METHOD FOR AUTOMATED INTEGRATION OF ECOMMERCE MARKETPLACES WITH EMAIL SYSTEMS FOR CENTRALIZED PROCESSING OF CUSTOMER MESSAGES AND BUSINESS PROCESS OPTIMIZATION USING LANGUAGE MODELS OF ARTIFICIAL INTELLIGENCE

Abstract

This paper presents a method for integrating eCommerce marketplaces with traditional email services to enable centralized processing of customer messages using language models. The growing volume of inquiries across multiple communication channels (email, messaging apps, internal platform systems) places significant strain on customer support teams, leading to response delays and reduced customer satisfaction. In most cases, these inquiries are textual, often repetitive, and fall into common scenarios that can be automated. In response to these challenges, we propose a comprehensive approach that consolidates incoming messages into a unified email flow, where they are automatically processed. 
The method relies on centralized aggregation of messages sent by marketplaces (e.g., Amazon, eBay, Etsy) to seller-specific email addresses. Upon receiving a message, the system performs automatic classification to determine the type of inquiry (e.g., “delivery question,” “product complaint,” “refund request”). This classification is carried out by an LLM that analyzes both the current message and its conversation history. Based on the identified category, the system either generates a fully automated response (for common questions such as delivery tracking or return policy) or forwards a draft reply to a human operator for review and customization. 
A key feature of the system is its ability to summarize long message threads using a dedicated summarization module. This prevents token overflow and keeps the model focused on the most relevant facts. The generated response includes context such as order details (retrieved via API) and support knowledge base articles. If the model exhibits high confidence, the reply is sent automatically to the customer. In more complex or uncertain cases, the message is escalated to a human support agent. 
A prototype of the system was implemented and tested using real-world customer messages from marketplaces. The model achieved ~95% classification accuracy across predefined categories, and ~60% of inquiries were resolved automatically without human intervention. As a result, the average response time for typical requests was reduced by a factor of 5–6, and the overall workload on support staff decreased by more than 60%. 
The proposed solution holds practical potential beyond eCommerce, including application in banking, telecommunications, healthcare, and any domain where high volumes of text-based customer interactions occur. The architecture supports modular extensions, multilingual support, and integration with live chat and messaging services.  
Future work will focus on expanding the system’s capabilities, including continuous learning from past interactions, more robust fallback mechanisms, and better confidence estimation. As the volume of digital communications continues to grow, the presented approach offers a path toward next-generation.