Can AI and AI-Hybrids detect persuasion skills? Salesforce hiring with conversational video interviews

Abstract

We develop an AI and AI-human-based model for salesforce hiring using recordings of conversational video interviews that involve two-sided, back-and-forth interactions with messages conveyed through multiple modalities (text, voice, and body language). We derive objective, theory-backed measures of sales performance from these modalities leveraging recent advances in body language analysis, conversational analysis, and large language models (LLMs). These measures serve as explanatory variables in our AI model. Our key contribution to the broader research on persuasion and influence is that we show how to use conversational videos to capture features related to (i) two-way conversational interactivity; (ii) real time adaptation and (iii) human body language, with minimal measurement error relative to extant survey-based approaches that suffer from recall biases. We use rubric-based scores by panels of sales professionals (correlated with hiring decisions) to isolate a candidate's "latent sales ability;" and use these as outcome variables to be predicted by the AI model. The AI model achieves reasonable predictive accuracy, yet incorporating human judgments into an AI-Human hybrid model enhances its effectiveness-improving workforce quality by 67% over random selection. While the content of what is spoken is most important in prediction, conversational interactivity, sellers' real-time adaptation to the buyer, and body language also have good explanatory power. Finally, in terms of performance-cost trade-offs, the addition of just one human professional evaluation in the hiring loop in combination with AI is optimal. Further, using human input based on only the two early stages of the interview in a task-based hybrid model is the most cost-effective in improving performance.

Keywords: AI-human, Video Analytics, B2B, Salesforce, Machine Learning