A new benchmark measures whether AI sales agents can survive a buyer who actually pushes back.
Researchers have released SmartSalesBench, built to test sales chatbots against buyers who argue, stall, and object instead of cooperating. Most dialogue benchmarks assume the person on the other end of the chat wants the task done; this one flips that assumption and scores agents on how they handle resistance. The setup pits five agent profiles against twelve buyer personas across fifteen smartphones in six price tiers, with conversations running up to twenty turns. Across four controlled experiments repeated three times, the team logged 13,680 sessions, with both the sales agents and the buyer personas generated using retrieval-augmented generation for grounding.
Most chatbot benchmarks assume the user already wants help. Sales doesn't work that way - the buyer's job is partly to resist, and an agent that can't handle a skeptical buyer is useless at the one task it exists for. SmartSalesBench pairs automated LLM judging with human review - 50 raters, a 0.71 agreement score, and a 0.81 correlation with its primary outcome metric (BRS) - for a more credible read on whether these agents can actually close, not just chat.
Whether any agent clears that bar yet is a separate question - and the one this benchmark seems built to keep asking.