Beyond Keywords: Intent-Aware Message Routing for Gmail and Slack Using Large Language Models

Abstract In many corporate business organizations today, message misrouting has become a persistent problem among professional team members, because communications frequently reach unintended recipients and will therefore require manual redirection, this leads to delays in response times and disruption to workflow. There are existing message automation tools designed to handle this like Gmail filters, Slack workflows, Zapier, and Make but they rely largely on keyword or condition matching, which can perform poorly given the natural variability of message phrasing in real communication. This study presents the design and implementation of an intelligent message routing and triage platform that applies natural language classification to automatically process and route incoming Gmail and Slack messages. The platform integrates the LLaMA 3.1 large language model, which was accessed via the Groq API, with the Gmail and Slack APIs, enabling semantic rather than keyword-based classification. The design of the system allows users to define routing rules in plain English through a web-based dashboard, while a background polling service monitors connected inboxes every five minutes for new messages. Each incoming message is classified according to semantic intent against user-defined rules, triggering one of three actions (forwarding, automated reply, or ignoring) or, where no rule matches or the classifier flags the message for user action, escalation to a human reviewer, with every event logged for auditability. The system was developed iteratively across four development cycles and implemented as a full-stack TypeScript application using Next.js, Express.js, and MongoDB. The evaluation comprised functional testing, integration testing, and a preliminary feasibility check of the classifier. Functional testing achieved a 100% pass rate across all component categories (34 tests). The feasibility check used 14 author-constructed messages covering four of the six supported rule categories: 12 were matched to the intended rule (85.7%; 95% Wilson confidence interval 60.1–96.0%) and two were not matched and were escalated to human review, whereas a simulated keyword-matching baseline matched 5 of 14 (35.7%; 16.3–61.2%). Because the set is small, was written to avoid the rule wording, and was labelled by the authors, these figures illustrate the system's behaviour and do not establish comparative accuracy. Both unmatched messages involved indirect scheduling language not explicitly captured by rule conditions, indicating that classification performance remains sensitive to rule specificity. The key contribution is a documented system design, including an abstain-and-escalate decision rule, for embedding large language model classification into multi-channel business productivity tools, with particular relevance for Nigerian and African small and medium-sized enterprises seeking affordable, accessible, and intelligent communication automation.