different ability levels would respond to each option in a multiple-choice item based on the targeted content, keyed response, and inferred functionality of the distractors. AI-generated option probabilities are then converted into item difficulty estimates and compared with operational item statistics. The session will discuss evidence from an IT certification assessment and practical applications for item review, pretest planning, unscored item evaluation, and more efficient calibration workflows.