BoardroomAI Enables Human-Steerable Multi-Agent Decision Making

Organizational decision-making is a dynamic process where evidence, constraints, and human priorities constantly change. Current multi-agent systems often operate in a "fire-and-forget" manner, where humans provide an initial problem and receive a final response after internal agent deliberation. This limits human involvement in the evolving decision process. Researchers have developed BoardroomAI to address this, treating the human as a continuous participant who can intervene at any stage. This human-agent coexistence is operationalized through four key components: a typed decision graph (representing evidence, assumptions, risks, decisions, and dependencies), an intervention compiler (converting human actions into graph updates), dependency-aware propagation (identifying affected subgraphs and reactivating relevant agents), and an evaluation framework. In synthetic experiments, dependency-aware propagation matched exhaustive impact computation while inspecting only a fraction of nodes. An exploratory pilot showed selective repair recomputed necessary nodes, preserved unaffected ones, and produced valid updated decisions in many cases. This framework highlights the need for "decision-sufficient context closure" to ensure agents have enough information for synthesis after human interventions, moving towards more collaborative and steerable AI deliberation.