A new benchmark says AI still can't figure out why customers actually buy.
Researchers built SDR-Arena and a companion 50,000-example corpus called SDR-Bench, drawn from customer success stories across 22 industries and 3,500 enterprises, to test whether AI models can reconstruct the specific arguments that won a real sale. Models see only the information available before the deal closed and must guess which selling points mattered, scored against a weighted nugget-recall metric. The best performer, Claude Sonnet 4.6, recovered just 55.8 percent of the winning content, a ceiling that held across other model families and that longer, costlier research pipelines could not break. A follow-up test traced the failure to retrieval, not reasoning: hand a model the facts a human researcher would dig up and its score jumps; let it search the web on its own and it comes up short.
That undercuts the pitch that AI can personalize sales outreach as well as a seasoned rep. Two separate reviews by professional sales reps rated only 48 percent of AI-written pitches as usable without edits, a sign that sounding personalized and being right are not the same thing, and that the real bottleneck is finding the correct facts, not writing persuasively once you have them.
If your AI sales tool promises to know exactly what will close the deal, ask it to show its sources first.