AI-driven decision support to improve task handover through mental model reconciliation

In preparation for submission to CSCW 2027. Handover is one of the most frequent sites of miscommunication (and downstream errors) across industries ranging from software engineering to to manufacturing to healthcare. We present an LLM-enabled pipeline that addresses handover by combining two complementary sources of evidence: system-visible observations from the task session, and the handover report notes provided by the outgoing user. Our pipeline reconciles state information from both input sources by converting each to a knowledge graph, aligning and merging them, and using the resulting unified representation as the basis for LLM-generated handover reports, which can be presented alongside the strategic guidance that only the user can provide. Our work presents three contributions: (1) A conceptualization of task handover as a model state reconciliation problem, framing the handover report as a lossy compression of an agent’s task model and identifying state information coverage as the primary failure mode; (2) A system pipeline that operationalizes this framing explicitly, combining system-side task observations with human-authored notes via deterministic knowledge graph merging and LLM-assisted filtering to produce concise handover reports with improved information preservation; and (3) an empirical evaluation using a simulated cognitive stacking task, introducing an evaluation framework for handover performance and demonstrating improved information coverage and reduced misinformation in resulting reports.