A Multi-Agent Large Language Model Framework for Marketing Decision-Making with Auditable Attribution Analysis

This paper addresses the challenges of inconsistent indicator definitions, complex attribution paths, and difficulty in tracing decision-making criteria in multi-channel marketing environments. It proposes a multi-agent large-scale model framework for marketing decision-making, constructing an end-to-end analysis loop around problem parsing, definition alignment, and auditable output. The system consists of a problem agent, an Metric Agent, and a decision agent: the problem agent maps natural language requirements into structured task specifications, clarifying target indicators, time windows, analysis dimensions, and constraint sets; the Metric Agent performs field mapping and indicator standardization on multi-source event data, unifying window filtering, deduplication rules, and cross-platform merging strategies, and providing validity judgments and consistent aggregation criteria for fine-grained slices; the decision agent generates executable resource allocation suggestions under budget and business constraints, and simultaneously outputs a recalculated evidence chain, ensuring the traceability of suggestion sources, calculation paths, and definition versions. Comparative experimental results show that the proposed method maintains strong robustness and reliable delivery capabilities under a multi-dimensional evaluation system, particularly excelling in constraint satisfaction, verifiable output, and auditable coverage. It provides an interpretable, traceable, and implementable decision support solution for marketing attribution assessment and ROI optimization.