An Architectural Proposal for Self-learning Artificial Intelligence Agents for Automated Query Resolution in E-Commerce

Abstract

E-commerce platforms increasingly rely on conversational agents to support customers, evolving from rigid rule-based systems to advanced assistants enabled by large language models (LLMs). These modern systems show strong abilities in reasoning and understanding context. However, the automation of post-purchase interactions such as delivery updates, return requests, or dispute handling remains limited, especially for small and medium-sized online businesses. Present-day solutions largely address pre-purchase guidance and basic FAQs, leaving a significant gap in after-sales automation.
This paper reviews major academic studies and industry-grade chatbot architectures, highlights limitations observed in current implementations such as the Aurora Shop assistant, and introduces a self-learning agent specifically designed to automate complex post-purchase workflows. The proposed architecture brings together LLM-based decision planning, the model context protocol (MCP) for controlled tool execution, and a Neo4j-powered knowledge graph for transparent and verifiable reasoning. A proof-of-concept model demonstrates how the system can manage real-world, multi-step customer issues without requiring heavy infrastructure. The work aims to offer a practical, adaptable, and transparent solution suitable for academic research settings as well as smaller e-commerce enterprises.