Intent Inference Mechanisms in Proactive AI Assistance Systems for Knowledge-Intensive Workflows

Abstract

Proactive AI assistance systems increasingly move beyond reactive command-response interaction by inferring user intent and initiating support during complex knowledge work. This study proposes an intent inference mechanism for proactive AI assistance in knowledge-intensive workflows, integrating contextual event representation, seven-class intent modeling, confidence-gated intervention, and workflow-level validation. The framework models seven intent classes: information seeking, source comparison, citation validation, synthesis writing, decision preparation, workflow recovery, and task planning. Results show that the model achieved the strongest recognition performance for information seeking, with 0.91 accuracy, 0.89 macro-F1, and 0.07 calibration error, followed by synthesis writing, with 0.88 accuracy, 0.86 macro-F1, and 0.08 calibration error. Workflow recovery remained the most difficult class, producing 0.76 accuracy, 0.73 macro-F1, and 0.16 calibration error due to ambiguous pause and restart behavior. Feature contribution analysis showed that semantic features had the highest mean weight at 0.290, followed by operational features at 0.253, contextual features at 0.210, temporal features at 0.156, and historical features at 0.091. Proactive assistance acceptance was highest for citation checks at 74%, source summaries at 68%, and decision summaries at 65%. Compared with no assistance, the proactive intent-aware condition reduced average task completion time from 47.6 to 38.9 minutes, decreased redundant actions from 31.4 to 17.9 per task, and increased cognitive continuity from 71.2 to 86.4. Error analysis further showed that confidence-gated intervention reduced total inference-related error from 28.4% to 16.7%, with premature intervention declining from 9.6% to 4.1%. These findings indicate that proactive AI assistance becomes effective when intent inference, confidence calibration, timing control, and bounded intervention are treated as an integrated augmented intelligence mechanism.