Large Language Models (LLMs) are increasingly integrated into code editors to provide AI-powered code suggestions. However, a significant proportion of these suggestions are ignored, leading to wasted resources, increased latency, and user frustration. In this work, we introduce a lightweight behavioral pre-filtering mechanism that predicts the likelihood of suggestion acceptance based solely on real-time developer telemetry—such as typing speed, file navigation, and editing activity—before triggering the LLM. Our method was deployed in a production-grade Visual Studio Code plugin used by professional developers in a naturalistic setting over four months. Empirical results show that the acceptance rate improved from 18.4% to 34.2%, while the total number of LLM suggestion requests was reduced by approximately 35%. This demonstrates that behavioral cues alone can be leveraged to meaningfully optimize both user experience and system efficiency in LLM-assisted coding workflows.