AdversarialCI: A Buyer-Adaptive Multi-Agent Framework for Evidence-Grounded Competitive Intelligence

Competitive intelligence (CI) tools in enterprise software procurement are predominantly static, producing generic battlecards that fail to account for individual buyer priorities. We introduce AdversarialCI, a domain-agnostic multi-agent framework that simulates an adversarial court proceeding between vendor advocate agents. Verdict dimensions are dynamically weighted by a structured buyer profile termed the "plaintiff". The framework consists of three core components: (1) a multi-source evidence ingestion pipeline with automated verification and confidence scoring; (2) a three-round adversarial court simulation where LLM-powered advocates argue solely from verified evidence; and (3) a plaintiff-weighted judge that produces dimension-level verdicts, an overall winner, and a "swing factor"—the single buyer constraint most likely to change the outcome. We demonstrate AdversarialCI in the database vendor selection domain across five vendors, achieving 87.9% accuracy against a manually curated ground truth dataset. Critically, the system produces different verdicts for different buyer profiles on identical vendor sets, validating that the framework is genuinely buyer-adaptive. This represents the first system to combine adversarial multi-agent simulation with plaintiff-specific verdict routing for competitive intelligence. The architecture effectively addresses the tendency of LLMs to generate unsupported claims by constraining agents to a verified evidence bank and making source citation mandatory .