Agentic Agent, Better Agent: Evidence from Agentic AI in Telemarketing

Large language models are increasingly being deployed as agentic AI systems that autonomously manage customer interactions, yet their effectiveness in persuasion-intensive telemarketing settings remains largely unknown. We examine whether agentic AI can become a better agent in outbound financial telemarketing through a collaboration with a leading Chinese FinTech firm. The firm deployed three telemarketing agent types in parallel: human agents, a standard LLM-based agentic AI system, and a retrieval-augmented generation (RAG) system grounded in verified internal knowledge. Using a quasi-experimental design and 7.41 million unique customer interactions, we find that standard agentic AI increases the odds of same-day loan initiation by 96.0% relative to human agents, while RAG-enhanced agentic AI increases those odds by 213.9%. RAG-enhanced agentic AI also outperforms the standard system by 39.8%, highlighting the value of grounding agentic AI in verified organizational knowledge. To understand why agentic AI performs so effectively in this context, we draw on the competence-warmth framework to examine the underlying mechanisms. The evidence suggests that agentic AI delivers both greater warmth, reflected in a more positive and more stable emotional tone, and greater competence, reflected in the superior performance of the RAG-enhanced agentic AI, stronger effects among competence-sensitive customers, and a widening AI advantage as human agents accumulate working hours. Although the AI advantage attenuates among customers more likely to recognize the AI identity, it remains economically substantial relative to human agents, suggesting that the effectiveness of agentic AI is robust to AI identity detection. These findings show that agentic AI can outperform human agents in trust-dependent financial sales and demonstrate the business value of retrieval augmentation in customer-facing agentic systems.