Preprint

Sales performance in real-time, high-pressure environments depends critically on a representative's ability to respond persuasively and promptly to prospect objections and buying signals. Traditional sales coaching models are retrospective — feedback is delivered after the call concludes, too late to influence the interaction. This paper presents PitchSense, a production-deployed, browser-based SaaS platform that delivers real-time AI coaching to sales representatives during live calls. PitchSense captures speech via the Web Speech API, routes each prospect utterance through a secure Next.js serverless proxy to Anthropic's Claude Sonnet large language model (LLM), and renders contextually appropriate coaching responses — termed "whispers" — on the representative's screen in under 2 seconds. Claude classifies each utterance as an objection, buying signal, or neutral using full semantic understanding, bypassing brittle keyword-matching approaches entirely. Post-call, the system generates a composite performance score (0–100), talk ratio analysis, letter grade, and a self-building playbook of top conversational moments. We evaluate PitchSense across five performance dimensions: end-to-end latency, signal classification accuracy, whisper response quality, call scoring consistency, and user experience. Preliminary architectural evaluation estimates a mean end-to-end whisper latency of 1.73 seconds, objection classification accuracy of 87.3%, and a System Usability Scale (SUS) score of 81.4. Note: The experimental results in Section 5 are preliminary simulations based on the system’s architectural characteristics; a full empirical user study is currently in progress. This work establishes PitchSense as a viable real-time human-AI collaborative coaching architecture with significant implications for conversational AI, enterprise productivity, and sales performance research.