Boston Consulting Group is taking an unusual approach to AI training, teaching its customer-facing agent, Jamie, what not to do. The firm is feeding the AI both the strategies of its strongest sellers and the blunders of its weakest ones. Japjit Ghai, a managing director at BCG X, described the method on a company podcast.
This dual-input system is designed to refine Jamie's sales performance by learning from negative examples. Rather than relying solely on ideal scenarios, BCG is incorporating real-world failures to prevent the AI from repeating common mistakes. The approach mirrors a broader industry trend of companies remaking their org charts using AI agents trained on human workers.
BCG draws from three data streams to train Jamie: its own internal expertise, client knowledge of their business, and a company's existing sales call transcripts. The agent analyzes patterns from high-performing sellers' call transcripts to understand effective customer engagement. Simultaneously, it studies poor sales interactions to learn which behaviors to avoid.
Ghai stated that the agent studies the best sellers' call transcripts and how they engage with customers. He also confirmed the team is actively teaching Jamie not to replicate the worst seller experiences. The training methodology raises questions about how effectively an AI can differentiate between nuanced successful and unsuccessful behaviors in complex sales environments.
Critics might argue that training an AI on negative examples could inadvertently teach it suboptimal patterns. An AI trained on human errors risks amplifying those flaws if the boundaries between good and bad behavior are not clearly defined. The success of this approach hinges on BCG's ability to precisely codify what constitutes a poor sales decision.